OpenAI's text watermark covers EU users only

Today's record holds two systems for marking AI output and two live disputes over AI-generated work, and no mark had anything to do with how either dispute surfaced.

Two marks, both shipped with their limits stated

OpenAI will add an invisible watermark to ChatGPT's text output for users in the EU only, to comply with EU rules, saying a regional approach gives it room to learn before taking the change wider (PCMag). The ledger's record for OpenAI puts ChatGPT at more than 900 million weekly users as of March 2026 (OpenAI), so the mark arrives scoped to the jurisdiction that asked for it rather than to the product.

Google DeepMind published the research version of the same idea, together with the ways around it: a watermark embedded in AI-designed protein sequences and structures, with true-positive detection above 99.8% at a 0.1% false-positive rate across 489 sequences expressed against three targets (Nature). The same paper states that the sequence watermark can be stripped by resequencing through ProteinMPNN, and that the structure watermark does not survive force-field relaxation.

That is the honest shape of the technique on today's record: strong detection while the mark is intact, and named procedures that leave it absent.

Both disputes were caught by people reading the underlying data

Elsevier is investigating two 2026 papers in the Journal of Neurolinguistics that claim to analyse public OpenNeuro brain-imaging datasets, after the researchers who collected the data reported that participant counts and measures in the papers do not match the datasets and said the papers appear to have been generated by AI (Retraction Watch). What caught it was a mismatch against the source data, noticed by the people who produced the source data.

Anthropic said its Claude model found a reverse transcriptase sitting inside a tandem repeat array, which it named an array-associated RT, after a 21 hour search of DNA sequence databases (Anthropic). The work is an unreviewed preprint, no gene-editing activity has been shown, and a researcher who had shared unpublished work on the same enzymes with the public version of Claude has questioned whether the model drew on it, which Anthropic denies.

That second dispute is about an input, not an output. A mark on what a model emits carries no information about what went into it, so the question raised against the enzyme result sits outside what either watermark published today is built to answer.

The route that is actually in place is the slow one

MDPI retracted a 2023 Atmosphere paper disputing the link between carbon dioxide and warming, citing unsupported conclusions, inaccurate interpretation of the data and reference attribution problems; the paper had been submitted and accepted in under a month and had drawn 26,000 views before the September 16 notice (Atmosphere). No AI claim attaches to that paper, which is the point of citing it here: it dates the alternative to a working mark. Acceptance took weeks, the correction took three years, and the readership was already counted when it landed.

What today's record supports

• A ChatGPT text watermark exists for EU users, and the regional scope is a stated choice about where to start, not a limit of the method.
• A protein watermark detects above 99.8% at a 0.1% false-positive rate, and its own authors name two procedures that remove it.
• Neither investigation now open was triggered by a mark: one came from a dataset comparison, one from a researcher's account of what they had shared.
• The one correction that has completed on this record took three years and followed 26,000 views.