Unsealed Briefs in Authors’ Case v. Microsoft/OpenAI

Who Should Read

Backend or ML developers who build their own training or RAG datasets and worry about copyright risk

Core Mechanics

• Unsealed filings put the defendants' knowledge at the center
  • The Authors Guild announced on September 21, 2026 the filings released the previous Thursday.
  • It says they center the case on what OpenAI and Microsoft knew when they acted.
• OpenAI worried about optics of using LibGen
  • Per the article, researcher McCandlish wrote he feared 'openai uses copyrighted data from sketchy russian website' showing up on Hacker News.
  • Then-Research Director Dario Amodei replied that LibGen was 'a bit sketchier' as a training set.
• Microsoft allegedly knew from April 2019
  • The brief says Altman and Amodei showed an early GPT-3 to Bill Gates and disclosed LibGen use.
  • CTO Kevin Scott and others were also present.
• LibGen files were deleted in summer 2022
  • The brief says OpenAI deleted them under 'Project Clear'.
  • On June 15, 2022, VP of Research McGrew wrote in Slack that it was time to excise LibGen from systems and storage.

Community Reactions

• Comment: fair use law may matter more than quotes
  • One commenter thinks courts will weigh fair use doctrine and economic studies more than these statements.
  • They say the quotes may matter more for willfulness and damages after infringement is found.
• Comment: evidence of AI replacing authors is thin
  • One commenter says one study found no real impact and another pointed to competition among authors who use AI.
  • They say they have seen no direct evidence of AI replacing authors.
• Commenter's experience with LibGen data
  • One commenter who pulled LibGen datasets for a corpus says they were overwhelmingly in-copyright textbooks.
  • They say that undercuts the public-domain framing.
• Another filing excerpt quoted by a commenter
  • A commenter quoted Ryan Lowe's July 2020 risk assessment of using LibGen for a book-summarization project.
  • It put the chance of being unable to say where the books came from above 80% and of a Twitter kerfuffle at about 40%.

How to Apply

• Build a dataset provenance list
  • Today, list every corpus you use for fine-tuning or RAG with source URL, license, and collection date.
  • Flag any entry whose source you cannot explain.
• Ask whether shadow-library data is mixed in
  • Ask legal or your team lead whether data from sources like LibGen is in your pipeline.
  • If so, pause training jobs that use it and get legal review.
• Record dataset decisions in writing
  • Require an approver and rationale whenever a dataset is added.
  • Internal chats were quoted as evidence here, so raise concerns in official channels.

Try It

# Paste into Claude or ChatGPT as-is (the source is a lawsuit article with no method, so this is a simplified dataset provenance check)

You are a reviewer helping audit data provenance for AI projects. You are not giving legal advice; you help draft a checklist.

Below are the datasets we use for model training or RAG.
{{dataset list: name, source URL, collection method, license}}

For each dataset, produce a table with:
1. Is the source clear?
2. Is the license verified?
3. Any warning signs that copyrighted books or papers are mixed in?
4. Questions to ask the legal team

If information is missing, do not guess; write 'needs verification'.

Next: turn each 'needs verification' row into a team issue and assign an owner.

Terminology

Related Papers

• The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge

  Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.

• 90% of US working hours are spent on tasks not covered well by AI benchmarks

  Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage

• Postgres 19 delayed after AI reviewers find more bugs than humans can fix

  LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted

• Show HN: Gemini 4 Argon plays StarCraft

  LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.

• A Plausible Explanation Is Enough to Mislead Multimodal Judges

  Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.

• Neural networks are clean

  Looking at neurons directly shows most have one main job, challenging superposition.

Related Resources

• https://authorsguild.org/news/ag-v-openai-top-execs-knew-mass-book-piracy-was-illegal/
• https://openai.com/policies/row-terms-of-use/