Claude text detection has entered private preview, with Anthropic handing verification API access to regulators, law enforcement, newsrooms, fact-checkers, independent researchers and European civil society groups. The method rests on SynthID-Text, published by Google DeepMind in Nature in 2024, and it only works for whoever holds the cryptographic key. The lab is upfront about where the mark disappears: short passages, code, and anything rewritten from scratch.
Key Takeaways
- The verification API opens to regulators, media, fact-checkers and researchers
- Models released after August 2, 2026 carry the mark, older ones are being backfilled
- The signal vanishes on short passages, on proofreading work and on code
Have an AI Sum Up This Article
ChatGPTWho Can Request the Verification Key
Anthropic opened access on Tuesday to Claude text detection, a verification tool it had kept in house until now. The eligibility list is named rather than open: regulators, law enforcement, media organisations, fact-checkers, independent researchers, educational institutions, European civil society groups and enterprises under compliance duties.
This is not a statistical detector. Rather than an arbitrary random number generator, Claude leans on a cryptographic key to pick between equally valid words, and that key is the only way to read the trace back. The lab walked through the method in August in its technical note on how the watermark works.
The underlying theory is borrowed. It comes from SynthID-Text, published by Google DeepMind in Nature in 2024, which itself builds on a proposal Scott Aaronson put forward in 2022.
The technical perimeter is drawn cleanly. Models released after August 2, 2026 carry the mark by default, and older models are picking it up over the coming months.
Nothing shifts for the person typing. Anthropic insists there is no slowdown, no extra cost and no personal identifier buried in the output, and that a reader picks up nothing at all.
The mark itself is not new. We covered it on August 12, when the invisible signal spread across every text the lab’s models produce, C2PA-signed files included. What changes this week is the right to read it.
The Mark Fades on Short Text and on Code
Here is the part the lab documents without dressing it up. The signal weakens as soon as there are few lexical choices left to exploit, which means short passages first of all.
It also thins out on dense factual content that forces specific terminology. A sentence naming a work, a date and an author leaves almost no room for substitution, so almost no trace behind.
Proofreading is the most awkward case of all. When Claude edits human prose, the original text dominates and the mark is close to absent.
Anthropic also admitted it cannot separate a text the model wrote from a text the model heavily reworked. The question a newsroom cares about most is exactly the one the tool leaves unanswered.
Code falls in the same bucket, since exact syntax offers no alternatives. A full rewrite wipes everything, which makes the tool useless against the most common way of hiding AI involvement.
Those limits matter more because the volume keeps climbing. We reported in late August that AI writing now shows up on one web page in three, a mass that Claude text detection covers for a single model vendor only.
More articles on Horizon
- Claude Fable 5.1 Codes Better and Costs 25% Less
- AI Video Has Replaced China’s Short Drama Actors
- Claude Code Limits Drop 17% on September 14
The 190 Signatories of the EU Code Come Next
This opening is not a standalone commercial move. It follows from the European AI Act, which close to 190 organisations backed by signing the Code of Practice in July 2026.
Anthropic is therefore front-running an obligation aimed at everyone. Whichever lab ships a verification API first implicitly sets the shape the others will have to match, including the choice of a vetted list over open public access.
For newsrooms the arrival is real but partial. An unmarked text proves nothing, since it may come from another vendor, from a model that predates August 2, or from a rewrite. Claude text detection settles one question out of three.
Observers have flagged a second blind spot. Choosing a synonym by key rather than by meaning could weigh on text quality, and traceability raises its own problem wherever AI use is restricted by contract.
The issue already touches the most exposed editorial ground. The argument over declaring AI in award submissions surfaced during the 2026 Pulitzer Prize round, with no verification tool available to settle it.
The preview stays private for now, with a wider rollout promised and no date attached. As long as access runs through case-by-case clearance, verification remains an institutional instrument rather than a reader’s reflex.
Follow the story on Horizon.
