The New York Times Launches a Strong Case Against Microsoft and OpenAI

Original Article ↗ Hacker News Discussion ↗

New York Times’ lawsuit against Microsoft and OpenAI over the use of its articles to train and power AI systems has become a focal point for unresolved questions about copyright, fair use, and AI-generated content. Commenters dissect the legal theories—from verbatim reproduction and DMCA claims to whether AI services unlawfully compete with or substitute for news outlets—and debate if training on copyrighted data should require licenses or be protected as fair use. The outcome is seen as potentially reshaping business models for both AI developers and content producers, with implications for open-source models, global competitiveness, and future copyright regimes.

Scope of NYT’s Claims

  • NYT isn’t only challenging training; the complaint targets:
    • Models that can output NYT articles or large excerpts verbatim.
    • “Synthetic search” / browsing features (Bing Chat, ChatGPT Browse) that fetch NYT content in real time and paraphrase or quote it.
    • Use of NYT content without a license, including alleged removal of copyright management info (DMCA §1202).
  • NYT frames OpenAI/Microsoft as creating products that directly substitute for NYT (including Wirecutter), harming its business model.

Fair Use, Training, and Verbatim Reproduction

  • One camp: training on copyrighted text may be fair use; occasional regurgitation from broadly trained models shouldn’t, by itself, be infringement.
  • Another camp: parts of the case are “easy” for NYT because the systems allegedly reproduce articles almost verbatim or act as paywall workarounds.
  • Debate over whether the key question is training-as-fair-use or downstream use and market substitution (with Warhol precedent cited for “use not transformative”).

RAG, Paywalls, and Competition

  • “Grounding” / retrieval-augmented generation allegedly:
    • Copies NYT pages (including paywalled ones), feeds them to the model, then outputs paraphrases or long quotes.
    • Strips attribution and affiliate links (e.g., Wirecutter), turning NYT into an uncompensated backend.
  • Some argue this resembles Google snippets and might be fair use; others note Google links back and already pays licensing in some jurisdictions.

Liability, Libel, and Hallucinations

  • Thread disputes whether hallucinated NYT “quotes” can support libel or trademark claims:
    • Some say “actual malice” or reckless disregard could attach to companies that know models regularly fabricate.
    • Others call this weak: damages are hard to prove and LLMs’ outputs are explicitly caveated as fallible.
  • Strong disagreement on calling outputs “lies” vs “hallucinations” vs “fabrications.”

Human vs Software Learning and Personhood

  • Large subthread on whether “humans learn from copyrighted works, so AI can too” is a valid analogy:
    • One side: software is just a tool; only humans/corporations can be liable, and scale + exact copying matter.
    • Other side: if automating what humans may legally do is allowed in other tools (e.g., Photoshop), it’s unclear why LLMs should be treated differently.
  • Additional debate on corporate personhood, possible future AI personhood, and whether humans are “special” in law and morality.

Economic, Competitive, and Policy Implications

  • Some expect a settlement and licensing regime (possibly collective or compulsory licensing) rather than a “kill AI” precedent.
  • Others warn:
    • Heavy licensing could entrench big tech and big media, locking out open-source and small players.
    • Strict US rulings may advantage jurisdictions that loosen training rules (e.g., references to Japan/EU debates).
  • Several commenters think NYT’s endgame is leverage for a better licensing deal, not principled opposition to AI.

Technical and Open-Source Notes

  • Commenters discuss:
    • Temperature=0 and crafted prompts enabling verbatim regurgitation; unclear how common this is in normal use.
    • Possibility of post-processing (plagiarism checks, forced citation, RAG with explicit links) to mitigate infringement.
  • Some predict lawsuits will mainly hurt proprietary US models, while open-source and foreign models trained outside US legal reach will keep advancing.