OpenAI adds invisible watermarks to ChatGPT text in the EU

Detection rates were substantially lower for math content according to OpenAI. The model has less freedom to choose words when solving math problems. Longer passages might strengthen the statistical signal and improve detection, though results could still depend on the content. OpenAI provides no data to test that assumption about longer texts in math.

At 400 tokens, roughly 300 words, the detector identifies about 94 percent of watermarks in psychology passages. However, it only identifies about 60 percent of watermarks in math passages. This difference shows how subject matter affects the effectiveness of the watermark. Math content is harder to watermark than creative writing like psychology essays.

Editing the text also makes the watermark much harder to detect. OpenAI says replacing just 10 percent of words with synonyms cuts detection rates for 400-token passages from about 92 percent to 66 percent. Replacing a quarter of the words drops detection to 17 percent, making the watermark easy to evade. That weakness could work in OpenAI's favor regarding user privacy concerns.

Even limited editing sharply reduces detection rates significantly. Replacing 25 percent of the words pushes detection below 20 percent, even in 400-token passages. This suggests that users can easily remove watermarks by making small changes to their text. Users who do not want their ChatGPT use detected might switch to open-weight models if its watermarks become harder to remove.

A technical report explains how textGrain works in depth for engineers. Anthropic has not published detection rates for Claude's watermark yet. However, it says its system holds up well against edits and other attacks. OpenAI will release the full details of textGrain as open source soon. This allows researchers to study the method and find potential vulnerabilities.

OpenAI says watermarking does not hurt output quality in their tests. They cite tests of its frontier model Astra to support this claim. The company reports no significant performance differences with watermarking on or off across eight benchmarks. These benchmarks include GPQA Diamond, BrowseComp, and DeepSWE. Those results do not establish whether watermarking affects writing quality specifically.

Critics have also questioned whether Claude's watermark hurts the quality of text. OpenAI acknowledges this uncertainty but relies on their internal testing for Astra. Writing quality is a different metric from benchmark scores like GPQA or DeepSWE. Users might still worry about subtle changes in tone or style caused by the watermark.

OpenAI says a detected watermark reveals nothing about human creativity or editing. It does not establish ownership or assign responsibility to any specific user. The tool cannot identify who wrote the text or show if it is accurate. Failing to detect a watermark does not prove a human wrote the text either. The passage could be too short, have been edited, or come from an unsupported model.

OpenAI will limit detector access for now to only selected researchers and specialist organizations. They can apply through an application form on OpenAI's website. OpenAI will grant access case by case under the EU's Code of Practice. Anthropic takes a similar approach with its detection API for users who need it.

The tool will report only whether it detected an OpenAI watermark. It won't identify users or reveal their prompts or conversations. This privacy protection is crucial for organizations that use AI for internal tasks. Existing verification tools for images and audio remain publicly available on openai.com/verify. The Content Provenance API also helps verify media authenticity without revealing user data.

OpenAI says it is restricting access because the detector can flag unmarked text or miss watermarks. The company plans to expand access when they believe results can be interpreted responsibly. It has not said when that expansion might happen yet. Engineers must wait for official announcements before building custom detection tools.

Existing verification tools for images and audio remain publicly available for comparison. Users can check how image and audio watermarking compares to text watermarking. The Content Provenance API provides a standard way to verify content provenance across different media types. OpenAI's textGrain aims to bring similar reliability to written text.

OpenAI is adding invisible watermarks to ChatGPT text in the EU but makes it optional globally. This change affects how companies use AI models for compliance and transparency. Engineers need to understand detection rates before building monitoring systems. Managers must decide if the cost of watermarking fits their budget.

The technology uses statistical signals hidden in word choices rather than visible marks. Detection depends on text length and subject matter like math or psychology. Editing text significantly reduces the ability to detect watermarks. Quality tests show no major impact on benchmark scores.

Privacy limits mean the detector reveals nothing about user identity or prompts. Access is restricted to researchers until OpenAI feels it is safe to open up. Engineers should test their own workflows with synthetic data before relying on the tool. Managers can weigh the compliance benefits against potential evasion risks.

OpenAI's decision to make global API watermarking optional creates a split market. EU users get mandatory labeling while global users have choices. This difference highlights the growing influence of the EU AI Act on tech companies. Competitors like Anthropic are taking different approaches to meet similar regulatory demands.

Engineers should monitor detection rates as OpenAI releases more data on textGrain. They can compare these numbers with SynthID performance from Google's open-source project. Testing with short and long texts will show how length affects reliability. Math content remains a difficult area for current watermarking techniques.

Managers might consider the cost implications of enabling or disabling watermarks globally. Some organizations may prefer to avoid any potential detection risks entirely. Others might want the transparency that watermarks provide for public-facing applications. The opt-in model gives flexibility but requires active management decisions.

OpenAI plans to release textGrain as open source to foster community improvement. This strategy mirrors how many other AI technologies evolve through collaboration. Researchers can contribute fixes or enhancements to make the system more robust. The company hopes this will build trust among developers and users alike.

The EU's Code of Practice guides OpenAI in granting access to the detector. This framework ensures that access is given responsibly and securely. Anthropic follows a similar path with its own detection API tools. Both companies recognize the need for careful handling of AI transparency tools.

Detection rates vary widely between psychology passages and math passages at 400 tokens. Psychology gets about 94 percent detection while math gets only about 60 percent. This gap suggests that creative writing is easier to watermark than logical reasoning. Math problems have fewer valid word choices which limits the signal strength.

Replacing 25 percent of words drops detection below 20 percent in most cases. Even small edits can undermine the effectiveness of textGrain significantly. Users who want anonymity might find this feature less useful than expected. OpenAI acknowledges this limitation but relies on internal testing for quality assurance.

OpenAI's frontier model Astra shows no performance drop with watermarking enabled. Eight benchmarks tested to confirm that writing speed and accuracy remain stable. Critics argue that benchmark scores do not capture all aspects of writing quality. The impact on nuanced style or emotional tone remains unproven by these tests.

A detected watermark reveals nothing about human input or editing history. It cannot prove that a human wrote the text or edited it significantly. Failing to detect a watermark also does not confirm human authorship. Multiple factors like length and model support influence detection success rates.

OpenAI restricts detector access until they feel ready for broader release. Only selected researchers and organizations can apply through an application form now. This cautious approach protects against misuse while allowing expert scrutiny. The company will expand access when they believe results can be interpreted responsibly.

Engineers should test their own models with synthetic watermarks before relying on OpenAI's tool. They can build custom detectors to compare performance against textGrain. Understanding detection limits helps in designing better AI safety and compliance strategies. Managers need to balance transparency goals with operational flexibility.

The optional global watermark policy creates a complex landscape for API users. Some customers might disable the feature to avoid any potential detection flags. Others might keep it on for public trust or legal compliance reasons. OpenAI expects this choice to drive market differentiation among service providers.

OpenAI's textGrain technology represents a significant step forward in AI transparency. It combines statistical analysis with practical deployment considerations like edit resistance. The system aims to meet regulatory requirements without sacrificing user experience too much. Future iterations might improve detection rates across different content types.

Managers should evaluate the cost-benefit ratio of enabling watermarks for their specific use cases. Compliance risks in the EU are high, but global markets offer more flexibility. OpenAI's opt-in model allows organizations to tailor their approach to local laws. Engineers need to document how they handle AI-generated content for audit purposes.

OpenAI will continue to refine textGrain based on community feedback and research. The open-source release will invite scrutiny from the broader AI development community. This transparency builds confidence in the technology's reliability and safety standards. Users who wait for more data might make better informed decisions about adoption.

Source: https://the-decoder.com/openai-will-watermark-chatgpt-text-in-the-eu-but-makes-it-optional-for-api-users-worldwide/

Reflection's Beam beats Chinese models with less compute

Reflection released Beam, an open-weight model that matches GLM 5.2 and Qwen 3.8 on benchmarks while using three to four times less compute.