OpenAI will watermark ChatGPT text in the EU only, leaving Canadian users out for now

OpenAI says it will start embedding an invisible watermark in text produced by ChatGPT and Codex, but only for users in the European Union, where new transparency rules under the AI Act now require AI-generated content to be machine-detectable. In a blog post on Monday, the company said the watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans in the EU. “We are not making text watermarking a global default at launch,” it said, describing the regional approach as a way to learn from real-world use. That means ChatGPT users in Canada and elsewhere outside the bloc will not get the feature at launch.

Outside Europe, the change is opt-in for developers. Starting Monday, API customers anywhere in the world can choose watermarked text output for select models, though the setting stays off by default. OpenAI said it is also working with cloud partners to offer watermarking for its models accessed through their services in the coming weeks. For Canadian companies building products on OpenAI’s models, whether to watermark will be their own call.

The technology, called textGrain, adds an invisible statistical signal to the model’s word choices that a detector can then look for. Because the signal lives in the words themselves, it travels with the text when it is copied and pasted, and it does not identify the user, TechCrunch reported. A technical report released with the announcement, co-written with researchers from the University of Pennsylvania and Yale, describes using a secret key to sort the model’s next-word predictions, so that hundreds of small nudges add up to a pattern the detector can spot using only the text and the key. OpenAI said textGrain “matched or exceeded” other approaches it tested, including Google DeepMind’s SynthID for text, The Verge reported.

The company was frank about the limits. At a target false-positive rate of one per cent, its detector found watermarks in about 80 per cent of 200-token passages and about 95 per cent of 400-token passages on psychology content, but detection was substantially lower for mathematics, where word choice is less flexible. Editing weakens the signal: replacing 10 per cent of words with synonyms cut detection in 400-token passages from about 92 per cent to 66 per cent, and replacing 25 per cent cut it to 17 per cent. OpenAI said benchmark scores for its Astra model showed no meaningful difference with the watermark switched on.

Those weaknesses are why the detector will not be public at first. Approved researchers and expert organizations can apply for access starting Monday, with approvals made case by case, OpenAI said. The company also stressed what a watermark cannot show: it does not measure how much a human contributed, establish ownership, verify accuracy or identify the user, and “the absence of a detected watermark does not prove human authorship,” since text may be too short, edited, translated or produced by another company’s tools. Its image and audio verification tools, including a web checker and an API, remain publicly accessible.

The legal driver is Article 50 of the EU AI Act, which applies from Aug. 2, 2026. It requires providers of AI systems that generate synthetic audio, image, video or text to ensure outputs are marked in a machine-readable format and detectable as artificially generated, using solutions that are “effective, interoperable, robust and reliable as far as this is technically feasible.” The article exempts systems that only assist with standard editing, and separately requires organizations that publish AI-generated text on matters of public interest to disclose it unless the text has undergone human review and someone holds editorial responsibility.

OpenAI is not alone. Anthropic said two months ago that it would watermark text generated by Claude worldwide, a decision that drew pushback from some users who argued they had supplied “the instructions, context, decisions” while the model was just “the tool,” TechCrunch reported. Anthropic, Google, Meta, Microsoft and OpenAI are among companies that have committed to follow the EU’s code of practice on AI-generated content. OpenAI had previously built a text watermark but held off on releasing it, partly over concern that users would move to rivals that did not watermark, The Wall Street Journal reported in 2024.

In Canada, the closest benchmark is voluntary. Ottawa’s 2023 Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems asks developers of publicly available systems to “develop and implement a reliable and freely available method to detect content generated by the system,” with a near-term focus on audio-visual content such as watermarking. OpenAI does not appear on Innovation, Science and Economic Development Canada’s published list of signatories, which includes Cohere, OpenText, TELUS, Mila and the Vector Institute.

OpenAI said it plans to release textGrain as open source, will update its technical report in the coming weeks and expects to revisit each part of its approach as the technology, standards and evidence evolve.

Sources: OpenAI; TechCrunch; The Verge; EU AI Act, Article 50; Innovation, Science and Economic Development Canada.