Translating into Español. The English version is shown until it is ready.
OpenAI to add text watermarks to ChatGPT and Codex output in EU
Versión 3: New reports add the technical detail of how the watermark works and where it is applied, including that API output is off by default with institutions able to enable it, that support is planned for cloud-partner models, that the EU AI Act transparency rules took effect on August 2, that OpenAI plans to open-source the technology, and that detection rates fall to about 60% for maths texts; coverage also adds the debate over whether watermarking can distinguish AI-written from AI-assisted text, and notes that Korea has its own display duties under its AI Framework Act.
En breve
OpenAI said it will introduce a text watermarking technology called Textgrain to mark AI-generated writing, applying it to ChatGPT and Codex output in the European Union within weeks to meet the EU AI Act’s transparency rules. The company said the mark causes no meaningful difference in answer quality or readability, but detection falls sharply after editing — from 92% to 66% when one in ten words is swapped for a synonym, and to 17% when one in four is. The detection tool will go first to approved researchers and specialist institutions, and OpenAI also announced a pilot of visual ads for some US advertisers in ChatGPT’s image generation.
Leer la noticia completa 3 min de lectura
OpenAI said it will introduce a text watermarking technology called Textgrain to identify writing produced by artificial intelligence, responding to the transparency obligations for generative AI under the European Union’s AI Act. The company announced the plan on October 5 local time, and said the feature will be applied to ChatGPT and Codex output in the EU within the next few weeks, according to Yonhap, Dong-A Ilbo, Chosun Ilbo, AI Times, Electronic Times and Kyunghyang Shinmun. [ 1 , 2 , 3 , 4 , 5 , 6 ]
According to the reports, Textgrain inserts a subtle statistical signal while the model selects words, rather than embedding hidden characters or special symbols, leaving traces invisible to the eye that dedicated software can read; Dong-A Ilbo and AI Times said the signal is embedded in the text itself. OpenAI said the mark produces no substantive difference in the quality, performance or readability of answers. AI Times added that OpenAI said the signal partly survives copying and pasting, and that the company plans to release Textgrain as open source. [ 1 , 2 , 4 , 6 ]
The reports set out the technology’s limits. OpenAI said detection rates fall for short texts or content such as mathematical formulas, where word choice is less flexible, and that post-editing weakens the signal. Yonhap, AI Times, Electronic Times and Kyunghyang reported that in a 400-token passage, replacing one in ten words with synonyms cut the detection rate from 92% to 66%, and replacing one in four pushed it to about 17%. Kyunghyang reported detection of about 80% for 200-token psychology texts and about 95% for 400-token texts, but only about 60% for 400-token maths texts; AI Times reported roughly 80% at 200 tokens and about 95% at 400 tokens with a 1% false-positive setting. [ 1 , 2 , 4 , 5 , 6 ]
Detection will not be openly available at first. Yonhap, AI Times and Electronic Times reported that the detector will be provided first to approved researchers and specialist institutions that apply for access, and AI Times said it reports only whether a watermark is present, not users, accounts, prompts or conversation content. For developer API output the watermark is off by default, but institutions can enable it, and AI Times said support is planned for models delivered through cloud partners. OpenAI said a watermark does not show how much a human was involved, cannot determine ownership or legal responsibility, and that its absence does not prove human authorship, according to AI Times and Kyunghyang. [ 1 , 4 , 5 , 6 ]
The move follows similar steps by other developers. AI Times said Anthropic applied watermarks to Claude output globally from August, while Electronic Times described the Anthropic measure as applying in the EU, and Kyunghyang said Anthropic introduced text watermarking in August after user complaints that human-written text edited with Claude could be treated as AI-written. Kyunghyang reported that Google has applied SynthID to Gemini since 2024; Yonhap and Electronic Times noted the comparable SynthID-Text feature. AI Times said the EU AI Act transparency rules have applied since August 2. [ 1 , 4 , 5 , 6 ]
Electronic Times said attention is now on whether the measure spreads to countries with display obligations, including Korea, which has such duties under its AI Framework Act. Kyunghyang reported that schools and companies will have to decide how much weight to give detection results, and quoted Kim Myung-joo, head of the AI Safety Research Institute at ETRI, saying technical accuracy has limits, so a watermark alone cannot be used to conclude that AI wrote a text or to impose a penalty, though the mere possibility of inspection can have a deterrent effect and help screen suspect content. [ 5 , 6 ]
In the same announcement, Yonhap reported, OpenAI unveiled an advertising policy: a pilot test of visual ads aimed at some US advertisers when the image generation function is used inside ChatGPT. OpenAI said the ads would not influence ChatGPT answers and that conversation content would remain private. [ 1 ]
Por qué importa
OpenAI is the latest major developer to add machine-readable marking to AI text, after Anthropic and Google, in response to EU transparency obligations; Korean outlets note domestic display rules under the AI Framework Act and are watching whether the measure spreads beyond the EU. How reliably the mark survives editing will shape how far schools, companies and regulators can treat it as evidence. OpenAI’s simultaneous advertising pilot shows it testing new revenue inside ChatGPT while saying ads will not affect answers.
Datos clave
- OpenAI said it will adopt a text watermarking technology called Textgrain to identify AI-generated writing, in response to the EU AI Act’s transparency obligations for generative AI. [ 1 , 2 , 3 , 4 , 5 , 6 ]
- The feature will be applied to ChatGPT and Codex output in the EU within the next few weeks, OpenAI said. [ 1 , 2 , 3 , 4 , 5 , 6 ]
- Textgrain inserts a subtle statistical signal while the model chooses words, rather than hidden characters or special symbols, and is invisible to the eye but detectable with dedicated software, according to the reports. [ 1 , 2 , 4 , 6 ]
- OpenAI said the watermark causes no substantive difference in answer quality, performance or readability. [ 1 , 2 , 4 , 5 ]
- OpenAI said detection rates fall when text is edited: replacing one in ten words with synonyms cut the rate from 92% to 66%, and replacing one in four cut it to 17%. [ 1 , 2 , 4 , 5 , 6 ]
- Detection is also weaker for short texts and for content such as mathematics, where word choice is less flexible; Kyunghyang reported a rate of about 60% for 400-token maths texts. [ 1 , 4 , 5 , 6 ]
- The detection tool will first be provided to approved researchers and specialist institutions that apply for access, according to the reports. [ 1 , 4 , 5 , 6 ]
- OpenAI also announced a pilot of visual advertising for some US advertisers when the image generation function is used inside ChatGPT, saying ads would not affect answers and conversations would remain private. [ 1 ]
Confirmado por varias fuentes
- OpenAI said it is introducing a text watermarking technology called Textgrain for AI-generated writing, citing the transparency obligations of the EU AI Act. [ 1 , 2 , 3 , 4 , 5 , 6 ]
- The feature will be applied to ChatGPT and Codex output in the EU within weeks, OpenAI said. [ 1 , 2 , 3 , 4 , 5 , 6 ]
- OpenAI said the watermark makes no substantive difference to answer quality, performance or readability. [ 1 , 2 , 4 , 5 , 6 ]
- OpenAI said detection weakens after editing: 92% to 66% when 10% of words are replaced by synonyms, and to 17% when 25% are replaced. [ 1 , 2 , 4 , 5 , 6 ]
- Detection rates are lower for short texts and for content such as mathematics, where word choice is less flexible, OpenAI said. [ 1 , 4 , 5 , 6 ]
- The detection tool will first be given to approved researchers and specialist institutions; use for API output is off by default but institutions can enable it. [ 1 , 4 , 5 ]
- Anthropic and Google have applied comparable AI content identification features — Anthropic in August and Google’s SynthID earlier — according to the reports. [ 1 , 4 , 5 , 6 ]
Aún sin aclarar
- Whether and when the watermark will be applied outside the EU. OpenAI is quoted as saying there will be no global default rollout at the EU launch; Electronic Times says attention is on whether it spreads to countries such as Korea, which has display duties under its AI Framework Act, but no schedule is given.
- Whether Anthropic’s watermark applies worldwide or only in the EU. AI Times says Anthropic applied it globally from August, while Electronic Times says Anthropic applied it in the EU region from August; the documents do not agree.
- The exact extent, duration and participants of the advertising pilot. Only Yonhap reports the pilot, and it gives no further detail.
- The size of the text used in the detection-rate example. Yonhap and AI Times describe a 400-token passage while Dong-A Ilbo describes roughly 1,600 English characters; the documents do not reconcile the two.
- Whether a detected watermark means the text was written by AI. OpenAI says the watermark cannot show how much a human was involved, ownership or legal responsibility, and that its absence does not prove human authorship; Kyunghyang reports it cannot distinguish fully AI-written text from human text that AI partly touched.
- When approved researchers and institutions will gain access to the detection tool. The reports say the tool will be provided first to approved applicants but give no date; OpenAI says access may be widened as the technology develops.
- Whether the technology will meet the transparency requirement in practice. The documents state the legal obligation and OpenAI’s stated compliance, but no regulator assessment is reported.
Qué dicen los medios locales
Cronología, hora local
- OpenAI announces the introduction of Textgrain text watermarking in response to the EU AI Act, along with a pilot of visual ads in ChatGPT image generation, according to the documents. [ 1 , 2 , 3 , 4 , 5 , 6 ]
- Yonhap publishes its report, datelined San Francisco, on the watermarking plan and the advertising pilot. [ 1 ]
- Dong-A Ilbo publishes its report on the watermark for EU ChatGPT and Codex output. [ 2 ]
- Chosun Ilbo publishes its report on the watermark for text generated by ChatGPT and Codex. [ 3 ]
- AI Times publishes its report detailing the technology, benchmarks and API policy. [ 4 ]
- Electronic Times publishes its report, noting Korean display obligations and the question of wider rollout. [ 5 ]
- Kyunghyang Shinmun publishes its report on the limits of watermarking and the debate over its use. [ 6 ]