PANONDA
Últimas
← News
Regulação·3 min read

Anthropic tells investors it may not know what copyrighted work it trained on

A new Anthropic filing says its systems for respecting paywalls and robots.txt may not always work, and that it may be unable to verify whether training material is copyrighted. For publishers, the paper trail now matters more than the crawler setting.

By PANONDA Newsroom

Foto: AlphaTradeZone · Pexels

What happened

Anthropic disclosed in a filing reported on November 21, 2026 that its systems for respecting paywalls and robots.txt may not always work as intended, and that it may be unable to verify whether material used to train Claude is copyrighted.

The same filing warns investors that the company could face significant licensing costs if courts reject its fair-use defense. Anthropic said courts in the United States and elsewhere have not authoritatively decided whether fair use covers its training practices.

That is not a lawsuit, a settlement or a licensing announcement. It is a company writing down, for investors, the limits of the controls publishers have been told to rely on.

Why this one is different

Most publishers treat a robots.txt block or a paywall as a clean technical barrier: if the door is locked, the crawler didn't get in. Anthropic's own disclosure undercuts that assumption from the inside.

If the company doing the crawling says its exclusion systems may fail and that it can't always verify copyright status, then your robots.txt file is not proof your work stayed out — and it isn't proof the AI company can show it stayed out either.

The context is not theoretical. Anthropic previously paid $1.5 billion to settle a copyright claim involving pirated books, with Bloomsbury among the recipients. Canadian publishers have filed two additional lawsuits alleging journalism was used to train Claude.

What this story is not

It is not an admission of infringement. Risk-factor language in a filing is written by lawyers to pre-empt surprise, and "may not always work" is doing heavy lifting. Expect Anthropic to argue exactly that if it gets quoted back at them in court.

It also does not give you a number. The filing does not disclose a per-article licensing rate, a royalty formula, the total count of publisher works used to train Claude, or what any individual publisher could recover. Anyone telling you this filing sets a price is reading something that isn't there.

And it is not a change in the law. The legal question — whether fair use covers training on your archive — is still open, which is the one part of this the filing states plainly.

One adjacent item worth watching but not acting on: India's media industry has floated a "one national, one license, one payment" model, with blanket access to lawfully obtained content and royalties collected centrally when an AI product is commercialized. It's a policy proposal outside the US, not a rule you comply with.

The practical shift

For two years the publisher playbook has been configuration: block the crawler, wall the content, move on. This filing moves the center of gravity to evidence. In any future claim or negotiation, what you can document about access and ownership will matter more than what your server was set to do.

Do this before Monday

  1. Pull and archive your crawler logs. Blocked-request records, user-agent hits, dates. Store them somewhere that isn't a server you'll migrate next quarter.
  2. Snapshot your robots.txt history. Current settings prove nothing about 2022. Save the historical versions and the dates each one went live.
  3. Audit the older archive. Identify what was openly accessible before the block went in — that window is where exposure lives.
  4. Separate access evidence from ownership evidence. Keep a clean record of which pieces are copyrighted, publication dates, and who controls the rights. Freelance contracts included.
  5. Document your paywall configuration over time. Metered, hard, leaky-by-design for SEO — write down which it was and when.
  6. Send vendors a written question list. Crawler identity, collection dates, exclusion process, how paywalled pages are treated, and the mechanism for correcting or removing improperly collected material. Get it in email, not on a call.
  7. Treat licensing talks as commercial, not corrective. A disclosure that licensing costs "could become significant" is leverage context. It is not a statement of what you are owed.

Sources

Everything above was checked against these pages. Open them and see for yourself — that is why they are here.

  1. 1journalarta.com
  2. 2axios.com
  3. 3economictimes.indiatimes.com
Share

Get the stories by email

What changes in AI for people who publish — sourced, and to the point.

No spam. Unsubscribe anytime.