On 2 December 2025 the US District Court for the Southern District of New York, in The New York Times v OpenAI, denied OpenAI’s motion for reconsideration, maintaining its decision requiring OpenAI to produce a sample of 20 million user chat logs.
The dispute: which logs are “relevant”
The core dispute was not whether to produce logs but how broad the production should be.
OpenAI argued that only logs relating to the plaintiffs’ works — conversations whose outputs involved copyrighted works — should be produced. The court declined this narrowing. The judge held that the 20-million-log sample is relevant “in its entirety”, on the basis that even output logs that do not directly implicate the plaintiffs’ works still have probative value for assessing the model’s behaviour patterns, output probabilities and whether systematic reproduction exists.
The logic is that in a copyright infringement case, proving whether a model tends to systematically output protected content requires looking beyond the samples that already went wrong to the broader output distribution. Logs that “look fine” are precisely the comparison data for judging the severity of the problem.
Why this order matters to AI companies
The order directly affects OpenAI, but it reveals a risk structure that applies to all AI providers.
First, user chat logs can become litigation evidence. AI providers typically use user conversations for model improvement, safety review and quality analysis, meaning logs are retained long-term. Once a company is sued, those logs can be claimed as discovery material — including the personal information they contain.
Second, “relevance” may be read more broadly than companies expect. Companies assume data unrelated to the dispute should not be produced, but in cases about model behaviour, courts may require a broader set to establish a comparison baseline. This reading can significantly expand the production burden.
Third, discovery and privacy duties are in tension. Producing logs may involve user personal information and third-party content. Companies must arrange to reconcile disclosure duties with user privacy, typically through protective orders and de-identification. See our note on major personal information processor obligations and our cross-border data service.
Four practical points
First, revisit the chat data retention strategy. Whether to retain raw conversations long-term, for how long, and which fields, should be determined by business necessity, litigation risk and privacy duties together — not by default full long-term retention.
Second, build a litigation hold mechanism. Once litigation is anticipated, a company needs the ability to freeze and export relevant data by scope. Without a well-structured data architecture, discovery cost and risk multiply.
Third, design logs with layering and de-identification. Distinguishing “data usable for model improvement” from “personally identifiable data” at the architecture stage can reduce the exposure of sensitive information when disclosure is required. See our data compliance service.
Fourth, fold copyright litigation risk into data governance. The reason chat logs became the issue is ultimately the copyright dispute over training data. Training data provenance and authorisation chains remain the upstream risk point. See our note on the GEMA v OpenAI Munich ruling.
A trend judgment
From the Munich court finding that model “memorisation” of protected works infringes, to the New York court requiring 20 million chat logs, the focus of AI copyright litigation is expanding from “whether training was lawful” to “how the model runs and how data is retained”. For AI companies, data governance is no longer merely an internal compliance matter but a strategic asset that determines litigation position.
For AI companies preparing to go global, planning data retention strategy together with litigation risk is more economical than responding after the fact.
This article is based on orders in The New York Times v OpenAI before the US District Court for the Southern District of New York and public reporting. It is for general reference only and does not constitute legal advice.