Selling to Labs
Article 53 in one page for data owners
You are not regulated by this. Your buyer is — and that is why it now governs your sales process. The EU AI Act places obligations on providers of general-purpose AI models, two of which reach directly into the data supply chain: a public summary of the content used for training, and a copyright policy that honours machine-readable rights reservations. A buyer who must publish a description of its training data cannot accept a corpus you cannot describe.
The two obligations that reach you
| Obligation on the provider | What it becomes at your end |
|---|---|
| Publish a sufficiently detailed public summary of the content used to train the model, following the template the AI Office provides. | Your corpus needs a describable identity: what it is, where it came from, how it was collected, roughly how large, and under what arrangement it was obtained. "A proprietary dataset" is not a summary line. |
| Maintain a copyright policy, including identifying and respecting reservations of rights expressed by machine-readable means. | Anything you gathered from third-party sources must be shown to have honoured opt-out signals at the time of collection — not retroactively. |
| Documentation to downstream providers and, on request, to the authority. | Your representations end up inside a document your buyer may have to hand a regulator. Expect them to be drafted as warranties, not as marketing. |
There are lighter obligations for models released under genuinely free and open-source licences, and additional obligations for models deemed to carry systemic risk. Neither changes the seller-side conclusion: the buyers with the biggest budgets are the ones with the heaviest disclosure duty.
Why this changed the market rather than just the paperwork
Before a public-summary duty existed, a buyer's rights review was an internal risk judgement — and internal risk judgements can be made with indemnities and a shrug. A published summary is different in kind: it is a durable, public, checkable statement. Anyone can read it and compare it against what they know about a corpus.
That converts provenance from a negotiating point into a gate. It is the single clearest reason rights carries the heaviest weight in our rubric, and why deals now die at counsel that would have closed three years ago.
What to prepare, in the order it gets asked for
- A one-paragraph description of the corpus a buyer could paste into a public summary. Domain, content type, collection method, time span, approximate scale, and the legal basis under which you hold it. Write it yourself; if you leave it to the buyer, they will write something conservative and it will constrain the deal.
- Source-by-source accounting. Which portion came from your own operations, which from users, which from third parties, which from public sources. Percentages, not adjectives.
- Evidence that rights reservations were honoured for anything collected from the open web: which signals you checked (robots directives, TDM opt-out declarations, site terms), when you checked them, and what you excluded as a result. The timing matters — honouring an opt-out today does not cure a collection that ignored it in 2023.
- A per-record provenance annotation, so that any subset can be described and filtered rather than argued over. This is the artefact that turns a week of counsel time into a query; see chain of title for the schema.
- Your position on future changes. Buyers increasingly ask what happens if a source withdraws consent or a rights holder objects after training. Have an answer: notification, replacement, or an agreed carve-out.
Two mistakes to avoid
Assuming it does not apply because you are not in the EU. The obligation follows the buyer placing a model on the EU market, and the large labs are placing models on the EU market. Your corpus is inside their compliance perimeter regardless of where you are.
Treating the summary as the buyer's problem. It is the buyer's obligation and your bottleneck. The sellers who close fastest are the ones who hand over a description that requires no editing — because at that point you have removed work from the person deciding whether to buy.
The practical test: could your buyer describe this corpus accurately, in public, without asking you a single follow-up question? If not, that gap is what is standing between you and a signature.
This is a seller-side orientation to obligations that fall on someone else, written at a level of generality that survives amendment. It is not legal advice, deadlines and template requirements have moved before, and your buyer's counsel will have their own reading. Get your own advice on anything specific.
Read next: How to run a trajectory capture programme — building an asset with clean rights from the first day instead of retrofitting them.