Someone else’s lawsuit, your contracts: how the Times case reached regulated AI buyers

The week

On December 27, 2023, The New York Times sued OpenAI and Microsoft in federal court in Manhattan. It alleged that the companies had used millions of its articles without permission to train the models behind ChatGPT and Microsoft’s Copilot products, and that the models could reproduce Times content closely. The complaint sought damages and the destruction of models and training data built with Times works. It joined a growing set of copyright suits against AI developers brought by authors, artists and other rights holders during 2023. In the months before, several vendors had announced commitments to defend enterprise customers against copyright claims arising from their AI products, including Microsoft in September and OpenAI in November.

Copyright litigation between publishers and AI developers looks remote from the work of a bank, hospital or agency. They do not train frontier models and are unlikely to be named as defendants. They are, however, heavy buyers of AI products whose legal footing the suit challenged, and three of their own choices set how much of that dispute could land on them. Each corresponds to a different channel through which someone else’s case reaches a non-party, and together they cover the routes: the contractual channel, meaning the indemnity a buyer relies on; the publication channel, meaning the content it puts out under its own name; and the evidentiary channel, meaning the data its vendor holds that a court can order preserved.

The indemnity a buyer relies on

In the months before the suit, Microsoft and OpenAI had each announced commitments to defend enterprise customers against copyright claims arising from outputs of their services. Both promised to pay resulting judgments, on conditions. Other vendors made similar promises. They shifted real risk from buyer to vendor, but their value lay in the terms.

Procurement and legal teams can test a commitment with five questions. Which products and license tiers does it cover, and does it reach outputs from fine-tuned or customized models? What customer conduct voids it, such as disabling content filters or deliberately prompting for protected material? Is it capped, and does the cap sit inside the agreement’s general liability limit? Does it cover only copyright, or also trademark and other claims? And is it incorporated into the agreement that actually governs the organization’s use, rather than a public web page the vendor can revise?

Embedded AI raises a further question. When a software vendor builds features on a model provider’s service, the provider’s promise runs to the software vendor, not to its customers. The buyer should ask whether that vendor passes the protection through or offers its own, and get the answer in the contract. Many organizations discovered that their most widely used AI features came through this route.

The content it publishes

An indemnity covers claims against the buyer arising from vendor outputs. It does not change who publishes. A regulated organization that uses generative AI to draft customer letters, patient education material, policy explainers or public reports issues them under its own name. If a passage closely reproduces a protected work, the publisher carries the reputational exposure, and the indemnity’s conditions decide whether the vendor shares the legal one.

Output practices can match use to risk. Internal drafting and summarization carry little exposure. Material published externally at scale, such as marketing, educational content and anything that summarizes news or research, deserves a human reviewer who checks for close reproduction and confirms that sources are credited or licensed. Recording which tool and version produced each published item costs little and lets the organization respond quickly to a claim. Compliance teams in banking already review marketing for accuracy, and health systems already review patient materials for clinical correctness, so the copyright check fits into existing review steps rather than requiring a new one.

The data its vendor holds

Litigation brings discovery and preservation duties, and those can reach data a vendor holds about all of its users, including organizations that are strangers to the case. An earlier article in this series treated AI chat histories as records held under a vendor’s retention policy. The Times suit added a sharper possibility: a court in someone else’s case ordering that data kept longer than either the vendor or the customer intended.

For regulated organizations this cuts against core obligations. Bank confidentiality duties, HIPAA, and agency records and privacy rules all assume the organization controls where its data goes and how long it stays. The protections are contractual and architectural. Enterprise agreements can offer zero or minimal retention of prompts and outputs, and organizations sending sensitive data should use them. Agreements should state how the vendor will respond to a preservation order or subpoena affecting customer data, including notice to the customer where the law allows. Health systems need a business associate agreement in place before any patient data reaches the service. And consumer tiers, which typically carry different retention and use terms, should be blocked for confidential work through technical controls as well as policy.

The same channel runs the other way when the organization itself is a party to litigation. A bank in a lending dispute or an agency in a records case may need to preserve its own staff’s AI prompts and outputs under a litigation hold. That is only possible if the organization knows which AI tools staff use, whether their logs are retrievable, and who can place a hold on them. Records managers should add AI services to the systems covered by legal hold procedures.

Buyers did not need to predict the outcome of a case that would take years. They needed contracts that allocated risk clearly, review steps that limited what they published, and data terms that kept their information out of other parties’ litigation.

With hindsight

The Times case moved slowly. In April 2025 the judge allowed its core claims to proceed, and it was centralized with other copyright suits against OpenAI in New York. The evidentiary channel became concrete in May 2025, when a magistrate judge ordered OpenAI to preserve ChatGPT output logs it would otherwise have deleted. OpenAI said the order did not reach enterprise customers or those with zero-retention agreements, and the obligation was later narrowed. Other cases produced the first substantive rulings. In June 2025 a federal judge in California found that training on lawfully acquired books could be fair use while letting claims over pirated copies proceed. That case, against Anthropic, settled later in 2025 for about $1.5 billion. Vendor copyright commitments spread across enterprise AI agreements. For regulated buyers, the exposure ran mainly through contracts, publication practices and data retention, as the three channels above suggest.