OpenAI is adding text watermarking in ChatGPT and Codex

OpenAI has begun embedding an invisible, machine‑readable watermark into the text generated by its flagship models, ChatGPT and Codex, marking a significant step in the tech company’s effort to comply with emerging regulatory frameworks in Europe. The new feature, dubbed “textGrain,” is being rolled out initially to users in the European Union, where the AI Act – a comprehensive set of rules governing artificial intelligence – is set to become law later this year. By embedding a subtle signal that can be detected by specialized tools, OpenAI hopes to provide a method for downstream platforms, researchers and regulators to verify whether a piece of text was produced by an AI system, without altering the visible content for readers.OpenAI’s announcement positions textGrain as a direct response to the EU’s requirement that “high‑risk” AI systems be equipped with mechanisms that enable traceability and attribution..

The company says the watermark is designed to be “invisible to human eyes” while remaining detectable by automated analysis, allowing auditors to confirm the origin of a text snippet even after it has been edited or republished. According to the firm’s technical brief, the watermark is embedded at the token‑level during generation, using a probabilistic approach that subtly biases the choice of words in a way that is statistically significant yet imperceptible to users.The rollout follows a broader industry trend toward watermarking and provenance tools for generative content. Google DeepMind’s SynthID, introduced earlier this year, embeds a similar signal in large‑language‑model outputs, and Anthropic announced its own watermarking system in August, both citing the EU AI Act as a catalyst..

OpenAI claims that its textGrain “matched or exceeded” the detection rates of these earlier approaches in internal benchmark tests. In side‑by‑side evaluations, watermarked text was identified with a true‑positive rate of over 95 percent, while false‑positive rates on human‑written text remained below 2 percent, figures that the company says are comparable to, or better than, competing solutions.OpenAI has also released a suite of open‑source tools that enable third parties to scan documents for the presence of the watermark. The detection algorithm, made available on GitHub, can be run locally or integrated into larger compliance pipelines..

OpenAI stresses that the tools are “designed for transparency” and do not require any special access to the model’s internal weights, a move that analysts see as an effort to build trust with both regulators and the broader public. The company’s blog post notes that the detection process is computationally lightweight, taking only a few milliseconds per kilobyte of text on standard hardware.Despite the technical promise, the introduction of mandatory watermarking has sparked debate among privacy advocates, developers and content creators. Critics argue that the presence of a hidden identifier could be misused to track or censor user‑generated content, especially if combined with large‑scale data‑mining efforts..

Others worry that the watermark could be stripped or spoofed, undermining its purpose. OpenAI has responded by emphasizing that the watermark is not a “digital signature” that can be used to attribute authorship in legal contexts, but rather a probabilistic marker intended solely for detection of AI‑generated text.OpenAI’s own performance metrics suggest that the watermark does not meaningfully degrade the quality of the generated output. In a set of benchmark tests spanning language fluency, factual accuracy and code generation, the company reported no statistically significant difference between watermarked and unwatermarked text..

For Codex, the code‑generation model, the team measured the same pass‑rate on standard programming problems, indicating that the subtle token‑level adjustments do not impair the model’s ability to produce correct and efficient code. OpenAI, however, cautions that the watermark “does not guarantee” that every piece of AI‑generated text can be detected, acknowledging the possibility of false negatives under certain conditions, such as heavy post‑processing or translation.The EU’s AI Act, which aims to impose strict obligations on high‑risk AI systems, includes a provision that requires providers to embed “technical measures” that enable the attribution of AI‑generated content. While the legislation is still in the final stages of approval, companies are already positioning themselves to meet the forthcoming requirements..

OpenAI’s early deployment in the EU is therefore both a compliance measure and a strategic move to set a de‑facto standard that could influence regulations in other jurisdictions. Industry observers note that the company’s approach could pressure rivals to adopt similar watermarking schemes, potentially leading to a fragmented landscape of proprietary detection tools.From a business perspective, the addition of watermarking may also affect the way developers integrate OpenAI’s APIs into third‑party applications. Platforms that rely on user‑generated content—such as social media sites, educational tools and content‑moderation services—will need to account for the new signal in their pipelines..

OpenAI has indicated that the detection API will be free for non‑commercial use, but enterprise customers may be offered premium support and higher‑throughput detection services. The company’s pricing page now lists “compliance add‑ons” as optional features for large‑scale deployments, hinting at a future revenue stream tied to regulatory adherence.The broader implications of AI watermarking touch on the ongoing conversation about the responsibility of AI developers to mitigate misuse. By providing a built‑in method for identifying synthetic text, OpenAI aims to make it harder for malicious actors to pass off AI‑generated disinformation as human‑authored content..

Yet the effectiveness of such technical safeguards depends heavily on adoption by downstream platforms and the willingness of regulators to enforce detection requirements. As the EU prepares to enforce its AI Act, the coming months will likely see a surge in both technical innovation and legal scrutiny surrounding watermarking technologies.OpenAI’s rollout of textGrain underscores a pivotal moment in the maturation of generative AI: the shift from experimental novelty to regulated utility. As the company expands the watermark to other markets beyond the EU, it will be watching closely how detection tools are integrated into real‑world workflows and whether the balance between transparency and privacy can be maintained..

The success of this initiative may well shape the future of AI governance, setting benchmarks for how AI‑generated content can be responsibly identified without compromising the user experience that has made large‑language models so popular worldwide..