New York Times v. OpenAI (Dec 2023): The Copyright Case That Defines AI Training Liability

Incident summary

On December 27, 2023, The New York Times Company filed a complaint against Microsoft Corporation and the OpenAI family of entities in the United States District Court for the Southern District of New York (Case No. 1:23-cv-11195). The complaint, published in primary form on the Times Company's own asset domain, runs to roughly 70 pages and includes exhibits documenting the alleged verbatim reproduction of Times articles by GPT-4 in response to prompts.

The complaint advances several causes of action including direct copyright infringement, contributory copyright infringement, vicarious copyright infringement, common-law unfair competition by misappropriation, trademark dilution under the Lanham Act, and violation of the Digital Millennium Copyright Act through removal of copyright management information. The Times seeks statutory and actual damages, an injunction against further infringement, and what the complaint describes as the destruction of GPT models and training data that incorporate the alleged infringing material.

Microsoft and the OpenAI defendants have responded with answers, partial motions to dismiss, and counter-claims. The case is on a multi-year litigation track. Coverage by the Times itself, published the day of the filing, remains a useful overview written for general readers; the complaint PDF is the canonical primary source for any specific legal claim cited from this matter.

Failure technique

The factual pattern alleged in the complaint is that OpenAI used Times articles, including current and archival reporting behind the Times paywall, as part of the training corpora for GPT-3.5, GPT-4, and downstream products including Microsoft's integration with Bing and Copilot. The complaint alleges that the resulting models can be prompted to reproduce Times content verbatim or near-verbatim and that retrieval-augmented chat features surface Times content in ways that bypass the paywall.

From a training-data-engineering perspective, the complaint highlights a class of choice that was treated as engineering convenience in 2020 and 2021. The Common Crawl dataset includes commercial news content. Many model developers used Common Crawl directly or via derived corpora without auditing for inclusion of paywalled or licensed material. The Times complaint asserts that the failure to audit and license is the heart of the alleged liability.

From a product-engineering perspective, the complaint highlights the question of whether retrieval-augmented generation that surfaces excerpts of paywalled content under fair-use framing is sustainable. Most retrieval-augmented systems treat retrieved snippets as background context, not as primary content delivery; the line between the two is contested in ongoing litigation across multiple cases.

Impact and consequences

Direct litigation impact remains pending. The case is on a multi-year track. The complaint is the canonical primary record of the Times's claims; rulings, settlements, and counter-claims are recorded in the docket and tracked by legal media.

Industry impact is already visible in the public record. Several frontier labs, including OpenAI and Anthropic, have announced or expanded direct licensing programs with news publishers since the filing. Publisher coalitions on both sides of the licensing question continue to publish positions on AI training, fair use, and access. The United States Copyright Office released a multi-part report on Copyright and Artificial Intelligence that addresses many of the doctrinal questions implicated by the case (primary source linked above).

Compliance and engineering impact is concentrated in training-data governance roles. Documented provenance of training data, licensing records, and an opt-out mechanism for publishers have become baseline expectations for AI training operations in 2025 and 2026, even before the Times litigation resolves.

Lessons for builders

Treat the training corpus as a licensing surface. The convenience of crawling the open web does not eliminate the licensing question for paywalled or rights-protected content. Document provenance for every dataset that enters training.

Maintain a publisher opt-out mechanism and honor it. Robots.txt directives, AI-specific user-agent strings, and contractual licensing all need handling in the training pipeline. Failure to honor an opt-out is one of the cleanest litigation hooks the complaint surfaces.

Audit retrieval-augmented generation for licensed-content exposure. If the system can surface excerpts of paywalled material in user-facing output, the team should have a documented licensing or fair-use posture, reviewed with counsel, before the surface ships.

Build training-data governance as an engineering function, not an audit-time reconstruction. The Applied AI roles that own this work are AI Data Engineer, AI Compliance Officer, and AI Governance Lead. Documenting provenance after the model trains is harder than documenting it while the data is assembled.

Mitigations

What builders should put in place to address the failure pattern. Each mitigation maps to operational practice the relevant Applied AI roles own.

Related Applied AI roles

The Applied AI roles whose day-to-day work would have prevented, detected, or contained this incident.

• AI Data Engineer: An AI Data Engineer designs and operates data pipelines specifically for AI training and serving.

Related AI Decipher Files

• EU AI Act Implementation: First Horizontal AI Regulation Goes Operational
• Samsung ChatGPT Data Leak: When Consumer AI Became an Enterprise Exfiltration Channel

Frequently asked questions