You write a report, ask an AI tool to edit it and publish the revised version. If a detector finds a watermark, has it established that the machine wrote your ideas? ’s and ’s announcements leave that distinction worth examining.
A text watermark is a statistical signal introduced during generation. It can provide evidence that a supported model was involved, but it does not measure a person’s intellectual contribution. The signal, authorship and content quality need separate judgments.
On October 5, 2026, announced that eligible and text in the EU will receive invisible watermarks over the coming weeks. explained its system in August. We checked both announcements, technical documents and the discussion on X. Sources checked on October 6, 2026.
What OpenAI announced and what Anthropic announced earlier
OpenAI’s official post describes two paths: an EU rollout in and , and worldwide opt-in for select API models. It does not say that all text is already watermarked.
The technology is called textGrain. Initial text-detector access is limited to approved researchers and expert organizations; it is not a public checker available to any user.
published its explanation on August 14, updating detector information on September 1. It uses a version of SynthID-Text and also offers detection in private preview.
Scope differs. Claude’s documentation describes worldwide marking for supported models, including the API and Code. Support depends on the model; check before assuming it covers every output. starts with a regional product rollout and API opt-in.
How AI text watermarking works
A model generates , which can be words or word fragments. When several continuations are plausible, sampling can leave a statistical relationship with a key. A detector tests that relationship across the passage.
The textGrain technical report describes a method for controlling how much sampling randomness the watermark conditions. Detection needs the text and key, without knowing the strength setting used during generation. and textGrain use different algorithms.
says it adds no hidden characters. Checking for zero-width characters therefore does not test this watermark. The SynthID-Text paper distinguishes marking during generation from detectors that later try to classify writing by its style.
C2PA is different again. Content Credentials are provenance metadata attached to supported files. Metadata and a pattern within words require different checks.
Detecting a signal does not establish who did the work
Claude’s documentation explains that a watermark does not identify the user or distinguish generation from editing. Light proofreading may leave little signal; extensive rewriting gives the model more words to choose.
That matters for a business report. A person may research the issue, check sources and write the argument before asking AI to reorganize the prose. Detection observes a signal from that process, without establishing editorial responsibility or the source of each idea.
A negative result leaves questions too: which model was used, whether it supported marking, how much text remains and what transformations it underwent. Treating “not detected” as “written by a person” requires information the result does not contain.
How well textGrain survives editing
publishes an evaluation of 400-token English responses: detection fell from about 92% to 66% when 10% of words were replaced with synonyms, and to 17% when 25% were replaced. These are evaluation detection rates, not the probability that AI wrote a particular document.
Short passages and text with few wording alternatives are harder to detect. This chart does not compare against or establish performance in Spanish. Signal survival depends on the content, transformation and detector.
says it sees no meaningful performance differences in its evaluations. The SynthID-Text authors also found no significant feedback differences in an experiment covering roughly twenty million responses. These evaluate their respective systems; they do not replace testing your task and configuration.
What people question about text watermarks
The system’s usefulness is questioned given its limitations and EU regulation. The debate focuses on the benefit obtained when editing can weaken the signal. An announced rollout also needs to be distinguished from a completed one: refers to the coming weeks and does not establish that every translation or rewrite always removes the entire signal.
Another objection concerns detection requirements influencing a writing tool’s word choices. The drop from 92% to 17% cited in the debate comes from ’s evaluation. That figure is supported under those conditions; disagreement with the design and its costs does not establish a loss of quality by itself.
These objections concern the same announcement and offer perspectives on the debate, without constituting independent measurements of the system.
What the AI Act requires and what to keep when publishing
AI Act Article 50(2) requires machine-readable marking of synthetic content, with exceptions for standard editing or outputs without substantial alteration. It does not prescribe textGrain. Article 50(4) and the Commission’s summary separately address disclosure for text on matters of public interest, with an exception tied to human review or editorial control and responsibility for publication.
The provider’s technical marking and the publisher’s disclosure obligation are separate matters. A detector cannot determine how the rule applies to a particular publication by itself.
For a report or article, our practical recommendation is to retain the draft, record how AI was used, and keep the changes and final review. If authorship is questioned, that record provides context a binary label leaves out. The visual example is fictional; we have not run a detector.
Before adopting one, ask which marks it recognizes, how it was evaluated and how a doubtful result can be reviewed. Assess those answers separately from commercial promises to “detect any AI.”
You can explore LetBrand’s AI tools to organize your work. When publishing, choose a model around the task and keep a review you can explain: checked sources, traceable changes and a person responsible for the result.