Suppose a photographer receives an email confirming that her work was used to train an AI model. The training is already finished. The answer reaches her after her chance to prevent that use has passed. A useful accountability process would give her a way to act while the decision can still be changed.
People should not have to investigate by hand to catch up with an automated process whose consequences may already be hard to reverse. That is the operating problem underneath the debate about AI training transparency. A company can collect material, process it and use it while the person who created it is still trying to find out whether any of that happened.
In Publicly Available, Andy Beach examines the distance between a disclosure requirement and an answer a creator can use. A photographer can read a training summary and still have no way to establish whether her work was included. Andy argues for records that support meaningful, model-specific answers, including an honest account of where the evidence ends.
His disagreement centers on the voluntary response described in paragraph 16 of the Commission’s explanatory notice. That paragraph also explicitly permits providers to disclose more than the required minimum. This piece makes the case for volunteering now and for what doing it well would require. Labs should use that room to build records, notices and controls that let affected people influence a use before it becomes hard to change. Enforceable obligations remain necessary; companies can build the mechanisms while those obligations develop.
Broad collection and rapid development serve understandable goals. Negotiating separately with millions of people would create enormous friction, and conflicting instructions and incomplete supplier records make individual decisions harder. Standardized controls give companies a way to handle recurring decisions consistently. A useful system honors standing instructions, makes retained records searchable, notifies people when a decision needs their attention and connects their response to controls that can change the affected use.
Make the records usable
A practical first step would cover new material a company collects directly. Preserve the source, the usage instructions encountered and the decisions made about collection and training, and give people a public place to search those records by domain or page address.
Starting with direct collection is a deliberate limit. It gives a company a defined set of records and decisions to make inspectable without presenting a domain search as a complete account of how a work was used. It addresses the narrow question first while preserving Andy’s warning that an answer about one website can leave other routes unexplored.
The result should explain what the company can establish: whether material was collected, excluded, waiting to be used or used in a particular model’s training. The information should come in a structured format so people can use their own tools to interpret it.
This is a lookup of retained records. Working out whether an image was used by examining a trained model is a different technical problem. Image matching across acquired collections could become an added capability as the service develops, with its limits clearly described.
Historical collections and third-party datasets need further work. A photograph may arrive through a licensed archive or a copy published elsewhere, so the system needs to explain which routes its records cover. A missing record must stay distinguishable from confirmation that a work was never used.
Personal data, security-sensitive details and legitimate confidential information need protection. Where public disclosure is limited, an affected person should still have a defined route to learn what the company can confirm about their work. Commercial sensitivity deserves the same discipline: a company should explain why withholding a particular record is necessary and how an affected person can still get a meaningful answer.
A company can publish what it can reliably search and expand coverage from there. Useful access should begin before the entire history has been rebuilt. A searchable register still leaves the creator responsible for checking, though.
Make intervention operational
Standing instructions should do the first part of the work. Companies should identify and respect applicable machine-readable rights reservations, publish how they interpret those signals and apply them during covered collection and before training. Website owners should not have to register the same refusal again.
A website owner should also be able to verify control of a domain, manage preferences and subscribe to notices. Those notices could confirm that an exclusion was applied, describe a proposed change in use or flag an unresolved conflict. Domain verification establishes who can manage collection preferences for a website. It does not settle ownership of every work published there, so a creator challenging the use of work hosted elsewhere needs a separate review path.
For future collection or training covered by the service, a notice that needs a decision should arrive early enough to influence the use and offer a direct way to respond. Silence should not be presented as affirmative permission. Where a use requires an affirmative decision, the workflow should remain on hold until that decision is made.
The goal is to reduce the owner’s monitoring burden. Making someone refuse the same use repeatedly would simply move that burden into an inbox.
The test is what happens after the person acts. A refusal button that opens an ordinary support ticket can leave the same timing problem in place. When a verified website owner excludes material from a covered future use, the instruction should reach the relevant controls without a prolonged exchange. The owner should then receive a status showing whether it has been applied, is pending or needs review.
For the initial service, intervention means stopping covered collection or a future training use. It does not promise to reverse training already completed. That limit makes early notice and working controls more important.
A disputed ownership claim or a concern about a third-party collection may need review. Those cases need a process that can hold the affected material when the concern warrants it. A public report should not give anyone automatic authority to stop an entire training operation. It should give a legitimate concern a practical route to the people and systems that can change the affected use.
This proposed service goes beyond the legal rights available in every jurisdiction today. Its initial scope should stay simple: new direct collection, searchable source records, standing preferences and notices when something needs attention. A service covering every AI provider would need wider participation and shared infrastructure. The first company does not need universal agreement to make its own conduct easier to see and influence.
A company should know how it handles conflicting instructions and what happens when the evidence is insufficient. Those rules should be public enough for affected people to understand the decisions being made about their material. The IETF’s AI Preferences working group is developing ways to express AI usage preferences and attach them to content. Its charter leaves enforcement, auditing and training transparency outside its scope. Companies still need to connect the signal to a decision and keep evidence of what followed.
Build for imperfect information
Missing and inaccurate information will remain part of the work. A source may be misidentified, a restriction recorded incorrectly or a supplier’s inventory incomplete. Public visibility gives affected people a way to identify those errors, and effective intervention gives the company a way to correct them before they compound.
That creates a more credible basis for accepting bounded uncertainty. People can see what is unknown, challenge the record and observe what changes. They have something concrete against which to judge the company’s conduct.
The consequences should depend on what is missing or wrong. An absent descriptive field may have little effect on whether a use can proceed. An unresolved restriction may be decisive. Transparency cannot supply missing permission, and publishing a gap does not discharge the responsibility to address it.
The record should show which rule applied, whether the instruction took effect and what remains unresolved. Legal teams can make those responses workable by setting evidence requirements, resolving recurring ambiguities and giving operational teams clear authority to hold material, correct a record or prevent reuse.
A problem with a supplier’s inventory should lead to corrected records and a better delivery specification. A valid complaint that exposes a policy gap should change the workflow. Each resolved case should leave the company better equipped to handle the next one.
This also changes who carries the investigative burden. A creator may have no relationship with the supplier that delivered her photograph. The lab selected that supplier and accepted the collection, so it should use that relationship to help establish what happened. Naming the supplier is useful. Leaving the creator to coordinate the rest of the investigation can preserve the imbalance disclosure was supposed to address.
Leadership decides how much of that imbalance the company is prepared to correct. Companies have reasons to protect their competitive position and control the pace of their decisions, but meaningful accountability requires sharing some of that control. A credible concern may interrupt a schedule or require a decision to be reconsidered. A fair process leaves room for the company to be inconvenienced, challenged and occasionally delayed while the outcome can still be changed.
Give people a reason to trust you
The distance between this standard and current practice may make the proposal sound optimistic. Public records, working controls and useful notices take investment. Parts of the foundation are already forming, though.
California’s AB 2013 requires covered developers to publish high-level training-data documentation, with an initial deadline of January 1, 2026. Google has announced that it aims to support a shared website AI-training preference signal in 2027. These address disclosure and an instruction a machine can read. The next operating responsibility is to connect them to dependable records, working controls and evidence of the result.
The customer and the creator are often the same person. A small business publishes product descriptions, a hobbyist maintains a blog and a customer may give original material directly to an AI service. Whether a given contribution enters training depends on the source, service, permissions and settings, but the roles overlap. A company can deliver a useful product while leaving its own customers unsure how their contributions are treated. A durable strategy has to address both experiences.
Trust becomes commercially meaningful when people have concrete reasons to participate. Useful records, understandable rules and controls that work give them evidence for that decision. AI companies should be ready to compete on capability, reliability, service and the quality of their operations. Making their treatment of other people’s material inspectable belongs in that quality.
The rest of Andy’s series will examine responsibility further along the chain. Labs already have a decision in front of them: how much practical influence will they give affected people while a use can still be changed? They can start with direct collection, expand the records they can reliably search and make exclusions work within that scope.
When collection is automated, accountability cannot depend on someone discovering the use by hand and pursuing an answer after the consequential work is done. It has to arrive while there is still something it can change.
Rebecca Louise Avery is a media operations leader and advisor to pre-Series A startups. She also writes the Integration Therapy blog, which explores the relationship between operations, culture, and leadership.



