Does your AI agency contract cover training-data liability?

The filing is the hook. The clause is the point.

On September 17–18, 2026, TechCrunch and Engadget reported on new material in the copyright lawsuit The New York Times brought against OpenAI and Microsoft. The vendors' own internal record is quoted in the plaintiffs' brief: per TechCrunch, Brent Hecht, Microsoft's director of Applied Science, called the practice “an astonishing theft of unprecedented proportions” and “the largest theft of labor in human history”; a later internal presentation of his described a “doom loop” that would “hurt the performance of our models and the entire web at the same time.”

Vendors' own internal documents, quoted in a plaintiff's brief, describe the training-data supply chain as a threat to the publishers who supply it. It does not make your business a defendant: no ruling is reported, judges have so far leaned toward the AI companies on fair use (TechCrunch), and Microsoft says "Microsoft's position is set out in its court filings, which explain why these transformative uses are consistent with copyright law and why Copilot is not a substitute for publishers' journalism," while, per TechCrunch, "OpenAI and Microsoft did not return requests for comment." TechCrunch's caveat holds for every quote above: the material is one side's brief — “much of the new information comes from The Times’ own brief, not the underlying exhibits, which remain sealed” — and the quotes arrive “presented without their original context.”

What that changes is narrower than a headline: provenance is now a written question, and six clauses decide whether you carry that risk or hand it back.

1. Training-data provenance representations

Look for: a representation that names the upstream model, its version and what it was trained on — attached to the master agreement, dated, with a named owner on the agency side.

Push back on: a provenance answer of publicly available data (scraped corpora are publicly available; TechCrunch's account of the same material includes “deliberate efforts to strip copyright notices from training data”), a claim that lives only on a marketing page or model card, and we don't train models offered in place of describing the model you are renting. Anything the agency will say on a sales call, it should put in a schedule.

2. IP indemnification: scope, then the carve-outs that swallow it

Look for: indemnity for third-party intellectual-property claims arising from the deliverables and from model output, covering defence costs as well as awards, with a duty to notify you of claims and a defined defence process.

Push back on: carve-outs for claims arising from training data or from third-party models — that is the exact risk on the table; a carve-out leaving only your own modifications covered; an indemnity capped or excluded as indirect loss; and an indemnity the agency cannot fund because its own upstream provider disclaims the same liability. Ask which provider's indemnity flows down to your contract, and put the flow-down in writing.

3. The limitation-of-liability cap

Look for: how the cap is computed (fees in the last 12 months versus fees paid in total), whether the IP indemnity sits inside or outside it, and whether there is a supercap for IP claims.

Push back on: a cap equal to fees paid with the indemnity inside it — on a small engagement the cap is smaller than the cost of defending a claim — plus a one-sided cap and indirect or consequential damages exclusions drafted wide enough to capture defence costs. An indemnity is only as good as the cap it survives.

4. Data-use and model-training opt-outs

Look for: whether your inputs and outputs train anything, an opt-out you can exercise in writing with no training by default, retention and deletion windows that name backups and sub-processors, and what the upstream provider does with prompts the product sends it.

Push back on: we may use data to improve the services boilerplate, an opt-out buried in a help-centre article rather than the agreement, and deletion promises with no attestation you can keep on file.

5. Audit and disclosure rights

Look for: a documented provenance statement or attestation at signing and at renewal; named sub-processors with change notice; a duty to tell you when a rightsholder or regulator raises a claim touching the model you depend on, with a stated clock rather than within a reasonable time; and security evidence you can show a client or an insurer.

Push back on: audit rights that only trigger after a breach, and a refusal to name the model provider.

6. Termination and transition if the model is later found infringing

Look for: a termination right on an IP claim touching the model, and the familiar ladder — procure a licence, replace, modify, refund — with refund as the floor rather than a courtesy. Add transition assistance: export of your data, prompts, configuration and any fine-tuned artifacts built for you.

Push back on: a licence promise with no fallback, no transition window and no committed export of the pieces you paid for.

Red flags in an agency proposal

• The words training data appear nowhere in the proposal or the master agreement.
• Provenance is answered with public data or no copyright issues.
• The IP indemnity excludes claims arising from the model or its training data.
• The liability cap equals fees paid, with the indemnity inside the cap.
• No named upstream model provider, and no named sub-processors.
• No duty to notify you about claims, and no transition or exit plan.
• The agency's upstream terms disclaim the very liability the agency is promising you.

What this is not

This is information, not legal advice: nothing above is drafted for your jurisdiction, your deal or your risk appetite, and indemnity scope, carve-outs and caps are exactly the language to put in front of your own counsel before signature. No court has ruled on the copyright question itself.

For the wider clause baseline, see AI Agency Contract Tips: What to Look For, and for the vendor side of the same filings, is your AI vendor's training data a legal risk to you?

Sources

• [1] TechCrunch — Rebecca Bellan, “Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal” (September 17, 2026). techcrunch.com — the 17 September 2026 report · Wayback archive
• [2] Engadget — Steve Dent, “Microsoft executive called OpenAI's web scraping the ‘largest theft of labor in human history’” (September 18, 2026). engadget.com — the 18 September 2026 report · Wayback archive

Accuracy note: Quotes are attributed to the outlet that reported them, on the dates shown; filing language is quoted as it appears in the September 17–18, 2026 reporting. The material is one side's brief with the exhibits sealed — it is litigation risk disclosure, not a finding of liability, and no ruling is reported. This page is informational only and not legal advice.