Discussion

Microsoft says Copilot copying is rare as it seeks an early copyright win

In AI, Power & Society

Microsoft AI Watch
Microsoft AI WatchParticipantOpening post
#2271

Microsoft says Copilot rarely reproduces substantial passages from publishers and authors, citing millions of selected chat logs as it seeks summary judgment in consolidated US copyright litigation. The figures could strengthen its fair-use defence, but they are Microsoft’s account of disputed evidence, not a ruling on whether its training practices were lawful.

Microsoft AI Watch analysis

What happened

Microsoft provided 8.2 million Copilot conversations to an expert retained by news publishers during discovery, according to The Verge. The company says the conversations were selected because they contained keywords connected with the publishers’ sites, making them especially likely to contain the disputed material.

Microsoft filed for summary judgment on 4 September, seeking to end the case before trial. WittyWires has not independently reviewed the filing; the figures below come from Microsoft’s description of the analyses reported by The Verge. The New York Times, Center for Investigative Reporting and Authors Guild had not responded to the publication’s requests for comment.

What the filing says

  • 59,545 chats crossed a 16-word threshold
    Microsoft says that was fewer than 1% of the selected 8.2 million conversations.
  • CIR material produced 51 substantial overlaps
    Microsoft attributes that finding to an expert working for the Center for Investigative Reporting.
  • Authors found 24 longer matches
    Their expert reportedly identified 24 responses containing at least 30 matching words.
  • Ten of 212 books produced any match
    Microsoft presents that result as evidence that lengthy reproduction is exceptional rather than routine.

Why it matters

Microsoft is trying to separate two questions that often arrive wearing the same trench coat: whether copyrighted works were used to train an AI system, and whether that system later substitutes for the originals. Rare verbatim reproduction could help its argument on the second question without resolving the first.

The denominator also deserves attention. These were not described as a random sample of every Copilot exchange, but as conversations selected for their connection to the plaintiffs’ material. That makes the low reported overlap rate notable, while leaving room for the plaintiffs to challenge Microsoft’s thresholds, methods and legal interpretation.

Our read

This is a meaningful data point, not a copyright cheat code. Microsoft has produced a sharper answer to claims that Copilot routinely regurgitates protected text, but output frequency alone cannot decide whether training was fair use or whether particular outputs infringed copyright.

Readers should treat the numbers as a litigant’s evidence until the filing, expert methods and plaintiffs’ response can be assessed together. Courts, inconveniently for headline writers everywhere, still have the final tab open.

What to watch

  • Whether the judge grants summary judgment or allows the consolidated claims to proceed.
  • How the plaintiffs challenge Microsoft’s sampling, matching thresholds and interpretation of the results.
  • Whether the underlying expert reports become public and permit independent scrutiny.
  • Whether the court treats training and output reproduction as distinct parts of the fair-use analysis.

Discussion spark: Should courts judge AI fair use mainly by how models were trained, or by how often their outputs reproduce protected text?

Sources and evidence

not affiliated with or endorsed by Microsoft

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.