Andy Beach's Engines of Change

Andy Beach’s Engines of Change

Publicly Available

The EU AI Act makes AI labs publish training data summaries for copyright holders. Answering a photographer about her own work is still optional.

Andy Beach's avatar
Andy Beach
Nov 15, 2026
∙ Paid

Obligation, Part I of IV. What each player in the AI chain owes the people whose work goes into it, from the labs to the app makers, the users and the government.


Any company that brings a new general-purpose AI model to the European market, open-source models included, now has to publish a summary of what it trained on. The law says why. The summary is there so people with a stake in that data, copyright holders by name, can protect their rights. So picture a photographer opening Google’s summary for its Gemma 4 models and finding the field that asks which websites were scraped. The answer reads “publicly available websites across educational, government, legal, and research sectors.” She still can’t tell whether her site is one of them.

Now suppose her site falls below the cutoff for the websites a lab has to name. Then a fully completed form could leave it out entirely. Europe saw that case coming. The Commission’s own guidance expects her to ask the lab directly whether it scraped and used material from her site for this model, and it encourages the lab to answer “in good faith and on a voluntary basis.” Publishing the summary is mandatory, while answering her is a courtesy. The rule was written for her, and the lab doesn’t have to answer her. That gap is what this series is about, and it starts here, at the top of the chain.

Read the Form

Most of these forms answer her question by category. The way to see it is to read them as she would, asking what each field lets her do next.

Google’s form gives her sectors, broad kinds of sites like education and government. The field asks for domain names, the actual sites, and gets back a description of the kinds of sites. The rest of that field covers media types, languages, how Google’s crawlers behave and that they respect robots.txt. Some of that may help someone. None of it tells a photographer where to look.

OpenAI’s summary for GPT-5.5 gets her one step further. Under large public datasets it names Common Crawl, a public archive of the web she can search herself. That gives her a lead. Its own domain field, a few pages later, falls back on categories too, academic repositories, legal and government resources, community sites. So she has a place to start and no confirmation that her site is in it.

This isn’t one company’s habit. When the French AI publication ActuIA checked a sample of 24 summaries from seven providers in September, 21 included the domain section and not one of those 21 named a site. The other three left the section out. Eleven of the 24 came from a single provider with a lot of repeated wording, so I’d treat it as a pattern and not a census. The pattern is still clear.

Then there is the form that shows what the alternative looks like. The summary for Apertus, the open model from the Swiss AI Initiative, links its main training sets by name and version number, FineWeb-Edu v1.0.0, FineWeb-2 v2.0.1, The Stack v1.2. It answers no to the question of whether the provider ran crawlers of its own. It links the filtering scripts on GitHub and says the data can be reproduced with them. A reader holding that form can open a named dataset at the version that was used, search it and rerun the filtering to see what survived. Nobody holding the Gemma form can do any of that.

Independent grading lines up with what the forms show. The AI Accountability Lab at Trinity College Dublin scores each summary on how much it discloses and on how usable it is for the people it’s meant to serve. As of November 9 it had found 80 summaries, graded 64 and flagged another 46 models as possibly missing one. GPT-5.5 scored C+ for transparency and D+ for usefulness. Apertus scored A and A+. The grades measure potential usefulness to a range of stakeholders. They don’t test whether any photographer has found her work. They confirm what the forms already let you feel.

Where the Law Stops

It would be unfair to read every gap in these forms as a lab dodging a clear rule. Much of the coarseness is by design.

Article 53 of the AI Act requires a sufficiently detailed summary built on a Commission template. Recital 107 explains who it’s for, parties with a legitimate interest including copyright holders, so they can exercise and enforce their rights. The same recital sets the limit. The summary should be comprehensive in scope and not technically detailed, with trade secrets taken into account. The Commission’s explanatory notice says it plainly. The template doesn’t require disclosure of the specific data and works used to train the model.

The template does ask for something specific about the web. Labs are supposed to list the top 10 percent of the domains they scraped, ranked by how much content came from each, with a lower bar for smaller firms. That ranking is by volume, so a small photo archive scraped heavily could outrank a famous site. The 10 percent is also a floor, and a lab is free to list more. But the floor is where the photographer’s case gets hard. If her site sits below it, a fully compliant form could leave her out without anyone breaking a rule.

The Commission anticipated exactly that. In paragraph 16 of the notice it recommends that, for domains not listed, labs let people with a legitimate interest ask whether content they made available there was scraped and used for training. It adds that this doesn’t affect the other remedies copyright holders have under EU law. Then it uses the words that decide the piece. Labs should do this “in good faith and on a voluntary basis.”

I want to be fair to that choice. A guest piece here later this month will argue that the system is evolving and that simpler answers may be the practical ones, and the notice gives that case real support. Regulators saw the photographer’s problem, built a way to help her that doesn’t require publishing whole training sets, and left room for labs to do better. That’s a reasonable design. My disagreement is about one word in it.

User's avatar

Continue reading this post for free, courtesy of Andy Beach.

Or purchase a paid subscription.
© 2026 Andy Beach · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for writing, podcasting, and media founders.