ChatGPT texts are getting invisible watermarks: What does textGrain mean for marketing and students?

AI-generated text is becoming harder to distinguish from human writing. Since October 2026, OpenAI has been gradually introducing invisible watermarks for eligible ChatGPT text outputs in the European Union. The technology is called textGrain and is intended to help identify certain AI-generated passages after creation. How does it work, and what does it mean for online marketing, copywriters and students?

What is textGrain?

OpenAI developed textGrain as a technology for invisibly watermarking AI-generated text.

It does not add a visible logo. Nor does its watermark consist of hidden characters or extra information stored in the text file. Instead, the system subtly influences the statistical selection of words or word pieces, known as tokens, during generation.

Across a passage, these choices create a characteristic pattern that specialised detection tools can look for.

The text normally remains readable as usual. Its statistical pattern can also survive copying into Word, Google Docs or a content management system.

On 5 October 2026, OpenAI announced the gradual introduction of textGrain for eligible ChatGPT and Codex text outputs within the European Union. For selected API models, watermarking is optional worldwide and disabled by default.

Important: Not every ChatGPT text automatically contains a watermark. The EU rollout does not mean all users or outputs worldwide are covered.

How reliably can AI-generated text be identified?

OpenAI has evaluated textGrain’s detection performance.

In certain controlled tests with English-language passages, including psychology-related material, it reported the following results at a target false-positive rate of 1%:

• Around 80% of watermarks were detected in passages of approximately 200 tokens.
• Detection reached approximately 95% at around 400 tokens.

These figures are not universal detection rates for all AI-generated text. Tokens are not the same as words.

Short passages, constrained wording and subsequent editing can make detection harder. Results were substantially lower for mathematical content with less flexibility in word choice.

Translation or extensive rewriting can also weaken the signal. False positives are possible, as are missed watermarks.

Failure to detect a watermark therefore does not establish that a human wrote the text.

The textGrain detector is initially not publicly available to everyone. Access is restricted to approved researchers and specialist organisations.

What does this mean for online marketing and copywriters?

textGrain adds a way to investigate the provenance of certain AI-generated content.

For copywriters and agencies, this does not make using ChatGPT inherently problematic.

Quality, factual accuracy and editorial responsibility remain essential.

The same distinction matters for search engine optimisation: Google does not automatically penalise content simply because AI helped create it. Low-value content produced at scale primarily to manipulate search rankings is particularly problematic.

An invisible watermark is therefore not an automatic SEO disadvantage. This is not a ranking guarantee, nor a claim that Google technically checks textGrain signals.

Businesses should nevertheless be able to track which content was created with AI assistance and subsequently reviewed by an editor.

Can universities identify AI-generated assignments?

textGrain is particularly relevant to students and universities.

ChatGPT is used for research, structuring ideas and improving the language of academic work.

Under certain conditions, invisible watermarks could help indicate whether a supported AI system generated a passage.

There are substantial limitations, however.

Not all AI providers use the same technology. Editing can change watermark signals, and a signal may be absent altogether.

The method therefore cannot reliably establish that an entire assignment was produced by AI.

Equally, a missing watermark does not prove every sentence was written by a human. A detected signal does not identify a person, account or prompt, and does not measure the human contribution.

For universities: A technical indication should not be treated as the sole evidence of academic misconduct.

Students need to follow their own institution’s rules on generative AI. Some institutions may permit language editing but require documentation of the tools used.

Those requirements apply regardless of whether an AI watermark can be detected. Academic rules differ between institutions and countries; this article does not describe a universal university policy.

Why is OpenAI introducing watermarks now?

The transparency requirements of the European AI Act are a significant part of the context.

Article 50 includes requirements for machine-readable marking of certain AI-generated content.

The relevant transparency obligations generally apply from 2 August 2026. For certain generative AI systems already placed on the market before that date, Article 111(4) provides a transition period until 2 December 2026 for compliance with Article 50(2).

textGrain is one technical approach OpenAI is using to support these requirements.

It is important to distinguish technical marking by AI providers from disclosure obligations that may apply to people or organisations using AI.

An AI-assisted advertisement, blog post or academic text does not automatically fall under an identical visible-labelling duty. Article 50(4) covers, among other things, AI-generated text published to inform the public on matters of public interest. Its statutory exceptions, including those concerning human review and editorial responsibility, must be considered.

This section concerns EU law. It is not a worldwide labelling rule, and its application should not be assumed to be identical in the UK, the US, Switzerland or other jurisdictions. Other local laws and institutional requirements may also apply.

Conclusion: More traceability does not mean complete verification

With textGrain, OpenAI is introducing a technology that can invisibly mark certain AI-generated passages.

It creates additional possibilities for online marketers, copywriters and universities to investigate the origin of text.

Its technical limitations remain significant.

A watermark can provide an indication. It is neither universal proof of AI use nor a reliable way to identify every AI-generated text.

For businesses and students, the priorities remain responsible AI use, careful review of content and compliance with applicable requirements.