OpenAI’s Invisible Text Watermark Comes to ChatGPT and Codex — With Limits

OpenAI is preparing to introduce an invisible text watermark called textGrain for eligible ChatGPT and Codex users in the European Union. The system is designed to embed a provenance signal into AI-generated text by subtly influencing the statistical pattern of words and word fragments used in a response.

There is no visible badge, warning or disclaimer attached to the finished text. To a reader, the output should look like ordinary writing.

The rollout is tied to the European Union’s AI Act and its growing focus on transparency around AI-generated content. OpenAI isn’t the only AI company exploring this approach. Anthropic has also announced a watermarking system for Claude that uses a similar statistical principle to Google DeepMind’s SynthID text technology.

But while AI text watermarking could make it easier to identify machine-generated content, it is far from a foolproof solution.

How Does OpenAI’s textGrain Watermark Work?

textGrain doesn’t add a hidden image, special character or visible marker to the text. Instead, the watermark is embedded in the way the model chooses words.

When generating a passage, the system can subtly adjust the probability of selecting certain words or word pieces. Individually, those choices are difficult to notice. Across a sufficiently long passage, however, they can form a statistical pattern that a specialized detector can analyze.

OpenAI describes the watermark as being embedded directly into the wording rather than appearing as a separate element that readers can see.

That distinction is important. You won’t be able to look at a ChatGPT response and identify the watermark yourself. Detection requires a separate system designed to analyze the statistical characteristics of the text.

Who Can Access the textGrain Detector?

At launch, OpenAI isn’t making the detector available as a public tool.

Access is expected to be limited to approved researchers and expert organizations on a case-by-case basis. OpenAI has pointed to the potential consequences of false positives and missed detections as a reason for restricting access.

In other words, the technology is intended to provide evidence about the origin of text, but OpenAI doesn’t want people treating an automated detector as an unquestionable authority.

What Does textGrain Mean for ChatGPT Users?

The impact of the watermarking system depends on how you use OpenAI’s products.

ChatGPT and Codex users in the EU

Eligible users may receive generated text containing the invisible provenance signal. Nothing obvious will appear in the response, but the text may be detectable by an authorized textGrain system later.

API developers worldwide

For developers using the OpenAI API, watermarking is available on selected models as an opt-in feature. This could help developers communicate the use of AI-generated content within their own applications and services.

Researchers and expert organizations

Organizations that need to study or evaluate AI-generated text can apply for access to the detection system. However, this isn’t the same as having a public AI-text checker that anyone can use.

Educators, employers and publishers

This group may need to be particularly careful with watermark detection.

A positive result should be treated as one piece of evidence rather than automatic proof that someone cheated, used AI improperly or didn’t write their own work. A detection system can provide a useful signal, but it cannot establish the entire history of how a piece of writing was created.

OpenAI’s Own Testing Shows the Limits

The most important part of textGrain may not be the watermark itself, but the limitations OpenAI acknowledges around detection accuracy.

According to a Unite.AI summary of OpenAI’s reported evaluations, detection reached approximately 80% for 200-token passages and around 95% for 400-token passages at a 1% false-positive rate under specific testing conditions.

Those numbers need context. Detection performance can change depending on the length and subject matter of the text, so they shouldn’t be interpreted as universal accuracy rates.

OpenAI also acknowledges that textGrain does not guarantee reliable detection.

That caveat becomes particularly important when the text has been edited.

Even Small Changes Can Weaken the Watermark

AI text watermarks aren’t necessarily permanent once the generated text leaves the model.

According to the same reported evaluation, replacing approximately 10% of the words with synonyms reduced detection for 400-token passages from about 92% to 66% under the test conditions.

Replacing around 25% of the words reduced detection much further, to roughly 17%.

That demonstrates a fundamental weakness in statistical watermarking: changes to the wording can interfere with the underlying pattern the detector is looking for.

The practical result is that the watermark may work best on relatively untouched AI-generated text. Once people substantially rewrite or edit that material, detection can become considerably more difficult.

What textGrain Cannot Tell You

Even when a detector identifies a watermark, there are important questions it cannot answer.

A textGrain result cannot:

• Verify whether the information in the text is accurate.
• Identify the specific person who generated the content.
• Establish ownership of the writing.
• Determine exactly how much of the text was produced by AI.
• Measure how extensively a human edited the original output.
• Prove that a person did not write the text when no watermark is detected.

That last point is especially important.

A failed detection doesn’t automatically mean the text was written entirely by a human. Likewise, a positive detection doesn’t necessarily tell you how the final piece of writing was produced.

At most, a positive result can indicate that an OpenAI system may have generated or processed some of the text.

OpenAI Isn’t the Only Company Exploring AI Watermarks

OpenAI’s move is part of a broader effort across the AI industry to establish ways of identifying machine-generated content.

Anthropic is pursuing a similar concept for Claude, while Google DeepMind has developed SynthID technology for watermarking AI-generated content.

However, these systems aren’t interchangeable.

A watermark created by one AI provider requires the appropriate detection technology from that provider. Anthropic’s watermark doesn’t automatically become detectable by OpenAI’s system, and vice versa.

That means the industry still doesn’t have a single, universal AI-text detector capable of reliably identifying content from every major AI model.

AI Watermarks Are Evidence, Not Proof

textGrain could make it easier for researchers and organizations to investigate the provenance of AI-generated text, particularly when the original output remains relatively unchanged.

But its limitations are just as important as its capabilities.

Detection accuracy depends on factors such as passage length and how much the text has been modified. The system can’t establish who wrote a passage, whether its claims are accurate or how much human input went into the final version.

For that reason, a watermark result should be treated like one piece of evidence rather than a final verdict.

OpenAI’s textGrain represents another step toward greater transparency around AI-generated content. But for now, the technology is better understood as a provenance signal—not a definitive AI lie detector.