OpenAI rolls out invisible textGrain watermarking for ChatGPT and Codex in EU to comply with AI Act
OpenAI will embed invisible textGrain watermarks into ChatGPT and Codex outputs in the European Union over the coming weeks to comply with Article 50 of the EU AI Act. The watermark embeds a statistical signal in word choices, detectable only by OpenAI's proprietary tool, which is currently restricted to approved researchers. API customers worldwide can enable watermarking today for select models. The system shows detection rates of about 80% for 200-token passages and 95% for 400-token passages at a 1% false positive rate, but editing significantly reduces detectability.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Common ground
- OpenAI's watermark is technically weak, with detection dropping from 92% to 17% after a 25% synonym swap, making it easy to bypass.
- The 1% false positive rate at 200 tokens is a real risk, especially in settings like universities where it could lead to unfair accusations.
- OpenAI controlling both the watermark and the detector, while limiting access to 'approved researchers,' creates a monopoly on verification and undermines transparency.
- The EU AI Act's vague language on 'machine-readable marks' allowed OpenAI to define compliance on its own terms, resulting in a system that checks a legal box without real accountability.
- The territorial limitation—deploying the watermark only in the EU—creates a two-tier internet, putting EU users and businesses at a disadvantage while non-EU users face no restrictions.
Points of contention
- One side argues the watermark is a deliberate 'compliance artifact' designed to satisfy regulators without catching bad actors, while the other sees it as a naive acceptance of a weak system that enables regulatory capture.
- There is disagreement on whether the 1% false positive rate is a manageable risk or a systemic threat to due process, with one side calling it 'not nothing' and the other warning of thousands of innocent people being flagged.
- One side believes the EU is a victim of OpenAI's strategic maneuvering, while the other argues the EU is complicit by not mandating open-source or cryptographic solutions.
Blind spots
- The debate overlooks how the watermark alters model outputs for EU users, creating a reproducibility problem for researchers and a competitive disadvantage for EU businesses.
- Cryptographic provenance—where the model signs its output with a verifiable key—is a technically feasible alternative that doesn't degrade with editing, but it was never seriously considered by regulators.
- The territorial limitation sets a dangerous precedent for other global regulators, who may now accept similarly weak compliance measures from tech companies.
WorldAttention’s read
This debate reveals that OpenAI's watermark is a performative compliance measure—technically fragile, controlled by the company, and deployed only in the EU. While there's consensus on its weaknesses and the risks of false positives, the deeper issue is regulatory capture: the EU accepted a black-box system from the very company it's supposed to oversee, instead of mandating open-source or cryptographic provenance tools. The territorial limitation creates a two-tier internet that harms EU users and businesses, while setting a weak precedent for global regulation. Ultimately, reliable text provenance at scale remains an unsolved problem, and this system does more harm than good by creating an illusion of control.
Reporting timeline
OpenAI to invisibly watermark ChatGPT and Codex text in EU under AI Act compliance
OpenAI will roll out an invisible watermarking system for text generated by ChatGPT and Codex in the European Union, complying with Article 50 of the EU AI Act. The watermark, called textGrain, embeds a statistical signal in the pattern of word choices rather than special characters, making it invisible to readers but detectable by OpenAI's proprietary tool. The detector, currently kept internal and shared only with approved researchers, achieved about 80% detection of 200-token psychology passages and 95% of 400-token ones at a 1% false positive rate. However, editing erodes the signal: swapping 10% of words for synonyms cut detection from 92% to 66%, and 25% swapping reduced it to 17%. Mathematics answers scored lower due to limited interchangeable words. The watermark does not identify users, prompts, or owners. Eligible users on all plans will receive the watermark over the coming weeks, while API customers worldwide can enable it today for select models.
Read sourceOpenAI to add invisible watermarks to ChatGPT and Codex outputs in the EU
OpenAI has announced it will begin adding invisible watermarks to text outputs from its ChatGPT and Codex models in the European Union over the coming weeks. The announcement, made via a post on the prediction market platform Polymarket, details a new policy aimed at identifying AI-generated text. The invisible watermarks are intended to be a technical measure to trace the origin of content produced by these AI systems. The rollout is specifically targeted at the EU market, with no mention of a global expansion at this time. The move comes amid increasing regulatory scrutiny of AI-generated content in Europe, particularly under the EU's forthcoming AI Act. The exact technical specifications of the watermarking method and how it will affect user experience have not been detailed in the announcement.
Read sourceOpenAI to watermark ChatGPT and Codex outputs in EU under new AI Act compliance
OpenAI will begin applying invisible text watermarks to outputs from ChatGPT and Codex in the European Union over the coming weeks, in response to the EU AI Act. The watermark is embedded in the wording itself, with OpenAI stating tests show no meaningful quality loss. However, the author expresses concern about detection requirements influencing how the writing tool phrases text. The post highlights limitations: in a test on 400-token passages, replacing 25% of words with synonyms reduced watermark detection from approximately 92% to 17%. Additionally, the watermark cannot distinguish how much a human contributed, meaning text written by a human and edited by ChatGPT can still carry the watermark. The author criticizes the EU's focus on this issue with sarcasm.
Read sourceShow 4 older updatesHide older updates
OpenAI expands content provenance to text watermarking for EU AI Act compliance
OpenAI announced it is expanding its content provenance approach to include text, responding to EU regulatory requirements under the EU AI Act. The company acknowledged significant limitations of current text watermarking technology. OpenAI already provides tools to verify whether images or audio files were created with its models, and this new effort builds on that work to help people better understand when content may have been generated or edited with an OpenAI model. In the European Union, OpenAI will begin watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Additionally, customers using OpenAI's API can enable text watermarking for select models worldwide starting today.
Read sourceOpenAI to roll out text watermarking for ChatGPT and Codex in the EU
OpenAI announced it will roll out text watermarking for its AI models ChatGPT and Codex in the European Union in the coming weeks. The watermarking feature will remain off by default in the API. The company stated that across benchmarks used to assess its latest frontier model, Astra, it does not see meaningful performance differences with and without watermarking. The post, shared on X, includes a comment from the user indicating skepticism about the claim, but the core announcement is attributed to OpenAI.
Read sourceOpenAI adds invisible watermarks to ChatGPT and Codex text output in EU
OpenAI is implementing invisible watermarks on text generated by eligible ChatGPT and Codex models in the European Union, according to a post on X. The watermarking feature will remain off by default in the company's API. Additionally, OpenAI is opening access to its text watermark detector for approved researchers and expert organizations. The post asks whether AI-generated text should be traceable, reflecting ongoing debates about AI content provenance and transparency. The move appears to align with EU regulatory expectations, though the post does not specify a timeline or technical details of the watermarking system.
Read sourceOpenAI Unveils textGrain Watermarking for EU AI Act Compliance, Rolls Out to API and ChatGPT
OpenAI has announced textGrain, a text watermarking technology designed to comply with the EU AI Act. The system embeds invisible statistical signals into the word selection process of AI models. Starting immediately, API customers can selectively enable watermarks on certain models. Within the next few weeks, OpenAI will add invisible watermarks to outputs from ChatGPT and Codex for users in the European Union. The accompanying detection tool is currently only available to approved researchers and expert institutions. OpenAI also released detection performance data and limitations, allowing readers to assess the current capabilities and constraints of the text watermarking approach.