Our corpus was scraped from public pages. Where does that leave us?
On solid ground, mostly. US courts have held that automated collection of publicly visible data is not unauthorised access. The exposure isn't the scraping — it's whether you can show what each source's machine-readable reservation said on the day you collected it. Under the EU AI Act that's the difference between honouring an opt-out and asserting that you did.
What proves it: a per-source collection log with timestamp and the reservation state as it stood at collection. If you have that, you're in good shape. If you'd be reconstructing it now, you're writing a story rather than producing a record.
Some of it came from behind a login. How bad is that?
It's a different question entirely — not access law, contract law. The moment a credential was used, the terms attached to that account apply, and courts have been clear that creating accounts to reach otherwise-unavailable data is where liability lands.
What proves it: authentication state per source — whether any credential, account or session was used to reach it. Most corpora never recorded this, which means most teams can't answer it in either direction. Not knowing is itself a finding.
We bought the dataset. Isn't that the vendor's problem?
Only if the vendor can produce records. Your position is exactly as strong as their documentation, and you inherit their gaps without inheriting their knowledge of where those gaps are. In diligence the question arrives at you, not at them.
What proves it: the licensing chain back to the original owner — not a warranty that one exists. A vendor who can show the chain is worth more than one who indemnifies you against it.
We licensed directly from rights holders. Are we clear?
Almost certainly, and it's the strongest position available. One thing worth confirming: whether the licence covers training specifically or only use. Agreements written before 2023 are frequently silent on it, and silence isn't permission once someone asks.
What proves it: the licence text with the training grant identified. If it's there, you're finished — this isn't a problem you have.
It's synthetic. Doesn't that avoid the whole question?
It moves the question rather than removing it. California's AB 2013 names synthetic data explicitly and requires disclosure of its use. What matters becomes what generated it, under what licence that model was released, and what went into the generation.
What proves it: declared lineage — generating model, its licence, and the inputs. Generated on permissively licensed weights with the licence archived, you can state it plainly. Generated through a commercial API, read that provider's terms on output use before you write the disclosure.
We acquired it. Nobody here built it.
The hardest position, and a common one. The people who made the collection decisions are gone, and whatever records exist weren't kept for this purpose. Acquirers have started pricing this risk directly rather than assuming it away.
What proves it: whatever survives — and an honest audit of what's missing is worth more than an assurance that nothing is. Documented uncertainty is defensible. Undocumented confidence isn't.
Honestly, it's mixed, and some of it we can't trace.
That describes most real corpora, and saying so isn't a weakness. The regimes now in force ask for a documented summary, not a perfect one. The failure mode isn't having gaps — it's not knowing where they are when someone asks.
What proves it: an inventory that separates the traceable from the untraceable. It's a smaller job than people expect, and it has to happen before any of the others matter.
None of this requires talking to us. If you can produce the records above, you can answer the questions coming your way. If you can't, that's worth knowing before someone else finds out.
The records named above are the ones our manifest format carries, and the verifier that checks them is open source — you can run it against a delivery from anyone, including us. tgs-verify on GitHub.