OpenAI is adding an invisible watermark to AI-generated text

OpenAI announced on Monday it is adding an invisible text watermark, called textGrain, to text generated by ChatGPT and programming agent Codex, a change that is mandatory in the EU and opt-in for API customers elsewhere. The watermarking responds to Article 50(2) of the EU AI Act, which requires providers to mark AI outputs in machine-readable form; providers that released models before Aug. 2 have until Dec. 2 to comply or face fines of up to 3 percent of worldwide annual turnover. OpenAI's detection tool will be limited to approved researchers and experts, and in testing it caught watermarks in only 80 percent of 200-word AI-generated passages, with detection falling to 66 percent after 10 percent of words in a 400-word passage were changed and to 17 percent after 25 percent were replaced.

OpenAI is adding an invisible watermark to AI-generated text Amanda Yeo https://mashable.com/author/amanda-yeo Read Full Bio https://mashable.com/author/amanda-yeo OpenAI https://mashable.com/category/openai is adding an invisible watermark to text generated by ChatGPT https://mashable.com/category/chatgpt and programming agent Codex, making it easier for machines to identify that it was written by AI. Announced on Monday https://openai.com/index/eu-text-provenance/ , the change will be mandatory in the EU, and is currently opt-in for API customers across the rest of the world. Large language models work by generating text one word at a time, determining what the next will be by looking at the ones that came before it. OpenAI's new text watermarking system textGrain https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf influences this by changing how AI models choose words, having them consider both preceding text and pseudorandom values determined by a secret key. This leaves a statistical signal in the generated content, allowing AI detection machines to determine what these pseudorandom values are and establish whether the text is dependent upon them. If it is, then it is flagged as AI-generated. You May Also Like This Tweet is currently unavailable. It might be loading or has been removed. https://twitter.com/OpenAI/status/2107164651478012412 OpenAI's text watermarking will roll out across ChatGPT and Codex in the EU over the next few weeks. It is also available to API customers globally across select models, though they'll have to opt in as it's off by default. However, OpenAI's text watermark detection tool will only be available to researchers and experts who apply for access. OpenAI stated that this restriction is partly due to limitations in its text watermarking, particularly concerning shorter passages. In testing, the tool only caught watermarks in 80 percent of 200-word AI-generated passages, with detection rates even lower when said text was on topics with rigid lexicons such as mathematics. Editing AI-generated text can also significantly reduce the efficacy of OpenAI's detection tool. The company stated that changing just 10 percent of words in a 400-word passage reduced detection to 66 percent, while replacing 25 percent of words lowered it even further to 17 percent. Terms of Use https://www.ziffdavis.com/terms-of-use and Privacy Policy https://www.ziffdavis.com/ztg-privacy-policy . "Text watermarking and detection remain early technologies with significant limitations, and views about their benefits and responsible uses are still developing," OpenAI said in its announcement. OpenAI further stressed that text watermarks cannot establish ownership of generated content, prove its accuracy, or identify the person who prompted it. The absence of a watermark isn't proof that content has been created by humans either. OpenAI's new text watermarks are in response to the EU's Artificial Intelligence Act. Originally signed in 2024 https://mashable.com/article/eu-ai-law , the regulation aims to "protect fundamental rights, democracy, the rule of law, and environmental sustainability from high-risk AI," according to a press release at the time. The EU AI Act has gradually come into force https://artificialintelligenceact.eu/implementation-timeline/ since 2025, with most of the articles now applicable as of Aug. 2 this year. This includes Article 50 2 , https://artificialintelligenceact.eu/ai-act-explorer/ article-50 which requires AI system providers to "ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated." Providers such as OpenAI who released their AI models before Aug. 2 have until Dec. 2 to comply with Article 50 2 , or could be fined up to three percent of their worldwide annual turnover. Last week, Reuters reported https://www.reuters.com/technology/openais-annual-recurring-revenue-nears-70-billion-axios-reports-2026-09-29/ that OpenAI's annualised recurring revenue is nearing $70 billion though they also acknowledged that such calculations can be misleading https://techcrunch.com/2026/05/22/how-vcs-and-founders-use-inflated-arr-to-kingmake-ai-startups/ . As such, a violation of Article 50 2 could potentially net OpenAI with a fine of over $2 billion. Earlier this year, Anthropic also announced that its AI chatbot Claude would watermark generated content https://mashable.com/tech/claude-watermark-ai-generated-content in accordance with the EU's rules. Unlike OpenAI, Anthropic's watermarks are being made available worldwide to all users, not just to those in the EU. Disclosure: Ziff Davis, Mashable’s parent company, in April 2025 filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems. Topics Artificial Intelligence https://mashable.com/category/artificial-intelligence ChatGPT https://mashable.com/category/chatgpt OpenAI https://mashable.com/category/openai