Most photographs will never live inside one company like that. They move through a system where the camera maker, the company that issues the digital certificate, the people who write the checking software, the platform that shows the result, and the newsroom or court that acts on it are all different organizations. Each makes its own decision. Whether a particular photograph gets believed comes out of all of those decisions together.

This piece is about that arrangement. Who decides what, what the published documents say those parties owe each other, and where the decision finally turns into something that happens to a photographer. Part I stated the hypothesis this series is testing, that a system which makes decisions people have to live with, and never says how those decisions can be questioned, uses up their willingness to go along with it. Part II does not test that hypothesis. Its job is to map who is making the decisions, because the question of accountability can’t be asked properly until that map exists.

The pug in the jet

In August 2025 Nikon added Content Credentials to the Z6 III through firmware 2.00. Content Credentials are the consumer-facing name for C2PA, the industry standard for attaching a signed record of where an image came from. When a Z6 III took a picture with the feature on, the camera signed it with a secret key and attached a certificate, a kind of digital ID issued by a certificate authority, that tells anyone checking the file whose key made the signature. Anyone checking the file later could confirm it came off that camera and had not been changed since.

Within weeks, a forum researcher who posts as Horshack found a way around it. As DPReview reported, he used the camera’s multiple exposure feature to feed it an image that did not come from its sensor, and the camera signed the result as a genuine capture. His demonstration image was an AI-generated picture of a pug flying a jet. It carried a valid Nikon credential.

Nikon suspended the service. It also announced that the certificates loaded onto cameras between the launch and the suspension “will be invalidated,” which meant every credential issued in that window, including those on photographs taken honestly by photographers who had done nothing wrong. As Digital Camera World put it, those images could no longer serve as proof of where they came from.

That was a reasonable response to a real forgery. It also depended on something Nikon did not control.

Invalidating a certificate only works if the software checking a photo goes and asks whether the certificate is still good. The tools most people used to check C2PA credentials at the time did not ask by default. PetaPixel found that the invalidated credentials kept passing in those tools, and only failed when someone changed a setting to force the check. Horshack filed a request asking the maintainers of c2patool, the standard’s command-line checker, to make revocation checking the default. Nikon had no way to make anyone else’s software listen.

So three different parties made three different decisions. Nikon decided to invalidate. The people who built the checking tools had already decided, some time earlier, not to check for invalidation unless asked. And whoever received one of those photos, whether a person, a newsroom, or a platform working behind the scenes, got whatever answer their checker gave, usually without knowing that the choice of checker mattered.

The result was more than one outcome. An honest photographer’s credential was now dead in one checker and valid in another, and the photographer had no way to know which checker any given editor, client, or court was running. The photographer could check their own credential and get an answer. What they could not know was which answer any particular recipient would get.

Why it is still true, for two different reasons

That was September 2025. Most of it still holds a year later, for two separate reasons that are easy to blur together.

The first reason is a design choice in the standard itself. A credential can carry proof, captured when it was signed, that its certificate was good at that moment. C2PA’s security guidance says a credential can be “validated indefinitely” even if its certificate is later revoked or expires. That is deliberate, and mostly sensible. An honest photograph taken this year should not stop counting in five years because a certificate was retired for some unrelated reason.

The main open-source library for reading C2PA credentials is c2pa-rs, maintained under the Content Authenticity Initiative, and c2patool is built on it. In its current settings, the option to go online and ask whether a certificate has been revoked since is off unless someone turns it on. With it off, the checker relies on the status information the credential carried when it was signed.

Nikon’s case is where that design and the maker’s intent collide. Nikon was not retiring certificates for a routine reason. Its camera had been shown to sign images it should not have signed, so Nikon wanted every credential from that window to stop counting as proof, going backward. A checker that judges a credential by its status at signing time sees those credentials as good, because at signing time they were. The same design that protects an honest photo from a later, unrelated revocation is the one that let the compromised window keep passing.

Leaving the online check off also has a real argument for it. The C2PA specification makes the check optional, and asking a certificate authority about a certificate tells that authority which certificate you are holding. When a camera uses a fresh certificate for every photo, as Google’s Pixel does, that means telling it which photo you are looking at. The library’s settings file does not say why its maintainers chose the default, but the tradeoff has a sound argument on both sides.

The second reason is a bug, and it runs in the opposite direction. With online checking turned on, c2pa-rs can report a good certificate as revoked. In one reproduction filed against the library, a photo signed by a consumer app was checked at a moment when the certificate authority’s clock ran about two seconds ahead of the checker’s. The library reported the certificate as revoked and treated the signer as untrusted. The same response, checked again two minutes later, came back clean. A second report describes the same problem with older “good” responses. As of this writing, a fix is under review and may well be merged by the time you read this.

Those are two different failures from two different configurations. With online checking off, a credential its maker has tried to withdraw retroactively keeps passing. With it on, a good credential can, under specific timing conditions, be reported as revoked. Put side by side, they mean a photographer still cannot count on the same answer from every checker.

Taken alone, none of this is a legitimacy problem. It is software disagreeing with other software, which happens in every distributed system while its implementations mature. What turns a technical disagreement into a question about authority is what happens when the answer reaches someone who has to act on it.

Read more

That is correct, and it is where the harder part starts. A promise of fair terms leaves three questions unanswered. Who controls the use? Who receives the money? Who can enforce the bargain? In the history he cites, the answers were not always the people who did the work.

What he gets right

Katzenberg asks Hollywood to stop spending its energy on making AI go away and to spend it on the conditions of use. He asks the labs to build with credit, consent and compensation for the people whose work trained their models. He says plainly that the content of fairness is a debate still to be had, once everyone agrees there should be terms at all.

The Animation Guild made the same calculation in 2024. Explaining its contract to members, the union said a ban on generative AI was never on offer and would have pushed the work to non-union shops. It bargained over the terms. So the useful question is what those terms let each party decide, collect and enforce.

The bargain of 1909

Katzenberg’s first precedent is John Philip Sousa, whose 1906 attack on “mechanical music” was part of a campaign for copyright reform. In 1908 the Supreme Court held that a piano roll was not a copy and owed no royalty. Congress responded the next year.

What Congress granted was a qualified right. Copyright owners had asked for the exclusive right to authorize mechanical reproductions. The 1909 Act gave them control over the first recording of a work. Once an owner had allowed one, any other manufacturer could record the same composition by paying the copyright proprietor two cents for each copy made.

There was a public reason for the design. Congress was concerned about monopoly control of recorded music, and the compromise did get owners paid. It also put the price in the statute, where it stayed at two cents until 1978.

The model has been considered again. The Copyright Office examined compulsory licensing for AI training in its 2025 report and concluded it would have significant disadvantages, because such licenses fix rates and terms that become hard to undo. Katzenberg does not call for one. He credits Sousa with changing the terms, and the terms that emerged included a rate that did not move for nearly seven decades.

The theater musicians’ union, the American Federation of Musicians, could bargain over recording-session wages, and it set scales for soundtrack work as early as 1928. It had no claim on the use of those recordings in thousands of theaters. By its own count, 22,000 theater jobs were gone by 1930 against fewer than 200 created. A payment tied to recorded music arrived in 1944, after the union refused to record for more than twenty-seven months. The money went into a fund that hired musicians for free public concerts, and the players on the records did not receive it.

Who owns the work

Katzenberg writes that the knowledge the labs need lives in people whose life’s work informed the models. For most studio artists that life’s work belongs to someone else. Under the work-made-for-hire rule, the employer is the legal author and owner of what an employee creates on the job, unless a signed agreement says otherwise. Permission to license the copyright generally rests with its owner.

The arrangement Disney and OpenAI announced in December 2025 shows how that works. It covered generating short videos and images with more than 200 Disney characters. The announcement said nothing about training models, and it excluded talent likenesses and voices. The Animation Guild said its members had never been paid for the licensing of those characters. The transaction was subject to definitive agreements, and Disney exited in March after OpenAI announced it would discontinue Sora. As announced, the proposal would have authorized character generation through an agreement between the two companies.

A studio’s permission to license an artist’s work is not the artist’s consent. Payment to the studio is not necessarily compensation to the artist. Ownership sets that default and nothing more. Writers and actors do not own the films they make either, and they are paid residuals because their unions bargained for them. An individual with enough leverage can negotiate participation in a personal contract.

What the contracts secure

The clearest individual veto in Hollywood covers a performer’s own face and voice. Under the SAG-AFTRA agreement, a producer who wants to use a digital replica of an identifiable actor generally needs that actor’s consent, separately signed, with a reasonably specific description of the use. There are exceptions, including routine post-production alterations.

Synthetic performers, which resemble no one in particular, work differently. There is no individual to ask, so the producer must notify and bargain with the union. If the producer breaks its commitments, the union can arbitrate for damages that are not necessarily limited to what a human performer would have been paid. If bargaining fails and the producer uses the synthetic anyway, the union can arbitrate what payment, if any, is owed. That money, negotiated or awarded, goes to the union’s health plan or a pension plan, at the union’s election. As with the musicians’ fund in 1944, the payment supports the workforce as a whole and does not reach an identifiable displaced worker.

On training, performers and writers have similar terms. When a producer licenses covered work to train a commercial AI system, it owes the union written notice and, on request, a discussion that may include remuneration. The 2026 Writers Guild contract keeps the protections won in 2023 and adds that duty. The guild had gone into talks asking for payment.

The animators’ agreement says a producer may require employees to use any AI system. It also protects them in several ways. Using AI does not reduce an employee’s rights or credit. The producer cannot require an employee to furnish prompts in a way that displaces a covered employee, and it must indemnify employees, subject to conditions, against claims arising from AI use on the job. It must give written notice in advance, and an employee can ask for a consultation about alternatives if the production schedule allows time.

Violations can be arbitrated with every remedy available except an injunction, and the arbitrator has no authority to restrict the AI use itself. A separate article preserves the producer’s “unrestricted right to make technological changes.” It provides displacement pay for eligible employees, separate from dismissal pay, that tops out at ten weeks for twenty or more qualifying years.

One clause in that agreement shows the pattern well. An animator can refuse to have her face or voice scanned, and consent cannot be made a condition of the job. The article grants no corresponding veto over licensing her studio-owned drawings for training.

Katzenberg restates his 2023 prediction that AI could cut the time and cost of world-class animation by as much as ninety percent, and he expects more films and more seats at the table as a result. He may be right about the films. A lower cost of production does not say who receives the savings, which can go to margin, to a larger slate, or to the people doing the work. Agreements like these help determine the workers’ share.

Three questions

I think Katzenberg is right to put the terms first. But if the promise is compensation for creators, payment to the company that owns their work is not enough. The bargain needs to secure a benefit for the people whose work makes the license valuable, whether through individual payments or collectively negotiated funds. Otherwise, the companies have settled their rights without delivering the compensation to creators that Katzenberg calls for.

So I would hold every future promise of fair terms to three questions. They are who holds the right to say no, who gets paid, and who can compel compliance when the terms are broken.

Run on his own cases, the test gives specific answers. In 1909 the owner controlled the first recording, the proprietor was paid two cents, and a court could compel payment. In 1944 a union withheld its labor, a fund was paid, and a contract could compel. Today an animator can refuse the scanning of her likeness without losing the job. The same agreement permits the producer to require her to use AI in covered work.

The Animation Guild agreement runs through July 31, 2027, and either side can serve notice to renegotiate by May 1 of that year. That negotiation is the next concrete opportunity to change these collectively bargained terms. Any promise of fair terms made before then can be put to the same three questions.

Three shows sent me back through the record, this year’s and 1948’s. On Inside the Stream, Will Richmond called the production commitment squishy and the per-film penalty cheaper than a film. On Cartographer’s Log, Evan Shapiro admitted he had expected structural remedies and then described a trustee reporting to the court with liquidation as a remedy. On Cord Killers, Tom Merritt read the terms straight and Brian Brushwood called the whole thing scaffolding so the people inside don’t take the full shock when the change everyone knows is coming arrives. All three were arguing about the settlement. I kept thinking about the other times Paramount, under very different leadership, had contended with regulators.

What the federal government decided

The Antitrust Division’s closing statement from June is worth reading in full, because it describes the problem and then declines to act on it. It examined streaming, linear television, and theatrical film.

On streaming it found the two companies were late entrants who would compete harder combined, and it said plainly that YouTube and TikTok are not substitutes under existing precedent. On linear television it found the market in decline and competition for live rights robust. On theatrical film it found smaller studios making bigger bets, cited A24 and Amazon and Lionsgate, and in a footnote invoked a 1974 Supreme Court case to warn against relying on historical market shares. It dismissed the labor theory in a paragraph. It called its own review rigorous, eight months, two million documents, led by career staff, with the states sitting in on depositions.

Then it listed AOL and Time Warner in 2001, AT&T and Time Warner in 2018, and Warner Bros. and Discovery in 2022, and observed that the legacy of those deals shows what happens when a merger’s commercial logic does not line up with the acquirer’s competitive incentives. It called Warner Bros. a repeated acquisition target, which is fair. And it cleared the fourth acquisition of the same asset in 25 years with no divestitures, no conduct remedies, and no concessions, concluding the deal would increase competition.

Four days later the Wall Street Journal reported that senior officials had closed the investigation before the career team could deliver its recommendation, and that the team had been leaning toward a lawsuit. In March the acting head of the Division had said the deal would not be fast-tracked for political reasons. Paramount’s chief legal officer is Makan Delrahim, who ran the Antitrust Division in the first Trump administration and who brought and lost the AT&T case the Division’s own statement now cites. None of that proves the decision was wrong. It establishes that the decision was a choice, made at the top, over the reported objection of the people who did the eight months of work. I cannot prove motive from outside. I can say that if you set out to design a clearance that looked political, it would look like this one, and I am disappointed the Division let it.

The case the states brought

A month later, twelve states filed under Section 7 of the Clayton Act. Their complaint defined three markets narrower than the ones the DOJ examined, wide-release theatrical distribution, distribution of anticipated top-grossing films, and licensing of basic cable channels to distributors. In each, the combined shares cross the concentration thresholds at which the government’s own merger guidelines presume harm. Five major studios take about 80 percent of the domestic box office, and the deal makes them four. The combined company and Disney would control 59 percent of basic cable. On July 20 the judge granted a temporary restraining order, writing that the states had shown “serious questions going to the merits”, and set a trial for March.

So the competitive case was strong enough to pause an $111 billion deal for two months. It was the case the federal government had the record to bring, with the depositions the states had only sat in on, and the resources of the Antitrust Division rather than twelve state offices. The states brought it alone, against a company that called it one of the weakest merger challenges in modern history, with a $7 million a day ticking fee due to start running against the company on October 1 and a bond motion for $1.88 billion pending against them. That is the position from which the settlement should be judged.

What the settlement is

The proposed decree is 32 pages, about a third of them signatures, and it is not yet entered.

It protects real things. The lots cannot be sold or closed for five years, and Shapiro is right that real estate was a large line in the $6 billion synergy case Ellison sold investors. Every film counted toward the release quota gets a 45-day theatrical window and a 90-day holdback from subscription streaming, and exhibitors keep the better of the two companies’ rental terms for three years. The quota is 30 theatrical releases a year for two years and 32 for three more, with floors on wide releases and independent films and at least half the slate produced or co-produced by the company. Miss it and the company owes $30 million per film to union funds, the Motion Picture & Television Fund, and a state antitrust fund, and an uncured shortfall means divesting Miramax. The two cable portfolios must be negotiated separately, and a material breach that survives cure means divesting six named channels. Pluto TV must stay free at current quality.

It leaves the market where it found it. The Monitoring Trustee is jointly selected by the company and a committee of state attorneys general, reports to that committee rather than the court, and is barred by the decree’s own language from any role in day-to-day operations, news, or content. Enforcement begins with a meet and confer within thirty days, mediation if either side wants it, and then a state seeking an injunction. The largest commitment, $300 million a year above 2025 US production spend, has no monetary penalty attached and no disclosed baseline figure. A force majeure clause excuses performance to the extent a recession, a strike, or a pandemic prevents it, and when the judge asked about it on September 24 the parties’ answer was the strikes and the pandemic. After two years the company can petition for relief if it can show the decree impairs its ability to compete and the change is not substantially likely to lessen competition. There is no requirement to run the studios as separate divisions. There is no divestiture on day one. Apart from a few unfinished obligations, it all expires at the end of the fifth calendar year after closing.

What the states could do alone

Rob Bonta said at the podium the settlement was “not a vote of support for this merger” and that he did not think the companies should combine. The Block the Merger coalition’s amicus brief opens by quoting his earlier insistence on structural remedies back at him. Shapiro reports that New York wanted to let the fee run. California settled ten days before it started, with Newsom and Bass pushing and a relocation threat on the table.

I am not going to absolve the states, and I am not going to pretend to be neutral about them either. They brought a challenge, won a restraining order on it, said only divestitures would do, and then took a decree with none. I understand why. Their case was narrower than the one the federal government could have brought, a March trial against a defendant with Paramount’s budget was a real risk, and the fee clock was theirs to lose. Settling for concrete protections instead of gambling on a verdict is a defensible call. It still disappoints me. They had the only restraining order anyone had won, they had the calendar for a few more weeks, and they had said in public that structure was the price. Whether another month would have bought a divestiture is unknowable from outside the room. I wish they had made the company find out.

But it is the second question. The first is why twelve state attorneys general were the only public enforcers in the room at all, carrying a Clayton Act case the federal government had the record and the mandate to bring, against a defendant that could outspend them and a clock that ran in the defendant’s favor. What they got is what a coalition of states can extract when it is the last line rather than the second. The protections went to the parties specific enough to write into a paragraph, unions, exhibitors, two lots, independent film, a news board. Competition is not that kind of party, and the enforcer whose job it is to represent it had already left.

The Wu of it All

I keep reaching for Tim Wu on this deal because he is where I learned to read large media deals. The Master Switch is a history of every American information industry and the pattern he found in all of them. Each one starts open, with many small players, and each one closes, ending in the hands of one company or a cartel that owns the channel and decides what moves through it. He called it the Cycle, and the film chapters run through Paramount. Adolph Zukor built the studio that owned its theaters at scale and used block booking to force exhibitors to take the whole slate to get the hits. It took the government until 1948 to break that, in United States v. Paramount Pictures, and the decrees that followed did two things. They separated the studios from their theaters, and they banned block booking and related distribution conduct. Structure and conduct, together. The Justice Department moved to end those decrees in 2019 on the argument that the market had changed, and a court did in 2020.

Six years later there is a new Paramount consent decree. It has the conduct half. The company may not tie one cable portfolio to the other in a carriage negotiation, which I see as the block booking prohibition rewritten for cable, and it must hold films in theaters for a set window. It does not have the structural half, because the only enforcer that could have insisted on it decided in June that none was needed.

The Curse of Bigness is Wu’s account of how that becomes the default. His argument is that antitrust in the Brandeis era treated concentrated private power as a danger in itself, which is where 1948 came from, and that from the late 1970s the Chicago School and then the courts narrowed the inquiry to effects on consumers, with structural remedies falling out of favor along the way. The current 2023 merger guidelines count wages, quality, and innovation alongside price, so the narrowing is not total, and Wu’s point is about what gets litigated rather than what the guidelines permit. The DOJ’s own closing statement is a fair example. It names the pattern, three prior owners of the same asset and the damage each did, and then treats the pattern as history rather than evidence.

Here is how I use him to read this deal, and this is my application rather than anything he wrote about Paramount. There are two kinds of fixes for a merger that concentrates a market. One is structural. You make the company sell something, or you stop the deal, so that two competitors keep existing instead of one. The other is behavioral. You let the deal close and make the company promise to act a certain way for a period of time. Wu’s history says the first kind is what changes an industry and the second kind is what regulators reach for when they have decided against the first.

Only the federal government could realistically have forced the structural fix here. It had the full record, the depositions, the budget, and a statute written for exactly this. In June it passed. Once that decision was made, the states were left with the second kind of fix, because that is the only kind a smaller enforcer can get from a company that has already been cleared. And a behavioral fix has a built-in limit. You can only write promises for parties you can name. You can write a paragraph for theater owners, for unions, for a studio lot, for an independent film fund. You cannot write a paragraph for competition itself, because competition is the thing two separate companies produce by existing, and by June there was going to be one.

So read the decree that way. Every page of it is a promise to a named party. The 32 pages are what the states could get once the only enforcer that could have asked for more had already said it would not.

Year six

This is where the deal touches the thesis I have been writing under for a year, that the model layer commoditizes and governance accumulates the durable value. Twelve states in effect bolted a governance layer onto a media company as the price of letting it exist. An output quota, a spend floor, a monitor, a negotiating firewall, an editorial board. That is public accountability, which is a different thing from commercial value, and the deal proves nothing about the thesis on its own.

What it sets up is a test, and the test has a date on it. The decree runs for five calendar years after closing. If the deal closes this fall, every obligation in it ends on December 31, 2031. The next morning there is no trustee, no quota, no spend floor, no firewall between the cable portfolios, and no court with a role in what the company does.

Some of what the company built to satisfy those obligations will still be standing on January 1, 2032, because a compliance team, an editorial board with three-year terms, and a windowing habit exhibitors have come to expect do not vanish just because the paper that required them does.

So the interesting question is what those leftovers are worth once nobody is making the company keep them. A compliance team that survives because the company finds it useful now answers to the company. An editorial board that keeps meeting because the board of directors sees no reason to disband it has exactly as much independence as that board wants it to have. A windowing practice that persists because it makes money is a business decision, and business decisions get reversed. Each of those can still be standing in 2032 and have lost, on the day the decree expired, every bit of its power to make the company do something it would rather not.

That is the split I want to watch. Governance that accumulates value for the people being governed is a constraint that outlasts the pressure that created it. Governance that accumulates value for the governor is a department. The Paramount decree is a five-year experiment on which one you get, and in 2032 the results will be sitting in plain view.

What happens now

Richmond’s framing is that the hard part starts now, and the hard part is the business. The company is raising $7.5 billion in debt on top of what it already carries. Colin Dixon’s Nielsen math has the combined company’s share of TV viewing down two points in nine months while NBCU gained. The synergy number was built partly on assets it can no longer sell. Warner Bros. staff are expecting layoffs in rounds, and the decree’s labor terms protect bargaining rights while saying nothing about headcount. Management will make the choices, but the balance sheet sets the menu, and the year-two petition is where that pressure has somewhere to go.

Tom’s open on Cord Killers was that streaming is now television, the consolidation phase is nearly over, and within a year or two the survivors will turn on each other for whatever comes next. That is Wu’s Cycle described from a living room, and it is the audience nobody in this sequence represented. The DOJ found consumers would benefit. The states negotiated for unions, theaters, and two lots. Pluto stays free, and on the paid services there is no price commitment and no bundling rule.

The judge has not signed. On September 24 she reminded the parties that a consent decree keeps the court involved, opened the docket to amicus briefs, and asked for responses to Senator Booker’s letter by September 28. Her questions ran to whether the settlement restores competition. That is the right question, and it was answered in June by people who are not in her courtroom. The terms are public. Read what the company agreed to, then read what the Antitrust Division wrote when it decided not to ask for more.

Four arcs with one progression, and a thing none of them examined. Each assumed a decision point and then described the machinery around it. Whether a provenance claim holds. Whether a credential counts. Whether a record is authoritative enough to act on. The work kept arriving at the moment of determination and kept turning to look at the pipes instead. That was the right order to work in, and it has now gone as far as it can.

The closing piece of The Control Layer put the gap directly. The layer can operate without legitimacy, but it cannot be governed without it, and no institution currently constituted is supplying what is missing. That was where the argument stopped.

So in this arc we take up the deferred question. The premise is that technical systems are beginning to distribute decision-making power. A system that determines whether something counts is exercising power over everyone who has to work with that determination, whether or not anyone involved thinks of it in those terms. Establishing that such a system works, that the engineering is sound, the cryptography current, the privacy properties real, settles a question about the viability of the system. It does not settle whether its determinations should be accepted, by whom, and with what recourse when they are wrong. Those are two questions, and the second one is what this arc will explore.

The hypothesis I want to test is that a system which makes decisions people have to live with, and never says how those decisions can be questioned, uses up their willingness to go along with it. That claim needs evidence this piece does not have, and later installments will either find it or not.

Part I has a narrower job. It shows that the question is now attached to a shipping product rather than to a scenario, and stays exact about what that product does and does not establish.

Which brings us to a camera.

On September 15, Apple published the design of Apple Reference Image, an opt-in camera mode on the main sensor of iPhone 18 Pro and iPhone 18 Pro Max. The system signs pixel data inside the sensor at the moment of capture, develops the resulting negative inside Private Cloud Compute (PCC), and returns an image carrying a post-quantum signature. It is not an implementation of C2PA. Apple’s post argues that attaching provenance after capture leaves a chain that can break at any editing step, and builds a separate path instead, one Apple operates end to end. It is a serious piece of work and I will get to why. The part I want to focus on is a short passage in the middle of Apple’s writeup, describing what happens when the system decides an image is not what it appears to be.

What the system issues, and what it takes back

Apple’s account describes a revocation system that can revoke individual photos as well as all photos from a specific sensor. During development, PCC computes a confidence score assessing whether the image has the physical characteristics expected of raw output from Apple’s sensors. Before the developed image is signed, PCC sends the photo GUID, the raw hash, the confidence score, and the sensor ID to a companion service, which records them, updates a running confidence score associated with that sensor, and confirms the sensor does not appear on a revocation list. If a low-scoring sensor is revoked, PCC stops signing its images. Devices fetch updated revocation lists on a regular cadence, and a client checks the list before it displays any reference image. Apple states that if a device is later found to be compromised, its images can be revoked and flagged retroactively.

Three different things can happen to a photograph here, and they do not mean the same thing.

The first case is an image with no credential at all, which follows from shooting in the ordinary mode. That absence describes how the photograph was taken and reflects nothing the system concluded about it. The second is a capture that goes through Reference mode and comes back without one, which can follow from a verification step or from the sensor already sitting on the revocation list. The third is a credential that was issued and later withdrawn, so the status changes underneath a viewer who may have already relied on the image.

Only the third involves taking back something the system already granted, and that case is the subject here.

Read more

We closed by naming four things we would watch for at IBC and promising to publish the verdict either way. The four tests are at the end of Part 1, for anyone who wants them.

Here it is, and it has three parts. The two of us independently arrived at the same conclusion about this market, from different evidence, and neither of us was looking for it. One of us got a significant claim wrong and is correcting it in public. And we agree on where the market is heading while differing on how long it takes to get there. All three of those are more useful to you than a victory lap would have been.

Lorenzo ran the same corpus method on IBC, and the order changed. At NAB, control led and AI came sixth.

I have analysed a corpus of 137 IBC 2026 product announcements, stripped of stopwords and vendor names, and ranked them by raw word frequency. AI leads, as expected, followed by control, IP and cloud. What’s more interesting to me, however, is that words like operational and integration are much higher in this ranking compared to my NAB Show 2026 analysis. That operational cluster suggests vendors are increasingly describing how their products fit existing operations and infrastructure.

The convergence

Lorenzo published his IBC notes on 17 September. The first section is titled Pressure, and the load-bearing sentence is that increasing pressure on media companies has translated into fewer buyers attending IBC, with some major broadcasters not attending at all this year and those who do attend staying for shorter periods.

I published a readout built from a week on the floor. Its first section argues that the pool of buyers is shrinking and that the buyers who remain are cutting the number of suppliers they work with.

We did not coordinate. The pieces went out a day apart. Caretta Research went into the show with the same position, that the industry is contracting, with consolidation removing buying organizations, survivors reducing supplier counts, equipment life being stretched, and the market decline concentrated in large hardware-heavy manufacturers while smaller software-focused vendors grow.

That is three reads and one conclusion. The interesting question at IBC 2026 was not which architecture wins. It was who is left to buy one.

This matters because it changes the test a product has to pass. Working is not sufficient. It has to earn money or take cost out, and it has to do so inside an organization that is deliberately trying to manage fewer relationships. A vendor with an excellent component and no answer to that second part is in trouble regardless of how good the component is.

My readout could only say that IBC publishes totals with no breakdown by role. Lorenzo has the mechanism behind the number.

IBC 2026 visitors were down 6% year-on-year. Even though granular data on buyer attendance is not publicly available, anecdotal evidence points in one direction. DPP’s David Thompson captures it well: consolidation and cost-cutting in the industry has led broadcasters to deprioritize trade show attendance, with some broadcasters like the BBC (as reported by Caretta Research’s Rob Ambrose) stopping all IBC travel, while others sent fewer people for less time. In line with both reports, I found vendor accounts of traffic flows inconsistent.

Aside from the economic conjuncture, there are other trends that are also working against attendance. Trade shows are becoming more regional (i.e., more dependent on local attendance), which means that IBC may have become more dependent on European attendance. The pool of European buyers has a larger share of public broadcasters, a sector facing government-led budget cuts - not just the BBC, France Télévisions has seen a 20% reduction in its budget over the last decade. And this regionalization is happening alongside a rise in local, conference-led, topical events, which means more competition. Moreover, products on the show floor have also increasingly become software defined, which weakens the traditional reasons to attend a trade event.

While we suspect that the number of buyers is down, the number of exhibitors has stayed relatively flat in the last few years, roughly meaning that fewer buyers are more stretched across the same number of vendors.

Where I was wrong

This next part is mine alone.

In the days after the show I drafted a follow-up to Part 1 arguing that the control layer had gone open, that the buy side had organized before the sell side could capture it, and that the margin had therefore moved one story up. I sent that outline to Lorenzo. It was wrong, and the error was basic enough that I want to state it plainly rather than quietly drop it.

I had been reading a vendor technical guide rather than the primary source. The EBU’s Dynamic Media Facility Reference Architecture version 2, published on 15 April, describes six horizontal layers, with Media Exchange as the fourth, sitting between the Container Platform below and Media Functions above. Orchestration, Control, Monitoring and Security are not layers in that model. They are verticals that cut across all six. Orchestration, in the EBU’s own phrasing, creates and destroys resources in each layer, and the Media Exchange layer contains its own control section.

So control is not a layer above the exchange that could go open or stay closed. It runs through the stack, and the EBU wants those interfaces open too, building on NMOS. The Story Object Model says the equivalent thing from its side, which is that nothing sits above the tools, no orchestrator and no central application. Jon Roberts of ITN, who ran the accelerator that produced it, goes a step further. “Control is not a layer here,” he told me. “It turns up at every level: the rights on an asset, the compliance position on a story, who is cleared to publish to which destination, whose judgement lifts a hold.”

I also had the direction of the economics backwards. The EBU states that the purpose of its six-layer model is to identify interfaces for interoperability “to reduce the risk of vendor lock-in and stimulate a multi-vendor economy.” That is the institution that built the open layer saying its goal is to weaken platform power, not to relocate it upward. I had no evidence against their stated intent, and the same architecture names a possible concentration point below the open layer rather than above it, since where an operator outsources to public cloud, the cloud provider controls the infrastructure layer and may supply the host and container platforms as well.

And I am dropping the scorecard. Our four August watch items were written as things to watch rather than as tests with thresholds. No minimum number of deployments, no definition of coordination, no required standards milestone, no pricing sample, no observation window. Grading them after seeing the evidence is not scoring a prediction, it is writing a favorable interpretation. That is a drafting failure and it is mine. What we can honestly report is below, and the reset at the end sets real thresholds for next time, before the evidence exists.

What actually shipped, described correctly

The most substantial open engineering at IBC 2026 came from two projects that get discussed together and should not be.

The Media eXchange Layer took the top creation award at the Innovation Awards as an open-source Linux Foundation project, with seventeen companies demonstrating it across the floor. IBC 2026 was not its origin, and that is the part most coverage has wrong. Linux Foundation intent was announced in April 2025. Early public SDK code was released in June 2025 and first interoperability was shown at IBC 2025. Version 1.0.0 shipped as a stable production release in early 2026, and the EBU’s stated goal for it was seeing first products for sale between NAB and IBC 2026. Products arrived and multiple vendors exposed the layer at the same show. That is a genuine adoption milestone, and it was a scheduled one. A standards-adjacent infrastructure project hitting the dates it published is rarer in this industry than a turning point would have been.

Be precise about what is stable. The production line is 1.0.x, carrying the Flow API for shared-memory access within a single host. Inter-host flow sharing is implemented in the 1.1 prerelease track, where Beta 1 and the Release Candidate are complete while the final release is not, and the Fabric API documentation remains marked as to be confirmed. When a vendor tells you cross-host is production-ready in their product, that is a claim about their own validation and support commitment rather than about upstream stability, and it should be evaluated on its own terms.

The Story Object Model published version 1.0 on 12 September, at the show, under Apache 2.0 for the schemas, tools and skills and CC BY 4.0 for the specification prose. Fourteen champion organizations, and they are not all broadcasters. The list spans news agencies and newspapers including AP, Reuters and The Washington Post, broadcasters including the BBC, ITV, Sky, Channel 4, Al Jazeera, NBCUniversal and Scripps, and standards and industry bodies including the EBU, SMPTE and the Global Creative and Security Community. It is days old, its reference implementation and integration guide are not in 1.0, and the accompanying skill library sits at 0.2.2 with its skills marked draft. It is not comparable to MXL’s maturity and the two should not be cited in the same breath as evidence of the same thing.

The consequential decision sits beside the schema. The consortium also published a format for newsroom rules and ten reference Skills, and I asked the authors why. Alex Bassett of NBCUniversal answered for the team. A Skill is a document, not a program. It is a newsroom rule written so a machine can check it. The vendor’s tool raises the flag, a person clears it, and the Skill never changes content or acts on its own. They put the format in the commons because a schema without a shared rule format is only a wire format, and because the rulebook was the most valuable thing any single participant could have kept, which made it the place capture would have started. In his words, they “gave away the grammar for writing a newsroom’s rules down and kept our hands off the sentences.” Jon Roberts, speaking for himself rather than the consortium, puts the principle underneath it in six words: “Standardise beneath the intelligence, compete above it.” Below that line sit the context data flow, its junctions to the standards either side of it, the shape of the rule book and the audit trail that travels with all of it. Tooling, models, agents, workflow and the final mile stay above it, where vendors and newsrooms are meant to compete. Morag McIntosh, who conceived the model with Alex & Jon, leads the strand the coverage has ignored: Story Archaeology, which takes one real story after transmission and traces how its context actually moved between the people and systems that handled it.

That is the one part of my first draft worth keeping, stated properly by the people who made the call. It is a statement of intent. Whether it held is what the second NAB 2027 test below will measure.

The adoption ladder

“Adoption” is doing too much work in the coverage. Five rungs can be told apart, and putting a vendor claim on one makes the conversation specific.

One, the code is stable. Reached at 1.0.x for the Flow API.

Two, vendors are shipping implementations. Reached at the ecosystem level. Whether a specific named product is generally available rather than in preview, beta or booth integration has to be asked product by product.

Three, multi-vendor interoperability is demonstrated. Reached repeatedly.

Four, a customer has run it in a real workflow. Reached narrowly. CBC/Radio-Canada ran selected Milano Cortina workflows on it, which the EBU describes as a proof of concept.

Five, mixed-vendor continuous production. We found no publicly documented case. Operators frequently keep production architecture private, so that is a statement about the public record rather than about what exists.

The Innovation Award citation reads “deployed by CBC/Radio-Canada.” IBC’s awards recognize completed projects in real working environments, so the word is defensible on its own terms. Nobody overstated anything. The problem is that “deployed” no longer tells a buyer which of rungs two through five they are being offered.

One counterexample belongs here rather than in a footnote. SVT’s Neo platform won the EBU Technology and Innovation Award this year after delivering more than eight hundred hours across ten parallel environments at the same Winter Games, and Neo is not documented as an MXL implementation. Software-defined production at Olympic scale does not require this layer.

Same destination, different clocks

We agree on where this ends up. We differ on how long it takes to get there, and we are going to leave that open.

My position is that the direction of travel is openness, and that the open question is what openness is for.

Here is the paradox at the center of the whole show. Open interfaces widen the set of suppliers a broadcaster could technically substitute into. Supplier rationalization narrows the set it is actually prepared to manage. The same buyer can push both at once, for the same reason, which is lower integration cost. For a vendor that changes what has to be proven. Interoperability gets you considered. Staying on a list that is being deliberately shortened is a separate test, and it is the one most vendors at IBC were not preparing for.

So the question is whether openness is being adopted to preserve substitution options, or to make a deliberately shorter list cheaper to live with. Nothing on the floor settled it.

Lorenzo’s position is that the supply side is answering it in the short term, and answering it toward consolidation. AI makes point solutions easier to reproduce, so vendors are stacking features together to make their products stickier and harder to substitute, and the pendulum is swinging from point solutions toward end-to-end platforms, through consolidation and through partnership, with some equilibrium likely rather than a full swing.

Those readings differ more on timing than on direction, and it is worth being honest about which evidence each rests on. Mine is a demand-side read from a week of buyer conversations plus the published architecture of the open projects. Lorenzo’s is a supply-side read from vendor announcements and vendor conversations. It is entirely possible that both are accurate about the half of the market each of us looked at, and that the interesting thing at NAB 2027 is which clock was right.

If the short-term tilt lasts, then all of this open engineering is how a smaller set of strategic vendors becomes easier to tolerate, and the practical beneficiaries are whoever survives the cut. If the swing back comes sooner, competition intensifies wherever those interfaces reach, though open interfaces remove only one barrier among several and certification, enterprise agreements, support dependencies, data gravity and operational familiarity can each keep a market concentrated anyway.

Here is Lorenzo’s side, in his words.

I see the journey towards equilibrium playing out from two sides.

Vendors are (rightly) worried about AI replicating point capabilities, particularly as a growing number of buyers increases investment in building technology in-house. To make their products more essential, vendors are taking one of two main routes. Some are going end-to-end, stacking features into platforms that also provide a control layer across them - let’s call this the closed front. Others are staying in the best-of-breed camp but building deeper connections with other vendors through partnerships, frameworks and interoperability initiatives - let’s call this the open front.

The difference boils down to ownership, but in an industry increasingly under pressure from cost, complexity and risk, buyers will expect end-to-end integration regardless of who owns the components. Failure to deliver it may push buyers towards the closed front even when they’d prefer best-of-breed solutions.

In the short-term, I expect buying behaviour to tilt towards the closed front, as vendors scramble to organize themselves through deeper collaboration that emulates the workings of end-to-end platforms. In the longer term, I expect a shift back towards the open front, particularly as some buyers may also retreat from their own build efforts.

About two-thirds of product announcements at both NAB Show 2026 and IBC 2026 used stacked platform language. This percentage slightly declined at IBC 2026, but too little to draw conclusions about the direction of travel.

What the buyers actually said

Agentic AI was, in my experience of the floor, the most repeated pitch of the show and among the thinnest evidence. Several vendors arrived with agent-facing control surfaces over Model Context Protocol, and governance language recurred across them. Each of those is an announcement or a demonstration. None is a documented buyer deployment. Lorenzo’s corpus adds a telling detail. Eleven of the twelve agentic announcements he tracked at IBC used orchestration language. That is the layer Jon Roberts would least want standardised. He holds that view more loosely than the rest of his argument, and says nobody can yet read where orchestration settles, but writing a standard into the middle of something moving that fast “looks like a poor bet.”

I went in expecting to write that operators are short of engineers. My coded sample says otherwise, and the difference matters commercially.

Of nine distinct operator and rights-holder organizations I spoke with across the week, four raised engineering capacity as a constraint. None of the four described a headcount shortage to be solved by replacing people. What they described was constrained access to the capacity they already have. One network news division had reorganized engineering into skill silos and now generates excessive ticketing for any task that crosses them. One broadcaster’s editorial teams are blocked from experimenting with new formats because standing up resources requires a full engineering team, where in the SDI era they could just use the studio. One league’s production operation runs on roughly twenty engineers, about ten of them key, with the knowledge concentrated in a few senior people. One European operator described the patching and maintenance burden of self-built systems.

Their remedies followed that diagnosis. Self-service provisioning to bypass the ticket queue. Knowledge transfer from senior engineers to the wider team. Interfaces that let existing control-room operators describe what they want in plain language. A supported product to buy instead of a system to maintain. Every one of those augments a fixed engineering base. The league was explicit that adoption among its own engineers is a bigger hurdle than the technology.

In the one conversation where an industry-wide shortage of cloud and AI engineers was stated outright, the speaker was a vendor. The two operators in that room described silos and queues.

Nine organizations is a theme rather than a market, and people who spend forty minutes discussing architecture with me at a trade show are disproportionately people currently changing their architecture. With that said, the commercial implication is sharper than the shortage story. These operators are not buying replacements for engineers. They are trying to get more out of the ones they have and to let non-engineers act without a ticket. A product framed as headcount reduction is answering a question they did not ask.

Business models under pressure

Lorenzo on what the pressure did to how vendors sell.

The pressure on media technology budgets was visible in how vendors are restructuring their business models. This is a consequence of changing demand patterns, which, aside from cost-cutting, are moving in two major directions, largely dictated by risk-reduction: less commitment and more predictability.

When it comes to commitment, some vendors such as Calrec and Nugen Audio introduced shorter-term licensing options. Others like Amagi positioned their products as zero-CAPEX alternatives. Broadly, more vendors are lowering the barriers to entry because buyers are refusing long-term lock-in, particularly due to AI compressing development cycles.

The predictability side of the equation is more complicated, and in some ways cuts against the commitment argument. For example, Suitest introduced flat-rate pricing alongside its existing usage-based model, giving buyers a more predictable fixed cost option. And cloud repatriation fits the same pattern. Some vendors at IBC were positioning on-prem and hybrid deployments as a commercial advantage.

These moves from vendors highlight that buying patterns sit on a spectrum, particularly at a time of pressure like this one, and it is important for vendors to provide different models and entry points to fit different types of buyers.

That is also the honest verdict on the last of our August watch items. We said we would look for pricing that bills like infrastructure. It did not appear. What appeared was a move toward less commitment, and it fits the buyer squeeze better than our prediction did.

On the buy-side, the most underdiscussed trend at the show, relative to its importance, was in-house builds. There is rising evidence that media companies, particularly in some areas such as Northern Europe, are increasing investment in in-house build capabilities. Recent DPP research points in that direction.

This is a threat to the supply-side of media technology, and another pendulum shift. Some history can help here. The sector’s propensity to build technology in-house increased between 2015 and 2020, driven by the move to direct-to-consumer platforms. The pandemic (and a few build failures) brought cost discipline back, pushing the pendulum towards buy through to around 2025. Now, build is back, propelled by the rapid diffusion of vibe-coding technology.

I remember studying this trend before 2020, when a broadcaster told me their build investment was focused on two main areas: building connections between tools and customizing interfaces. Today, their efforts may be focused on the latter, while the vendors build the backend, and enable the connections through deeper partnerships.

Partnerships

Lorenzo and I reached this one independently, from opposite ends. Mine came from the buyer side. A component vendor needs an answer to how it helps a customer manage fewer relationships, and that answer often runs through an OEM, a marketplace, a cloud platform or an integrator, so the component arrives without a new contract attached. Lorenzo’s came from the vendors.

As I said in my IBC 2026 notes, I didn’t expect that partnerships would be such an important theme at the show, but they were. The industry has talked about this for years, but often only as a decoration, not real intent. This year was different, and since publishing my notes I have seen many others commenting on the same thing.

What were the drivers of this trend? Surely, lower buyer attendance led to more vendor-to-vendor conversations on the floor. But, more importantly, vendors are realizing that they can’t do everything themselves. This requires a mental shift from “I am the best at X” to “I can work with Y and Z to deliver X.” Some vendors made that shift, but others were still in aggressive pitch mode. Mentality is not easy to change.

In the future, I think trade shows may become platforms to discuss and showcase vendor connections more formally. For example, if two or more vendors have built a solution that is reproducible and applicable across customers, they can turn that into a published framework the market can evaluate, adopt and reconfigure. That is where the mental shift becomes hardest: accepting that you will not always be picked and being fine with it.

Even vendors more oriented towards end-to-end are recognizing the value of formalizing partner relationships. The Accedo and MediaKind joint streaming solution is a good example of this. It brings the evidence of a buyer already using it and proposes the same model to the rest of the marketplace.

What agents are allowed to touch

Security is not a gap in this industry’s thinking. The EBU architecture carries Security as a vertical across all six layers, with zero-trust principles, authentication, authorization and accounting, least privilege and segmentation. IBC made security, authenticity and provenance a conference theme in its own right. The vendors exposing agent interfaces are not silent on authorization either, with permissions inherited from existing rights controls and human oversight stated as design intent.

The narrower question is what none of the published material answers. When playout, routing, ad sales and system configuration are exposed to agents over a documented protocol, what is the identity model for the agent as distinct from the user it acts for? Who audits the supply chain of the skills and tools it loads? What does least privilege mean for something whose next action is generated rather than configured?

Lorenzo’s framing is that a conversation about agentic AI ends up being a conversation about what agents are allowed to touch, and the Hugging Face attack is the reason a buyer should treat that as urgent rather than theoretical.

Lorenzo’s corpus puts a number on it.

Security and governance language appeared in 36% of the IBC sample, up from 30% in the NAB Show 2026 measurement. And that growth ran alongside the rise in agentic and MCP language at IBC 2026, suggesting that vendors may be starting to connect the two conversations. Half of the agentic-related announcements tracked at IBC 2026 also used security language.

The security conversation is crucially linked to the point vs. end-to-end debate. More components generally mean more attack surfaces, which is something not to disregard in the era of AI cyberattacks. Therefore, security architecture should become an essential element of any partnership discussion.

The security conversation is part of a larger conversation about trust. AI is eroding our confidence in what is real, which makes proofs of authenticity more valuable than ever. This applies to different areas, from C2PA provenance standards to vendors and buyers discussing their technology roadmaps in person.

Consolidation, on hold

Consolidation does two things to a vendor at once. It shrinks the pool of buyers, and it is supposed to create the integration projects that replace some of that lost demand. This year the first happened on schedule and the second got stuck in court.

Lorenzo expected industry pressure to produce vendor M&A near the show and it did not arrive. On the buyer side the consolidation is real and it is stuck. Paramount Skydance’s proposed acquisition of Warner Bros. Discovery has not closed. Twelve state attorneys general sued to block it in July, a federal judge granted a temporary restraining order, Paramount agreed to a no-close and no-integration stipulation while the litigation proceeds, and an antitrust trial is set for March 2027. Nexstar’s acquisition of TEGNA did close, on 19 March 2026, at $6.2 billion with six station divestitures required, and an April preliminary injunction prohibits further integration subject to certain exceptions while an appeal proceeds. RTL’s acquisition of Sky Deutschland closed with a publicly stated annual synergy target.

So buyer-side consolidation is happening and litigated, while vendor-side consolidation is expected and absent. The rationalization of duplicated systems that consolidated companies typically pursue, and the cost takeout that justifies the deal multiple, keeps being deferred. If you sell into broadcast groups, the buyer-pool contraction is real and the specific integration projects you may be forecasting against are sitting in court.

The reset for NAB 2027

Three falsifiable items and one watch item, written before the evidence exists, with thresholds this time. Where a test depends on something becoming public, it says so, because operators keep production architecture private and a test that cannot see past that should not pretend otherwise.

One. Public proof that MXL reaches the top adoption rung. A publicly documented deployment in which media functions from at least two different vendors exchange live media over MXL in continuous operation for thirty consecutive days or more, with the operator named. Thirty days exceeds any single event and requires surviving a maintenance window. Disconfirmed, as a claim about the public record rather than about reality, if by NAB 2027 no such case has been published while at least three further vendors have announced implementations.

Two. A SOM implementation outside the consortium. At least one production newsroom publicly identified as publishing or consuming SOM 1.0 traffic where that organization is neither a champion nor a listed participant in the Smart Stories project. Disconfirmed if all publicly identified implementations twelve months on are consortium members. We will also record whether SOM has a permanent home by then. Jon Roberts says the repository is where 1.0 lives today and not where it settles, that a neutral home is being worked on, and that it should not be handed to any single organization, including the ones that built it.

Three. The pendulum. This is the timing question, made testable. Lorenzo coded every product announcement at both shows as point solution or stacked platform, and about two-thirds used stacked platform language at NAB 2026 and again at IBC 2026. The five-point band is a judgment we are making now about what counts as a real move. It is not derived statistically. If that share rises by five points or more by NAB 2027, the short-term tilt toward end-to-end is holding. If it falls by five points or more while open interchange adoption rises, the swing back has started. If it stays within five points of two-thirds, neither clock has moved, and we will say so.

And a watch item rather than a test. The switching cost gets published once. One named operator running an MXL-based workflow reporting what it cost to replace a media function with a different vendor’s. One case cannot establish whether open interchange reduces lock-in across the industry. It would establish that substitution happens in practice and give the first public data point on its cost. It sits outside the falsifiable three because its absence means the question is untested rather than answered.

Andy Beach is Chief Alchemist at Alchemy Creations and writes Engines of Change. The full IBC 2026 readout this piece draws on is available to Alchemy readers.

Lorenzo Zanni is an independent media technology analyst and founder of Sense the Frame, a data-driven newsletter tracking what is happening at the intersection of media, technology and storytelling. His full IBC 2026 notes are here.

Disclosure. Andy Beach is interim CTO of Fukui Labs and of Secret Level, a venture partner at Hallstone Ventures, and an advisor to swXtch.io, V-Nova and SMPTE among others. SMPTE is a champion of the Story Object Model. No client, advisee or sponsor is cited as evidence in this piece.

The reason he gave for it working is the part worth slowing down on. AMC had assets spread across locations, storage systems, and formats. Consolidating all of that into one environment before doing anything else was not a realistic starting point. So they went up instead. Phillips said they were inspired by the media lake concept, the idea of moving up in the architecture, and that this was why they got into production so fast, because they could move without needing to worry about what their current state was.

The constraint came first and the architecture answered it. Paul Dulac of AWS, on the same panel, described the mechanism from the vendor side, with S3 holding the assets as a single source of truth and PRISM handing references and pointers to the systems downstream, so editors could work immediately.

Nobody on that panel used the word reversibility, but that is what Phillips was describing. PRISM delivers access. What building it protected was something else, which is the ability to change things later without having to fix everything underneath first. The alternative was to consolidate first, agreeing on one storage system and one set of formats and one place where everything lives. At AMC’s scale that is not a step before the project. It is a larger project than the one it was supposed to enable, and nobody sees a benefit until it finishes.

What He Was Actually Protecting

The industry calls this lock-in, and the label misleads. Lock-in describes a condition imposed by a supplier, usually through a contract or a proprietary format, and it implies a villain. Most of what actually constrains a media operation has no villain in it.

AWS, of all parties, makes the cleaner argument. In its own guidance on the subject, and in an archived whitepaper written to address the concern directly, the company argues that lock-in is a misleading term and that what organizations really have are switching costs. Choose Java, migrate later to Node, and you pay something. The cost has existed throughout the history of the industry and it attaches to every commitment anyone has ever made.

That useful reframing is correct but it is also self-serving. It concedes that the property exists while implying it can be assessed. You know what you committed to, so you can reason about what leaving would cost.

Engineers have a name for the neighboring idea. Technical debt is the cost of a shortcut somebody took on purpose, and it comes with an author, a date, and usually a ticket. What we are describing has none of those, because nobody chose it and nobody deferred it.

Here is where the reframing breaks. A component diagram tells you what the pieces are and how the signal moves between them. It does not tell you what any of those pieces costs to replace. Those are different questions and only one of them can be drawn. An architect knows a single view never captures everything, and the discipline has whole practices built around that limitation. The problem is not that architects are naive. It is that the industry has started reading component diversity as evidence of reversibility, but it has never been a measure of that.

Everyone Is Hybrid Now

Which matters more this year than last, because the stacks have stopped looking like anyone’s category.

Avid arrived at IBC 2026 with Media Composer 2026.8, and the announcement leads with openness. Broadened collaboration choice. Third-party storage support. An OpenTimelineIO 2.0 adapter for round-trip fidelity with other tools. Expanded APIs. A Google Gemini panel and an agentic orchestration layer sitting across the workflow. Read the feature list and you can see a company distributing itself deliberately across other people’s components.

Read the same announcement for what moved between tiers and something else appears. Shared Projects is the capability that lets several editors work on the same material at the same time, which is how any facility of size actually runs. On Avid’s own Nexis storage, that capability comes down to the standard tier. The supported path for doing the same thing on qualified third-party storage sits in enterprise, at the top of the product line. Both changes are in the release that announced broadened choice, neither is hidden, and it is all on a price list. It drew enough criticism to make the trade headline.

Akta comes at it from the other direction. Its pitch is consolidation in its purest form, a unified AI-first platform meant to replace the collection of tools a media company uses to move, understand, repurpose and measure content, all inside one cloud. Then in June the platform went generally available on Oracle Cloud Infrastructure, and the reason given, as reported at the time, was lower-cost egress and native routing making full cloud production practical. Egress pricing is what a cloud provider charges to move data out, and it is one of the most commonly cited reasons that further cloud adoption stays difficult for media distributors. The company selling freedom from stitched-together tools changed the ground underneath itself over what it costs to leave.

Two companies, opposite strategies, and in both cases the fact that matters is on a pricing page rather than in an architecture description. Neither is doing anything underhanded. Both are behaving rationally. And an operator comparing them on capability would see none of it.

Below the line, a broadcaster replaces one component, spends the next ten months on everything attached to it, and shows you the cost that no architecture diagram will ever display.

Read more

What the major media companies are running instead is a hold: extremely limited commercial model use on anything that touches content, over IP infringement concerns. I think that is a defensible position, and I will come back to why. But a hold on commercial models governs your own people. It says nothing about your vendors, and most media companies are not the ones calling the model. They buy localization and QC from vendors, and the vendors are the ones holding the API keys. Unless somebody wrote the hold into the vendor contracts, and often they have not, the policy stops at the building.

So the companies being most careful about models internally can end up with the least visibility into where models are touching their content.

The supplier inside your supplier

The vendor’s model is the bigger exposure, for two reasons. Far more of a media company’s content work runs through vendors than through its own API keys, so that is where most of the model output is. And nothing in the company’s own AI governance can see it, because that governance was in most cases written for employees.

In supply chain terms the model is a tier-2 supplier sitting inside a tier-1 vendor. Supply chains have handled that arrangement for decades, in manufacturing and in media, through disclosure requirements, flow-down clauses, a right to audit, and treating undisclosed substitution as a breach.

Buying localization and QC is usually the right call, and a tier-2 dependency is ordinary. Every manufacturer and vendor has dozens or hundreds of them and manages those on the critical path closely. What makes this one different is that it is new enough to be under-documented, so it sits outside anyone’s watchlist.

How much trouble a tier-2 model causes comes down to one thing, which is whether the vendor told you about it or not. There are two cases and they are not the same problem.

In the first, the vendor sells “AI augmented localization” or “AI assisted QC” as a named service. The dependency is disclosed, so it is supposed to go into the agreement: which models, what happens when they change, acceptance criteria, where the content goes and whether it trains anything, who does the Article 50 disclosure work given that the duty lands on you as the deployer, and what the fallback is when their supplier pauses a run.

In the second, the vendor swaps model output in under the existing rate card to cut its own cost, and does not tell you. Anyone who has worked with multiple vendors knows this is a possibility. The media company takes on every risk above without knowing it, at the old price, with no clause covering any of it. Quality drift shows up as a QC miss against the vendor’s SLA while the real cause sits two tiers down. The guild terms make it worse. Under the new SAG-AFTRA contract, once a picture starts principal photography on or after July 1, 2027 under performer contracts signed after that date, a digital replica may not carry that performer’s voice into a foreign language without the performer’s consent. A dubbing supervisor will notice the lead’s own voice coming back in German. They are far less likely to notice it on the fourth-billed role in a smaller territory, or to catch a human dub actor’s track blended with a synthetic one, and the dub card will not tell them because it is built from what the vendor reports. Whatever the vendor did, the consent obligation sits with the Producer no matter who ran the model. The vendor is absorbing an upstream risk it functionally cannot manage, under a human-priced rate card, hoping it all works out. In any supply chain that manages its tiers, this is undisclosed substitution, and it is likely grounds for termination.

The first question is not which models we use. It is which models our vendors use, and whether they told us.

Why nobody owns this supplier

Putting terms on the vendor’s model means somebody has to write the terms, and that is the same problem the media companies have not solved for their own use of models. Why is almost nobody running those five controls? For two very understandable reasons.

The first is how the model arrived. A localization vendor gets onboarded into the operation once it is contracted, and someone in operations owns the relationship. A foundation model gets an API key and an engineering owner, and in most operations I have seen that is where it stops. The specific model dependency is usually created by an engineering lead picking the best performer that week, and the legal review comes after the choice is made, on the contract rather than the workflow. A monthly API bill does not trigger anyone’s vendor onboarding process, no matter how pricey it is.

The second reason is the hold I opened with, and the concern behind it is real and unresolved. Andersen v. Stability AI, the lead case on whether training is infringement, is still headed for a jury, and the MPA’s agreement with ByteDance governs what the models produce, not what they were trained on. As The Next Web put it, a model that “politely declines to draw Iron Man” does nothing to resolve whether the film libraries taught it what Iron Man looks like. Refusing to put an unqualified supplier on your most valuable work is what a careful buyer does.

It is also only half of a sourcing decision. Buying anything from outside takes two answers: where in the operation are we willing to use this, and which providers are approved to supply it, on what terms. A hold answers the first question. Very few companies have methodically answered the second. There is generally no published and approved list of model providers for creative work, and no standard set of terms to attach to them. I am not a lawyer; I am reading these as an operations person asking what happens to the delivery when a clause gets tested. No-model-training and content-use clauses are standard now and they do what they say. IP indemnification is worth less than it seems, and the exclusions are where to read. No indemnity anywhere covers the model being confidently wrong, which is the failure that actually costs an operation money.

A hold with no approved list behind it does not hold for long, because the demand for the work does not go away while the policy is unfinished. Two things happen instead. Teams under delivery pressure go around the hold, or they lean on vendors who will do the work for them, which is how the tier-2 problem above gets created. And the technology organization, which already holds the enterprise agreements for the non-content uses of AI, ends up at odds with, and across the table from, operations, when those two together are the only people who could build the list and write the terms.

The last point is happening in real time at different media companies. Tech is on a mission to deliver solutions identified by the C-Suite to drive down costs using AI. Operations is working to manage costs while protecting quality, talent, resourcing, vendor relationships, and a whole raft of other considerations. They don’t feel like Tech is consulting them, so they’re in a stand-off position while searching for their own tech solutions.

Nobody owns this supplier. Technology holds the contract, Operations holds the consequences, and neither is managing the supplier.

Where to start

The Human Layer post closed with a diagnostic and this one should too. Find the dependencies first, starting with the ones you cannot see, then put the five controls around them:

• Ask every vendor on the critical path which, if any, models they use in delivering your work, under whose terms, and whether it is in the contract. Expect a surprise. Then put disclosure, change notification, content-use terms, guild flow-down, and a fallback into the next amendment.

• Inventory every production workflow of your own that depends on a foundation model, name the model, and mark each dependency sole-source or dual-sourced. Most companies cannot pull this from a system, so ask.

• Write the acceptance criteria for each model-dependent workflow, expressed as a pass rate on a reference set, and spec out the ability to score any new version against it before you cut over.

• Qualify a substitute for anything sole-sourced on the critical path, using your own titles and your own delivery specs.

• If you want to enable the use of AI for creative work, finish the sourcing policy behind any hold on model use: an approved supplier list, no-model-training and content-use clauses, indemnification read with its exclusions in mind, guild flow-down, and an on-premises or open-weights option where content cannot leave.

• When you commit a delivery date on model-dependent work, build in the fact that the model may not behave next month the way it does today.

• Add “supplier” to your miss attribution categories.

The discipline already exists

Andy’s closing line in August was that none of the four parties was fighting about model quality. They were fighting about the conditions attached to a model. Sizeable media companies already run vendor management functions that know how to qualify a supplier, write flow-down terms, and plan for a supplier failing. That function generally has not been pointed at the model in the middle of the localization workflow, or at the vendor quietly using one, because neither looked like a supplier when it arrived. NIST named the category in its Generative AI Profile, Value Chain and Component Integration, and wrote most of the checklist above as suggested contract terms.

A foundation model is a supplier. It does not arrive labeled as one, and it does not send the kind of invoice that gets a vendor manager’s attention, so nobody handed it to the people who do that work. Those people are already on the payroll and they know how to do this. Give them the relationship.

Sources and further reading:

• Andy Beach, Four Ways to Reach the Model (Aug 21, 2026), What the Badge Buys (Aug 14, 2026), The Other Ledger (Aug 28, 2026).

• The Next Web, ByteDance signs a copyright pact with Hollywood’s MPA to rein in its Seedance AI video tool.

• Andersen v. Stability AI, Joint Case Management Statement, N.D. Cal., Jan 13, 2026. The lead case on whether training a model on copyrighted work is infringement; not yet tried as of this writing.

• SAG-AFTRA, 2026 Theatrical-Television Memorandum of Agreement (also at AMPTP), read in full. MOA Section 39 (Artificial Intelligence) is a pointer, not the provision: it modifies Sections 64 and 64.1 of the General Provisions per Attachment D, and the operative text sits there. Dubbing consent is Section 64, Paragraph C: “Effective on motion pictures commencing principal photography on or after July 1, 2027, for which a performer’s contract is entered into on or after July 1, 2027, should the Producer wish to use the Digital Replica of a performer for purposes of adjusting the voice of the performer to a foreign language, it must comply with the consent, but not the compensation, requirements of Paragraph A. above.” Both conditions apply, and compensation is expressly carved out. Replica security is Paragraph G; transfer of rights is Paragraph H, under which a transfer does not relieve the Producer of its consent, use and payment obligations unless the transferee commits in writing to comply and to Union arbitration, and the Union does not disapprove the transferee’s financial responsibility within 21 days; the articulable business reason for scanning is Paragraph I. The training-license notice is Section 64.1, Paragraph D, triggered only where a Producer licenses covered performances to a non-affiliated third party for compensation to the Producer, for training a public-facing commercially available GAI system, and requiring written notice plus a meeting on request rather than consent. Ratified June 4, 2026; term through June 30, 2030. See also the Summary of Tentative Agreement, whose item lettering does not match the MOA’s.

• WGA, Summary of the 2026 WGA MBA, Section 4, Artificial Intelligence. Ratified April 24, 2026; term May 2, 2026 to May 1, 2030. 2023 Article 72 provisions carried forward per Know Your Rights: Artificial Intelligence. Both the SAG-AFTRA and WGA agreements require written notice to the union when a producer licenses covered work to an outside party, for compensation, to train a commercially available model.

• IP indemnification terms, read from each vendor’s published enterprise terms. The pattern across the five: all will defend against a third-party claim on the output, with conditions, and three of the five cover the model only when it is working alone, which is not how anyone runs a model in production. OpenAI Business Terms Section 13.1 (effective Jan 1, 2026; exclusions include combination with non-OpenAI products, modification, Customer Content, Customer Applications); Anthropic Commercial Terms Section K.1 (effective June 17, 2025; exclusions include Inputs, knowing infringement, modification, combination with non-Anthropic technology); Google Cloud Generative AI Indemnified Services (index, July 20, 2026) and the 2023 announcement; Microsoft Customer Copyright Commitment (Mar 20, 2026; required mitigations, including safety filters on and documented testing producible on a claim); AWS Service Terms Section 50.10, “Defense of Claims and Indemnity for Indemnified Generative AI Services,” read in full: the indemnity runs only to an enumerated list of Indemnified Generative AI Services, all of them Amazon’s own (the Nova and Titan families, HealthScribe, Personalize, Amazon Q, Transform, Lex, Kiro and others). Third-party models served through Bedrock are not on that list and are “Third-Party Content” under Section 50.12.1, governed by the model provider’s own terms. Section 50.10.2 carries seven exclusions, among them infringing inputs, interference with or failure to enable available filters and tools, and any fine-tuning or customization but for which the infringement would not have occurred. There is no combination-with-other-technology exclusion, and the indemnity is uncapped. Google Cloud Service Specific Terms, AI/ML section, read in full: Google runs two indemnities. The output indemnity applies only to “an unmodified Generated Output from a Generative AI Indemnified Service using only Google Pre-Trained Model(s), a Modified Google Model, or a Customer Adapter Model used with a Google Pre-Trained Model,” which is a combination-and-modification limit written as a positive scope condition, with five further exclusions covering known infringement, disabling filters or source citations, use after notice from a rightsholder, trademark claims, and missing rights in customer data used to customize. The separate training-data indemnity covers allegations that Google’s use of training data to create a Google Pre-Trained Model infringes, and is the only indemnity among the five that reaches the training question rather than the output.

• NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). Risk category “Value Chain and Component Integration”; suggested actions GV-6.1-004 (contracts and SLAs specifying content ownership, usage rights, quality standards, provenance), GV-6.1-005 (supplier risk assessment framework), GV-6.1-009 (procurement processes for embedded GAI), GV-6.2-003 (third-party incident response and notification), GV-6.2-004 (continuous monitoring of third-party GAI in deployment), GV-6.2-007 (liability assignment and GAI system changes over time, including fine-tuning, drift and decay).

• EU AI Act Article 50: paragraph 2 places machine-readable marking of synthetic output on providers; paragraph 4 places deepfake disclosure on deployers. Applicable from 2 August 2026 per Article 113.

On the surface, a model does not look like a supplier, it looks like software. It is licensed, it has an API, and it sits in the stack next to the transcoder and the asset management system. But what makes something a supplier is not how it is purchased. A supplier is an outside party whose work goes into your deliverable, on terms the supplier controls. The test turns on what the work is, not on what the technology is. The same model is software when it summarizes your ticket queue and a supplier when its output is deliverable subtitles or dubbing. Software is bought against a specification. You know what it is supposed to do, and when it changes there should be a version and a note saying what changed, even when the change was not your idea. Unlike true software, a model can produce the creative work product itself, a translation, a QC decision, a first-pass conform, and nothing you bought says what that output is supposed to look like. The company providing the model can change what comes out, or the conditions attached to it, without your consent. Regardless of the name on the invoice, that looks like a supplier relationship.

Because the models look like software, they are typically handled like software. Many media companies did the legal work at very significant expense, and from what I have seen of these deals, the IP, no-model-training, and indemnification terms were negotiated in detail. Every mastering, localization, and finishing vendor gets the same set of things before it receives any work: acceptance criteria, a service level, change notification, a qualified alternate, and a named owner in operations. The models get none of those terms or controls. The model makers will not volunteer for them either, and that’s really the crux of the problem.

In August, four news stories showed four ways the conditions attached to a model can change without the customer’s consent. None were about model quality, and none would have been prevented by a better license. Conditions changing mid-contract is a common situation in supplier relationships, which is why contracts and supply chain management exist.

The legal work on the models has been done but nobody has done the vendor management.

Four disruptions, one category

Andy’s “Four Ways to Reach the Model” read the August stories as four parties finding four ways to impose terms on a model. From the buyer’s side, each is a supplier disruption any supply chain manager would recognize.

The Motion Picture Association and ByteDance MOU from August 17 created a bunch of new IP guardrails across five different in-use services. The MPA, a downstream customer with leverage, forced a spec change on a supplier. Output filters, likeness blocking, and provenance credentials now arrive in the product whether you asked for them or not.

Twitch‘s “training for generative AI” setting was switched on for every account. Its chief product officer explained the default plainly: “If this was opt-in, nobody would opt in.” A default setting is a contract term that was not negotiated.

The transparency obligations in Article 50 of the EU AI Act became enforceable on August 2. Plenty of boards heard “delayed” out of the Digital Omnibus and assumed all of it had moved; it had not. Here the compliance clock is set by a regulator rather than by either party to the contract.

OpenAI paused part of its frontier training, hitting the brakes on Astra due to safety and security concerns. They stopped work on a model for their own reasons and on their own schedule, and every customer with a Q4 roadmap built on that release found out from the press.

Four parties changed the conditions of supply on a model inside three weeks without consulting a single customer first.

The vendor that will not sign

If a model is a supplier, the obvious question is why nobody has simply put it under a services agreement and been done with it. The answer is that Anthropic, OpenAI, and Google will not sign one. No output quality SLA, no change approval, no exit assistance, no acceptance criteria. Fair enough, from where they sit, but that also means the discipline and controls have to live on your side.

Physical supply chains have a category for this supplier: the standard finished part bought from a company much larger than you, for example a common Texas Instruments SoC chip used in countless consumer products. That supplier ships to its own roadmap, issues change notices on its own schedule, ends the life of the part when it decides to, and will not sign a buyer-specific quality agreement unless the buyer is enormous. It is not hostile, it is indifferent. Mid-size manufacturers manage it anyway, through incoming inspection, change notice and end-of-life reviews, and a plan for exit. The model vendor fits the category almost exactly.

The contract with the model maker gives you data-use terms and a conditional indemnity, and not much else. What it does not give you is a specification. A model is probabilistic: it produces variation by design, it can be wrong, it often is, and the terms say so. That part is not unusual. A resistor comes with a tolerance band, and a localization vendor quotes a QC pass rate rather than a guarantee. Both of those suppliers tell you the shape of the variation you are buying, and you plan the operation around it. Acceptance sampling exists for exactly that reason. The model vendor ships you the variation and no numbers to go with it.

Because you cannot contract for the control, you have to build it.

What you build instead

You build the same five things every mastering and localization vendor already gets before it touches your work. The difference is who produces them. A normal vendor gives you the service level and sends you the change notice; with this supplier you measure the level and detect the change yourself. Your supply chain people know how to do all five.

First, an acceptance criterion, written against real work and not a demo. Second, a tolerance stated as a rate on a reference set, which is the same method I use to set service levels in media supply chains, taken off a measured baseline instead of declared. Third, change detection, so a new version gets scored before you cut over to it. Fourth, a qualified alternate you keep current and feed work to periodically to keep it up and tested. And fifth, a process owner, because audit, review, respond, and fix only happens if it is somebody’s job.

Models are commoditizing in price, not in behavior, and the qualified alternate is the item that gets skipped. Technicolor showed what that costs from the vendor side: a supplier with more than a hundred and ten years of history was very suddenly liquidated in early 2025. The companies with a qualified alternate were better able to manage than those that had to rapidly figure out what to do next.

Not one of those five requires the vendor to agree to anything.

Where the failure lands

Say none of this gets done. There is no acceptance criterion, no measured tolerance, and no qualified alternate. This is what that looks like in an operation, and why nobody will call it a supplier problem.

Take one dependency chain. A localization workflow tuned over months to one model’s behavior, a QC step that leans on one vendor’s detection, an orchestration layer built against one provider’s API. The model changes, outside your control. Output drifts and the post-editors absorb it by hand until they cannot, the QC step starts passing what it used to catch, and the miss surfaces as late deliveries and rework. The dashboard charges it, correctly, to operations. Unfortunately this is a common issue, the “upstream system that changed its schema on a Friday and broke stuff” problem. What is new is that the party upstream is a foundation model vendor under no obligation to do things differently.

In the performance measurement work I have done for media supply chains, every missed delivery gets assigned a cause: materials, capacity, intake assumptions, downstream queueing, vendor. That assignment is what tells you where to spend money to fix it. The list of causes is set in advance, and none of the lists I have seen has a line for a supplier changing its own product. So the miss gets recorded against capacity, and the next quarter somebody funds more capacity to solve a problem that was never about capacity.

Until somebody adds the model to the list of things that can go wrong, its misses get charged to operations.

None of this requires new technology or a new function. It requires calling the model a supplier and handing it to the people who already manage suppliers. Almost nobody has, and the reasons are more defensible than they look.

Stay tuned this week for Part 2: why nobody owns this supplier, and why the models your vendors are already running on your content are the bigger exposure.

What connects these four is the layer. All of them install proof beneath the content rather than policing it above. What does not connect them is who gets to sign. Apple and the C2PA coalition have picked the same problem and opposite answers about whether the verifier is an open standard or a single company. Katzenberg and the labels have picked the same problem from opposite ends of the table, one building the licensed model and the other supplying it. The pattern is agreement on the architecture and a fight over the key. That fight is the story, and it started this week.

Disclosure: I have done consulting work on C2PA implementation, including with Secret Level. That work informs the view here and readers should weigh it accordingly.

Apple builds proof of origin and skips the standard

Source: Nieman Lab, September 9, 2026

At its iPhone event, Apple introduced Apple Reference Image for the iPhone 18 Pro and Pro Max. With the feature enabled, a photo is effectively saved twice, once as a normal editable file and once as a reference image carrying a unique ID in its metadata that attests it was captured by a physical iPhone camera rather than generated. Capture telemetry, including hardware IDs and sensor signatures, goes to Apple’s Private Cloud Compute, which produces a signature tying the reference image to that specific phone. Apple says it never sees the photo itself and that the process only runs on images the user chooses to sign. Apple calls the result a digital negative.

Nieman Lab’s Joshua Benton draws the line that matters. Nikon and Canon have adopted C2PA Content Credentials for the same job. Google supports C2PA on Pixel. Apple has not signed on to Content Credentials, and this launch is the clearest statement yet that it believes it has a better approach.

Why it matters

The provenance layer just got its first major defection, and it came from the company that ships more cameras than anyone. C2PA’s premise is that a credential should be checkable by any conformant party, a broadcaster, a platform, a court. Apple’s premise is that the verifier is Apple. The reference image is a receipt the photographer can produce later, but the entity that validates the receipt is the same one that issued it. That is a closed loop with a privacy rationale, and the privacy rationale is not nothing. It is also a bet that trust in Apple substitutes for trust in a standard.

For newsrooms and broadcasters this splits the workflow in two. A photo from a Nikon carries a credential the ingest pipeline can read. A photo from an iPhone 18 Pro carries a token only Apple can resolve. Both are proof of origin. Only one of them is portable. The industry spent five years converging on a shared verification layer, and the device most journalists actually carry now sits outside it.

The broadcast floor files provenance under security

Source: NewscastStudio, August 28, 2026

IBC opens today at the RAI, and the preview coverage tells you what the buyers are asking about. NewscastStudio’s survey of production network executives puts cybersecurity, system reliability, and content authenticity in the same sentence, as workflows move to IP transport, cloud platforms, and distributed infrastructure. Net Insight’s Andreas Eriksson names the handoffs between facilities, networks, and partner organizations as the highest operational risk, the points where responsibility and visibility are divided. In the same conversation, executives point to Content Credentials based on C2PA as the tool for verifying provenance and tracking modifications across a piece of content’s lifecycle.

The concrete work sits in the IBC Accelerator programme. The Stamping Your Content project, championed by BBC and ITN, built an open-source tool that inserts C2PA metadata at the moment of publishing and a companion tool to verify it, with fingerprinting and watermarking used to recover credentials that get stripped in transit. The stated next steps are packaging it as a plugin for the tools newsrooms already use and extending stamping to live broadcast.

Why it matters

Provenance has changed departments. A year ago it lived with standards and policy people, a compliance item with a conference panel attached. This week it is being discussed by the people who run transport networks, in the vocabulary of zero-trust access and handoff risk. When a broadcaster describes content authenticity as an operational risk at the handoff, it has stopped treating provenance as a label and started treating it as part of the plumbing, alongside encryption and access control.

The Accelerator work makes that shift executable. Stamping at publish, verifying on ingest, and recovering stripped credentials via fingerprint are the three functions a pipeline needs before provenance can be assumed rather than asserted. Read against Apple’s launch, this is the open-standard side of the same layer being built by the institutions that have to exchange media with each other every day. They cannot use a verifier that only one vendor can run. That is the practical reason C2PA exists, and it is the reason Apple’s choice will be felt on this show floor first.

Katzenberg’s answer to Seedance is a model he can license

Source: The Information, September 9, 2026

The Information reports that Jeffrey Katzenberg is planning a startup with Bill Peebles, the former head of OpenAI’s Sora app, and Sujay Jaswa, the former Dropbox CFO, to train AI video models for filmmakers. The company is unnamed and is in talks with investors including Andreessen Horowitz on a large funding round. Katzenberg’s WndrCo already sits on the board of Reactor, a real-time AI video company that raised $59 million in May.

The context is February. ByteDance’s Seedance 2.0 produced a viral clip of AI versions of Tom Cruise and Brad Pitt within days of launch. The MPA said the model had engaged in unauthorized use of US copyrighted works on a massive scale and demanded ByteDance cease. Disney sent a cease-and-desist. ByteDance pledged stronger safeguards. Seven months later, the industry’s most experienced studio operator is building the thing the studios were threatening to sue.

Why it matters

The move is not defensive in the way the headlines suggest. A Hollywood insider building an AI video model is not a hedge against Seedance. It is a decision about where the legal exposure gets resolved. Suing ByteDance polices the output. Owning a model trained on material you are entitled to use settles the input, which is where the courts have been slowest and the studios have been most exposed. The lawsuit route requires proving harm after generation. The ownership route makes the training set the proof.

Peebles is the tell. Sora’s head leaving OpenAI to build video models with a studio veteran says the frontier labs are not the natural owners of the licensed layer. The labs own the general capability. The people who own the rights and the relationships can build the version filmmakers are permitted to use, and that permitted version is now the product. The a16z conversation is capital arriving to fund a rights-native model rather than a general one, which is a different investment thesis than the one that funded the labs.

Suno retires its old models and ships the licensed ones

Source: Variety, September 9, 2026

Suno launched its v6 family, three models built with Warner Music Group, BMG, and the distributor Believe, and said it will retire every earlier model as v6 rolls out. The flagship v6 and the more experimental v6-wild go to Pro and Premier subscribers; v6-mini is free. Chief product officer Jack Brody describes the models as trained from the ground up on licensed data from partners plus user data, with revenue sharing for artists and labels, though Suno has declined to give specifics on the training set or the payment structures.

The sequencing is the story. Warner settled its copyright suit against Suno in November 2025 on terms that required Suno to launch more advanced licensed models and retire the existing ones. BMG, which never sued, signed a licensing deal in August. Believe’s partnership, announced the day before launch, opens TuneCore distribution to tracks generated on the licensed models. Universal and Sony remain in litigation.

Why it matters

This is the cleanest version of the pattern because the settlement wrote the product roadmap. The models trained on Warner’s catalog without permission are being switched off. The models replacing them were built with Warner. The lawsuit did not end in damages; it ended in a supply agreement, and the supplier is the plaintiff. Warner and BMG are now on the issuing side of the proof. Their license is the credential that makes a v6 output usable, and they get paid each time it is issued.

Katzenberg is building this from the studio side. The labels arrived at it from the catalog side. Both end at the same place, a model whose legitimacy is established at training rather than litigated at output. The gaps are real: Suno will not say what is in the data, the artist payment terms are undefined, and the two largest labels are still in court. But the direction of the settlement flow is set. The remaining suits are negotiating for a seat on the issuing side, not for the model to go away.

The layer is settled. The key is not.

Take the four together and the shared move is easy to see. The proof of what something is and where it came from is being installed before distribution, in the camera, in the publish pipeline, in the training set. Enforcement after the fact is not going away, but it is being demoted from primary control to backstop. That is the shift that will still be true in a year.

The disagreement is over who issues the proof, and it splits along a line the industry should recognize. On one side, an open credential that any party can verify: C2PA in the broadcast pipeline, and, in the rights layer, licenses that travel with the model and can be audited by whoever needs to. On the other, a verifier that is also the issuer: Apple resolving its own reference tokens, a single studio-backed model whose permission structure lives inside the company that trained it. Both produce proof. They produce different kinds of dependency.

Why now is straightforward. Seedance made output policing look hopeless in February. Warner’s settlement made the licensed-model route look profitable in November. IBC puts the people who have to exchange media across organizations in one building this week, and they have concluded that provenance belongs with security. Apple, reading the same landscape, chose to own its slice of the layer rather than join it. Everyone decided the receipt goes upstream. They did not decide it should be readable by anyone else.

Closing note

The next twelve months will be spent finding out whether the proof layer is a shared utility or a set of walled gardens with overlapping vocabulary. The early evidence points both ways at once, which is normal for a layer this young. Broadcasters will push for interoperability because their business is handoffs. Device makers and model owners will push for control because their business is the platform.

Watch three things. Whether Apple’s reference image can be read by any conformant C2PA validator without going through Apple. Whether the Accelerator’s stamping tools show up as defaults inside the newsroom software vendors on this show floor, rather than as plugins someone has to remember to install. And whether Universal and Sony settle with Suno on terms that look like Warner’s, which would confirm that the issuing side of the proof is where the labels intend to live.

If the answer to the first is no, the provenance layer has forked at the sensor. If the answer to the second is yes, the open side has a distribution advantage the closed side will have to buy. Either way, the question the industry is now arguing over is not whether the proof exists. It is who you have to trust to read it.

A year later, TheWrap reported that the partnership had run into complications. Two of them, and they are not the same problem. The first was scale. People familiar with the work told Roger Cheng and Jeremy Fuster that a model trained on Lionsgate’s catalog alone could not do what the partnership had promised, and that the same would be true of Disney’s. The second was rights. Whether Lionsgate held full authority over its own films, and whether ancillary claims belonging to performers, writers or directors would have to be settled first, was reported as an open question.

Nobody at Lionsgate conceded that the second problem stopped anything. The studio said its AI work was continuing on several fronts and that the Runway arrangement was never exclusive, and the partnership did in fact expand in June of 2026, with Lionsgate taking an equity interest and the work running across previsualization, storyboarding and finished frames. So this is not a story about a deal that died. It is a story about which part of the deal turned out to be hard, and the hard part was not the modeling.

Ray Seilie, an attorney at Kinsella Holley Iser Kump Steinsapir, put the underlying condition plainly, observing that each production will have a variety of interested rights holders. Everyone in the business already knows this. What is new is being asked to say who they all are before anyone has requested anything.

The thin volume of studio training deals has plenty of obvious explanations. Litigation posture, valuation disagreement, an unwillingness to arm a competitor. All of them are probably operating. Underneath them sits a mechanical problem that has nothing to do with strategy, and it is one the business has been carrying, quietly and successfully, for about seventy years.

The scale finding does not compete with that problem. If one studio’s library is too small to be useful on its own, the corpus that works spans more than one owner, and each owner added is another rights estate that has to answer the same question.

What the Archive Actually Sells

Anyone who has licensed archive footage already knows the shape of it.

Sony Pictures Stock Footage will license you a clip. What it will not do is tell you the clip is clear. The license covers the footage. Identifying the performers who appear in it, determining whether they were guild members, working out which residuals and reuse fees the intended use triggers, and paying them, all of that lands on the licensee.

This is not a loophole and nobody involved treats it as a scandal. It is how the business has functioned for decades, and there is an entire professional layer, researchers and business affairs departments and guild administrators, whose work is exactly this. The International Documentary Association has described the practice plainly, noting that archives often cannot themselves say whether a particular use triggers payments, or whether the people who worked on the material were union members at all.

What makes the arrangement function is the unit. Somebody wants twelve seconds for one production. That request generates a specific question with a findable answer. The clearance cost is irritating and finite, it gets paid, the deal closes.

The rights state of the asset is never resolved in advance. It gets assembled on demand, around a use somebody has actually named.

That is the system. It has never been a database. It has always been a research process, and it has always started when someone asks.

This Has Happened Before

Now hand that architecture to a buyer who wants the whole catalog.

Corpus licensing inverts the sequence. It asks for the answer before anyone requests a clip, across every asset at once, without a particular downstream use to make the question finite. The clearance layer that has served the business for seventy years has no procedure for this, because it was never asked to produce one. Everything it knows how to do, it does one request at a time.

It would be convenient to call this an AI problem. It isn’t. The pattern is older.

Google scanned tens of millions of books without seeking permission title by title, and the Second Circuit eventually held the scanning and the snippet display to be fair use. The litigation ran for years, and the resolution, when it came, arrived through fair use rather than through clearance, which is to say the question of how you would clear tens of millions of books was never answered. It was routed around.

The Copyright Office took the administrative version of the question seriously and proposed an extended collective licensing pilot for mass digitization in 2015. More than eighty comments were filed. By 2017 the Office concluded there was not sufficient consensus to proceed, and the pilot went nowhere.

Streaming produced the other kind of answer. As programs moved into subscription streaming, the guild agreements supplied collectively defined residual formulas for the new reuse market. Nobody reopened every performer agreement to negotiate the new market separately. The collective rule made the new use administrable across the whole body of covered work.

So the pattern is old. A use appears whose commercially interesting scale exceeds the granularity at which rights were originally administered, and the industry has to find some way out of the one-at-a-time system. Sometimes collective bargaining does it. Sometimes collective licensing gets proposed and dies in committee. Sometimes a party proceeds without asking and lets a court sort it out afterward.

One thing about the current version is genuinely different, and it is worth being precise about it. Google’s route never required constructing a grant. A studio selling training access cannot make that move. Selling requires affirmatively stating what is being conveyed, which means somebody has to determine, in advance and in writing, what sits inside the thing being sold.

Read more

Read more

On April 25, 2014, Lucasfilm changed what its own past was allowed to do. The statement was careful, and it did two things at once that most people read as a single action. It named George Lucas’s six films and The Clone Wars as the immovable objects to which all other tales must align, and it said that everything else, thirty-five years of novels and comics and games written to be consistent with each other and with the films, would no longer bind. The stated reasons were creative freedom for the sequel films and the preservation of surprise. Lucas, the statement said, had never been beholden to that material anyway. Then, in the same breath, it said the material was not being discarded. It would stay in print under a new banner, and the creators of new Star Wars would have full access to it.

So the record survived intact and lost its authority in a single act. Nothing was burned. Every fact about Thrawn and Mara Jade and the Yuuzhan Vong remained exactly where it had been the day before, available to anyone, and none of it could tell a new writer what was true. Having a history and being governed by it turned out to be two different conditions, and Lucasfilm had just pulled them apart in public, at a scale nobody could miss. Who had the right to make that call has an answer, and it is the boring one. The company that owns Star Wars, acting through the people running it. Thirty-five years of decisions were still sitting there, readable, ignorable, and available. What does a later decision-maker do with a body of work like that?

Start with the record everyone thinks they know. Lucasfilm kept a continuity database called the Holocron, and for years it was described, by fans and by the press, as the closest thing a story world had to a court, the place where canon was adjudicated, where every ship and planet and character was filed at a level of authority, where the franchise remembered what it had ruled. Its keeper was a man named Leland Chee, and Chee spent a good part of two decades explaining, patiently and on the record, what it actually was.

Believe it or not, the Holocron was a FileMaker database, and it worked. Built inside Lucas Licensing in early 2000, it grew to tens of thousands of entries covering characters, planets, vehicles, weapons, and organizations, and new additions were checked against it while Licensing enforced consistency through it in practical ways. It carried a top-level category for material originating with Lucas himself, which is authority represented in a field. But Chee’s own account is narrower than that reputation. He said the database’s job was to track continuity for Lucas Licensing and, to some degree, the company’s online arm, and that to his knowledge it was only rarely used for production. The famous canon levels, the tiers fans argued about for years, were a construct built for the database, and Chee said most of the people using it skipped that field and went to the source material instead. Lucas had never given directives about how anything was entered. Chee had no direct contact with him about continuity at all, and what he saw were notes from other people’s meetings. And in 2005, asked on the StarWars.com forums where the real canon lived, Chee described a version in George’s head at any given moment that nobody else knew, and said it was constantly changing.

That is a more interesting instrument than either the reputation or the debunking suggests. The Holocron could tell you where a thing sat in the pecking order. It could not tell you why it sat there. Those are different kinds of information, and the difference matters more than it sounds. A database field can hold a rank, this is top-tier, that is second-tier, and so on. What it cannot hold is the argument that puts it there. And the argument is exactly what a later decision-maker needs, because ranks only help when things disagree, and two entries at the same rank that contradict each other leave you with nothing to go on. That gap stayed invisible for as long as it did because the people who needed the argument could go find the person who had it.

Which leaves the question I am trying to address here. The record could say what ranked above what. It could not say why, and the why is what lets a past decision guide a future one rather than just sit there being true. So where was it kept? Anyone who has inherited a body of prior decisions knows the shape of that problem without needing the vocabulary. Tom Brevoort, who has edited at Marvel since 1989 and took over the X-Men line in 2023 at the end of a five-year era, was asked in his newsletter whether inheriting the X-Men’s continuity was a gift or a burden. He said it was always a little of both, that there are always stories he wishes had gone differently, and that past a certain point you accept that they happened and play the ball where it lies. Play, not leave. The prior decisions are on the field and you work from where they landed. Which opens the harder version of the question. What does it mean to play the ball where it lies when the ball came from a continuity that has been declared not to bind?

Lucasfilm answered that in public in July 2016.

At Star Wars Celebration in London, Dave Filoni announced that Grand Admiral Thrawn was coming to Rebels. Thrawn was the villain of Timothy Zahn’s Heir to the Empire, published in 1991, the book that effectively launched the modern Expanded Universe, and he had been Legends since the reset. Filoni and Kiri Hart, Lucasfilm’s head of development, had pitched the return, and reporting from the panel says the idea had been under consideration for some time. Zahn was writing a new canon novel about the character. A major figure was being lifted out of the record that did not bind and given standing once again.

Then Filoni distinguished the addition. Zahn’s novels had come with a specific piece of hardware attached to the character. The ysalamiri were lizard-like creatures that generated a bubble in which the Force did not function, and Thrawn used them to strip Jedi of their advantage, which is how an ordinary military officer with no Force sensitivity stayed credible as the antagonist of a Star Wars story. Take the creatures out and the problem of how Thrawn survives contact with a Jedi has to be solved some other way. Fans at the panel understood exactly what was at stake and asked whether the ysalamiri were coming too. Filoni said no, and gave the reason. Years earlier, during The Clone Wars, he had discussed the creatures with Lucas, and they had concluded that a living thing removed from the Force contradicted Lucas’s conception of what the Force was. Zahn later acknowledged that the ysalamiri had been one of the few notes Lucas ever gave him on the original books. Asked about Mara Jade, Zahn’s other great creation, Filoni said not likely.

Look at what that ruling required. Thrawn and the ysalamiri came from the same books and, after 2014, carried the same claim on the new canon, which was none unless Lucasfilm chose to restore it. A later authority went back into that body and did not treat the package as indivisible. It took the character and refused the device, and the ground for the refusal was a rationale that predated the ruling by years. That last part is where a skeptical reader will push, and it is worth being exact. Lucas’s conception of the Force was not merely an older creator’s opinion. It carried higher standing inside the canon regime Lucasfilm itself had reaffirmed two years earlier, when it named his films and his television series as the objects everything else must align to. Filoni was applying a principle that still governed the question, and the record of why the ysalamiri were excluded is that principle, remembered from a conversation and stated in public.

Lawyers have a word for what Filoni did there. When a judge is handed an old case that looks like it ought to decide the matter in front of them, and the judge explains why it does not apply here, or why part of it applies and part of it does not, that move is called distinguishing. The old case is not overturned. It stays on the books, still good law for everyone else. It simply stops reaching this particular set of facts. Courts can do this reliably because they have machinery built for it, rules about which decisions bind which judges, a habit of writing down reasons, a citation system that lets anyone find those reasons later, and a doctrine governing when an old rule may be set aside. Story worlds have almost none of that. What Lucasfilm had in July 2016 was one person carrying the reasoning that let him tell two things from the same book apart. Yesterday’s rulings become tomorrow’s constraints, and they did here, but notice where the reasons had been kept. Not in the Holocron, but in Filoni’s memory of a conversation with Lucas. Which also means it could have gone the other way, and nothing in the record would have objected.

The people making Rebels then did something quiet with what they had refused. Look at Thrawn’s office in the show and there are sculptures of lizards framing the desk. They are the ysalamiri, present, recognizable, kept. They cannot repel the Force, because in this version of Star Wars no such creature exists. So the show is doing two things with one prop. It is acknowledging where Thrawn came from, and for the part of the audience that knows the books, it is displaying the exact piece of that inheritance the new canon refused. The creature is in the frame. It has no power over anything. Cited without authority, preserved without standing.

So that is one company’s answer to what a later decision-maker does with a body of work that no longer binds. Lucasfilm goes back in and picks, restoring one piece at a time and leaving the rest where it lies. The obvious next question is whether anyone else works this way, and the short answer is that the other places worth looking have each landed somewhere different.

Read more

The part worth sitting with is not the number. It is that nobody at the newspaper was ever measured on municipal borrowing costs. No editor ran a report on the spread between what the county would h…

Read more

The interesting part is not that all four are the same fight, because they aren’t. Two are about what comes out of a model, one is about what goes in, and one is about whether the training finishes at all. What connects them is that each party could impose terms, and each had a completely different reason it was able to.

Hollywood signs its first tr…

Read more

And studio work rarely produces just one final, which proves the point rather than weakening it. The airline cut, the broadcast edit, the version with the flashing lights removed, the localized master for another territory. Each exists because someone has to make a deliberate editorial decision for a specific distribution context, and each leaves behind its own fixed artifact carrying its own standing. The pile of versions is a family of rulings. The line between an authorized variant and everything else stays crisp, because crossing it requires an act.

What the Artifact Was Doing

A finished master settles which edit exists. The sentence sounds trivial until you notice how much work it does. Between deliberate editorial acts, nothing about the work drifts. The version in the vault this morning is the version from last night, and everyone downstream builds on that assumption without checking. Localization teams, syndication packagers, rights lawyers, fans arguing on forums, all of them inherit a stable referent.

Canon itself was always bigger than any one of those artifacts. A story universe’s canon is the rules and the recall, everything held to have happened inside the world and the principles governing what can happen next. No single object contains it. The finished episode captures a piece. The script captures another. The story bible states rules the footage never shows, production notes record why a choice was made, and a source novel or an earlier film sits underneath an adaptation as precedent from another jurisdiction. The archive holds the evidence for canon rather than canon itself, the material a ruling can point to, and some of it outranks the rest. The artifact’s quiet trick was making that evidence cheap to keep and hard to accidentally revise.

That evidence base is why continuity work stayed periodic. The people responsible for a story world’s coherence did not sit in permanent session. They convened when something new was proposed, ruled on it, and dispersed, trusting the artifacts to hold everything steady in the meantime. The work was exception handling, and the medium kept the exceptions rare.

Franchise scale is where the old regime showed its limits. Lucasfilm’s Story Group is the standing body that aligns Star Wars storytelling across film, television, publishing, games, and everything else the franchise touches. Even with fully fixed artifacts, enough of them accumulating across enough platforms already requires standing human governance. Fixed artifacts never eliminated judgment. They bound the workload. The Story Group can operate as a review function rather than a permanent tribunal because its inputs arrive as discrete, deliberate releases. A film. A novel. A game expansion. Each one is a proposal that can be examined before it enters the canon.

Inputs arriving as deliberate acts is the condition everything below turns on.

Read more

Now, of course, none of these companies coordinated with one another. But when you see this news side by side, it becomes easier to see the moves companies feel forced into just to look credible on AI. The mark, the penalty for carrying it, and the exit from the system all shipped in the same week. Labeling was sold as a disclosure measure. It is functioning as a pricing mechanism, and the price is reach.

Anthropic starts marking at the model

Source: The Register, August 11, 2026

Anthropic confirmed that new Claude models w…

Read more

The thing was OpenEXR, the high dynamic range image format ILM built for its own pipeline and opened to the world in 2003, which grew into a widely used standard in professional visual effects, the container for the frames studios care most about. The people who built and carried it, Florian Kainz, Rod Bogart, Drew Hess, Piotr Stanczyk, Ed Hanway, moved on one by one, and by the later handoffs the work had settled into maintenance routine rather than active development. The studio’s engineering attention had long since shifted to bigger problems, because the format simply worked. Cary Phillips, the R&D supervisor it eventually landed on, describes inheriting it without a deep background in image science and with hardly any hands-on time in the modern tooling. And then a security vulnerability was reported in the library, and the team holding it had, in his words, so little experience with security issues that it was “ill-equipped to respond.”

The destination already existed, because someone else had been building one. Rob Bredow, ILM’s CTO, was serving with Phillips on the Academy’s Science and Technology Council, where an effort was brewing to safeguard the open software the whole industry runs on, and Bredow drove that effort into what became the Academy Software Foundation, a neutral home governed by the companies that depend on the code. In 2019, OpenEXR moved in. Phillips gave the public reason in the announcement, and it is a clear-eyed sentence about an institution’s own limits: “there’s a natural life cycle to software projects.” Architects leave. Expertise spreads. The ecosystem outgrows any single owner’s ability to keep up. Inside the studio, by his account, nobody argued against it. Quite the opposite. The one real condition was the counterparty. ILM would have hesitated to hand stewardship to an entity whose agenda might drift from its interests, and the Foundation, governed by seasoned professionals from the studios and vendors that live on the format, cleared the bar. Phillips signed on to chair the project’s technical committee for three reasons. It fit his council role. Two of the original developers are close friends, and he did not want to watch their creation wither. And it sounded like fun.

Read that sequence against everything this series has described, and the strangeness sharpens in both directions. This is not a rescue story. Nothing had been lost, no attic was involved, and the studio was still running its most important imagery through the format every day. A working institution looked at the people leaving, concluded its own custody had a horizon, and arranged for responsibility to outlive it, the first actor in four essays of accidents and defaults to behave like a preservation system on purpose. And honesty requires the other direction too. The Foundation was already being built, but the decision to move OpenEXR into it followed the scare, which readers of the second piece in this series will recognize as the oldest law in this machinery. Nothing moves until something forces it. What the move solved is also worth being precise about. A foundation can receive what a keeper is able to hand over, the code, the roadmap, the authority to decide. The judgment that leaves in the elevator is a different problem, and it has to be solved again, one successor at a time.

Where the survivors came from

The last piece in this series ended by saying that what survives tends to survive somewhere small, close, and tended. That described where survivors turn up. It did not explain why. What they actually share is that something with a different mandate was standing there when the first one stopped. The first three pieces studied the powers of persistence while they operated, the authority that decides what carries forward, the dependency that vetoes change, the judgment that knows when the rules stop applying. This one is about what happens when responsibility has to move.

The seam

Start with a number from the people whose entire job is keeping things. Researchers went through the global registry of research data repositories, the institutions built explicitly to preserve material for the long term, and identified 191 that had shut down. The median age of a repository at shutdown was twelve years. Fewer than half migrated their holdings to a successor before closing, and nearly half showed no sign of using any data-rescue strategy at all. Look at what that denominator means. These were organizations whose founding purpose was persistence, staffed by people trained in it, and their mandates still ended young, and in most cases the ending had no plan. Even the Digital Preservation Network, built to replicate holdings across organizations and geographies, with succession written into its design, wound itself down in 2018. Redundancy inside a mandate does not survive the end of the mandate.

The preservation field knows this about itself, which is why its standards read the way they do. A trusted repository is required to have a succession plan: a named successor, transferable material, enough context that a stranger could reconstitute the collection, the rights to hand it over, the capacity to receive it, and a trigger that fires while there is still something to hand over. The discipline learned, the hard way, that survival is not a property a system has. It is an arrangement between the system and whatever comes after it.

Now look at media. Almost nothing in it was built to that standard, because almost nothing in it was built as a preservation system. The studio pipeline exists to ship this year’s slate. The rights database exists to close this quarter’s deals. The playout chain exists to make air tonight. The format exists because somebody’s renderer needed it. Persistence, everywhere in this industry, is a side effect of systems designed for something else, and side effects do not get succession plans. Which means the moment the preservation field plans for arrives in media unplanned, and usually unannounced. Responsibility rarely gets put down cleanly. It erodes through a merger. It diffuses across a reorganization until no desk owns it. It evaporates when a team is cut, and the discovery that it evaporated comes years later, when someone reaches for the thing and finds that nobody has been carrying it. A great deal of media succession is somebody realizing, long after the fact, that nobody has the baton.

Read more

That assumption is starting to expire, and one of the moments that convinced me came at the Portland Video Tech Meetup this June. Oracle’s Gaurab R. Upadhaya, Matthew Young, and Ai Lou demoed a live soccer feed with an AI analyzer subscribed to the video like any other viewer. The analyzer watched the match frame by frame, detected goals, clipped replays, and published its findings back into the same session as an event track. Every subscriber received the goal events through the same publication-and-subscription architecture carrying the video. The machine held a subscription and a publication in the same session.

The obvious objectio…

Read more

Several years later, the railroad called him back. Not to run the old network. To help kill it. The microwave and copper era was giving way to cellular, and the railroad needed someone to teach a new generation of engineers how to decommission systems they had never seen built. Which order to cut. What each piece protected. What could be turned off before what, without leaving a stretch of track blind. What began as a consultation became years of contract work. The organization that had absorbed his retirement without a ripple discovered, a decade on, that it could not safely take his own work apart without him.

In theory, none of this should have been necessary. Railroads document. This is an industry whose documentation standards most media companies would envy: standards bodies, schematics, signal maintainer manuals, standard operating procedures, rulebooks with revision histories older than television. All the paper existed. The paper was not enough. They needed a human being.

The first two pieces in this series were about what happens to the things, the assets that survive on disk and go functionally dark, and the systems that cannot be removed because everything stands on them. This one is about what happens to the knowing, which was never in the things at all.

The paper and the person

Stay trackside, because the gap has a precise location. The schematics showed every circuit in the detection network. Nowhere did they say which order was safe to cut. Nowhere did they explain why one site was wired against the standard, or which readings the old hands had learned to distrust, or which workaround was protecting which failure. The paper told you what was true. The person told you when to trust it.

That is the split this piece is about, and every industry runs on it. Media’s version is familiar to anyone who has ever reached into an archive: the master that is technically present and practically wrong, the clause that appears in every contract for a reason nobody wrote down, the naming convention that encodes a workflow three reorganizations dead. Call the first kind stored knowledge and the second kind held knowledge. Stored knowledge is the record. Held knowledge is the judgment about the record, and decision history exists everywhere as workflow exhaust and nowhere as durable provenance. The record is flat. The judgment about it never made it in.

Read more

In none of these did anyone dispute the model file. Its contents were settled, downloadable, benchmarked. What people fought over was everything attached to it.

The industry puts its name to the position that weights are not the moat

Source: The New Stack, July 27, 2026

A letter urging policymakers to avoid premature restrictions on open-weight models went out on July 24 with roughly two dozen signatories. It now carries more than 130. Google, OpenAI, and Amazon signed after the initial release. So did Comcast, SpaceX, Intel, AMD, Databricks, Runway, Black Forest Labs, Resemble AI, and Bria. Anthropic is the only major American lab still outside it.

The lett…

Read more