Yes—OpenAI announced a statistical watermark for eligible ChatGPT and Codex text generated in the EU, with rollout planned over the weeks following its 5 October 2026 announcement. The mark is intended to be detectable by software, not visible to readers: it changes the model’s word-choice patterns rather than inserting hidden characters or unusual spaces. The EU AI Act requires machine-readable marking of covered AI-generated content where technically feasible, but it does not require OpenAI to use this particular method.
What OpenAI announced—and who it applies to
OpenAI’s 5 October 2026 announcement describes textGrain, a statistical watermark for eligible ChatGPT and Codex text output in the European Union. OpenAI said rollout would happen over the coming weeks. Because that was a prospective schedule, availability may differ by product surface or account; the announcement does not establish that the feature is already active for every EU user. OpenAI’s announcement
Separately, API customers worldwide can opt in to the feature for select models. It is off by default for those customers, so this is not a claim that all API-generated text is automatically marked. OpenAI’s announcement
What “invisible watermark” means
textGrain does not add a visible label or encode the mark as hidden characters, invisible spaces, or odd punctuation. Instead, it subtly alters how the model selects among plausible words or word pieces. Across a long enough passage, those choices can form a statistical pattern that a detector can look for. The mark is therefore in the wording itself and may be imperceptible to a reader. OpenAI Help Center: provenance signals
#1 Best Overall
This is different from metadata attached to a file: copying text into another document does not necessarily remove a signal embedded in word choice, but changes to the wording may affect what a detector can identify. OpenAI’s materials do not establish a reliable, universal way to remove or preserve the signal through every kind of editing, so a detector result should not be treated as a guarantee about a passage’s history.
Does the EU AI Act require ChatGPT to watermark text?
Article 50(2) of the EU AI Act says providers of systems that generate synthetic audio, images, video, or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The measures must be effective, interoperable, robust, and reliable as far as technically feasible, while accounting for content-specific limits, costs, and the state of the art. The Act does not name textGrain or prescribe one text-watermarking technique. Article 50 of the EU AI Act
Rank #2
The European Commission says these Article 50 obligations apply from 2 August 2026. Its Code of Practice is voluntary; the transparency obligations in the Act are not. The Commission describes the code as an adequate voluntary way to demonstrate compliance, while organizations using other means must show that their approach is adequate. European Commission guidance and Code of Practice
Provider marking is not the same as a user disclosure
The Act also has a separate rule for deployers—organizations or people using AI systems. Under Article 50(4), a deployer publishing AI-generated or manipulated text to inform the public on a matter of public interest must disclose that it was artificially generated or manipulated. The provision includes an exception where the content has undergone human review or editorial control and a person holds editorial responsibility, as well as a specified criminal-law exception. This is not a general requirement that every ChatGPT user label every private message or routine piece of writing. Article 50 of the EU AI Act and European Commission guidance
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →How well can textGrain detect ChatGPT writing?
OpenAI reported that, at a target false-positive rate of 1%, its evaluation detected watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. OpenAI said results were substantially lower for mathematics, where word choice is less flexible. These are OpenAI-reported evaluation results, not an independent test or a universal success rate for real-world text. OpenAI’s announcement
Performance depends on the material. Short passages provide fewer word choices to analyze, and constrained text—such as mathematical writing or code—may offer less room for the watermarking pattern to emerge. OpenAI’s Help Center says the Code of Practice does not require watermarks for outputs shorter than 200 tokens, approximately 150 English words, or for code snippets. That threshold is not proof that every longer passage will be marked or detectable. OpenAI Help Center: provenance signals
Rank #4
OpenAI cautions that performance under ideal conditions does not ensure reliable everyday detection. False positives and false negatives are possible. A positive result is evidence of a signal associated with a provider; by itself, it does not conclusively prove who authored a passage, whether a person edited it, or how it was used. OpenAI’s announcement
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Can you check whether ChatGPT wrote something?
Not with a general-purpose public OpenAI text checker, based on the company’s published access description. OpenAI says its text detector is initially available only to approved researchers and expert organizations, with applications reviewed case by case. The public verification tools described in its Help Center cover supported image or audio provenance, not general consumer text checking. OpenAI’s announcement and OpenAI Help Center: provenance signals
Recommended Free Tools
Text marking is also only one of several approaches under consideration. A 2026 European Commission technical report groups methods into watermarking, structural marking, metadata, logging, and AI-generated-text detection. These approaches can be compared by whether they distinguish generated from human content, withstand editing, work consistently across tasks, are accessible and interpretable to relevant stakeholders, and can operate across providers and technologies. European Commission technical report
What this means for readers and publishers
- Do not treat a detector result as authorship proof. The reported error risks and uneven performance matter, particularly for short or constrained passages.
- Do not assume every ChatGPT response is marked. OpenAI described eligibility limits, a staged EU rollout, and an API opt-in that is off by default.
- Separate the technical signal from legal disclosure. A watermark is a provider-side marking measure; Article 50(4) addresses certain deployers publishing public-interest text.
OpenAI itself described the technology cautiously: “Text watermarking and detection remain early technologies with significant limitations, and views about their benefits and responsible uses are still developing.” OpenAI, 5 October 2026
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
