Have you ever wondered how to tell whether a piece of text was written by a human or generated by AI? In a world where ChatGPT can write emails, essays, and even novels in seconds, that question is no longer just a philosophical pondering. It has become a legal issue, a technical challenge, and a battle over the control of information.
And OpenAI has just made its first move in that battle. The company has officially announced that it will roll out text watermarking—an invisible fingerprint embedded into every piece of text generated by ChatGPT and Codex—for users in the European Union. This is not just a technical update. It is a declaration about the future of transparent AI.
The EU AI Act: When Transparency Becomes Law
To understand why OpenAI has to do this, we need to look back at the legal context. The EU AI Act—specifically Article 50 on transparency obligations—took effect on August 2, 2026. This provision requires providers of generative AI systems to ensure that machine-generated content can be machine-readable.
In other words, if you use ChatGPT to write an article, anyone's system—from regulatory agencies to social media platforms—must be able to detect that the text was not written by a human. This is an unprecedented requirement, and it poses a massive technical challenge for all AI companies.
AI systems already on the market before August 2 have until December 2, 2026, to comply. This explains why OpenAI, Anthropic, and other companies are rushing to deploy their own watermarking solutions. And that is also why August 2026 marked a turning point as Anthropic became the first company to announce a global watermark for Claude.
textGrain: How OpenAI's Watermarking Technology Works
OpenAI calls its watermarking system textGrain. It sounds simple, but behind that name lies sophisticated technology developed in collaboration with researchers from the University of Pennsylvania and Yale.
Instead of inserting a hidden character or identification code into the text—things that can be easily detected and removed—textGrain works by subtly altering how the model selects words. Imagine the AI model predicting the next word in a sentence. It has hundreds of different choices, all grammatically and semantically plausible. textGrain uses a secret key to rearrange those choices in a certain order, and the model selects the word according to that order.
When hundreds of such small choices accumulate, they form a hidden statistical signal invisible to the human eye, but recognizable to a detection tool equipped with the secret key. And because the trace lies within the words themselves, it survives even when the text is copied and pasted to another platform.
OpenAI states that textGrain is "on par with or superior to" other methods they have tested, including Google DeepMind's SynthID. Benchmark scores show that the Astra model shows no significant performance difference when watermarking is enabled. The company also plans to release the technology as open source.
The Limits of Watermarking: When Thumbprints Can Be Erased
But here is the crucial part that anyone interested in AI transparency needs to understand: watermarking is not a silver bullet. And OpenAI itself admits it.
In internal tests, OpenAI released notable figures on textGrain's effectiveness. With a target 1% false positive rate, the detection tool finds the watermark in about 80% of 200-token passages and about 95% of 400-token passages for content like psychology. Sounds impressive.
But when text is edited, everything changes. Replacing 10% of the words with synonyms drops the detection rate from about 92% to 66%. And when replacing 25% of the words, the detection rate plummets to just 17%—rendering it nearly useless.
For mathematical content—where word choices are less flexible—the detection rate is significantly lower. Short paragraphs, math answers, and translated text are also harder to detect. And of course, if the text is generated by another AI company, OpenAI's watermark will not appear.
OpenAI frankly admits: "textGrain does not guarantee reliable detection." The watermark does not verify accuracy, determine text ownership, measure human contribution, or prove human authorship. The absence of a watermark does not prove text was written by a human—it may be too short, edited, or come from another AI tool.
Because of these limitations, OpenAI decided not to provide the detection tool to the public in the initial phase. Instead, they are opening applications for approved researchers and expert organizations who can help evaluate reliability and responsible applications. The tool will report whether it detects OpenAI's watermark without revealing user identity, prompts, or conversations.
Regional Strategy: The EU Is a Laboratory, Not the World
One of OpenAI's most notable decisions is that they are not making watermarks a global default. Instead, they chose a regional approach.
In the coming weeks, watermarks will roll out for eligible ChatGPT and Codex users across all plans in the EU. For API customers worldwide, the watermark is an opt-in option for select models, starting today—defaulting to off. OpenAI is also working with cloud partners to deliver watermarks for their model outputs through partner services.
OpenAI explains that this regional approach "gives us space to learn from practical use and feedback." This is a pragmatic strategy: test where the law requires it, learn from mistakes, and then scale globally if the technology proves effective.
This does not mean OpenAI opposes transparency. The company publicly endorsed the EU's Code of Practice on Transparency of AI-generated Content in June 2026. But they also recognize that text watermarks have inherent limitations, and immediate global deployment could cause unintended consequences.
The Watermark Race: Anthropic, Google, and Pioneers
OpenAI is not the first to enter the watermarking race. Anthropic announced a global watermark for Claude in August 2026, using Google DeepMind's SynthID. This decision sparked backlash from some Claude users who argued they provided "guidance, context, decisions" while Claude was merely a "tool."
Google DeepMind has also developed SynthID, its own watermarking system, and has expanded it to multiple domains. And OpenAI—which already used SynthID for images and audio—is now developing its own text watermarking technology.
It is worth noting that OpenAI previously built a text watermark but delayed its release. The reason given in 2024 was concern that users would migrate to competitors without watermarks. The company's survey showed that nearly 30% of ChatGPT users said they would use the service less if watermarks were implemented.
But the landscape has changed. With the EU AI Act in effect, delaying is no longer an option. And with Anthropic having taken the lead, OpenAI is playing catch-up—but on its own terms.
What Does This Mean for You?
You might be thinking: "I live in Vietnam, I don't use ChatGPT in the EU. What does this have to do with me?"
The answer is: a lot.
First, this is the future of global AI. When the EU sets a standard, the rest of the world usually follows. Other countries are closely watching how the EU enforces the AI Act and may adopt similar regulations. Vietnam, as a rapidly digitizing nation, will not stay outside this trend.
Second, watermarks affect how you use AI. If you use ChatGPT to write content for blogs, social media, or work, watermarks may be embedded into that text—even if you are not in the EU. This means others can detect that your content was AI-generated, even if you edited it.
Third, this is a story of privacy and autonomy. OpenAI's watermark does not identify users. It only reports that the text was generated by an OpenAI system. But in the future, as technology evolves, could it be traced back to users? Could governments demand access to watermark data to monitor citizens? These questions remain unanswered.
Fourth, this is about the value of human creativity. If all AI text can be detected, it means the value of human-written text will rise. Those who invest in writing skills, critical thinking, and original creativity will hold an advantage in a world where AI is increasingly common.
Unanswered Questions
OpenAI's watermark is an important step forward, but it raises more questions than answers.
Can watermarks be completely disabled through clever editing? OpenAI admits that replacing 25% of words can drop the detection rate to just 17%. Thus, those wishing to hide AI usage only need to rewrite the text slightly to bypass the system.
Will watermarks cause unintended side effects? OpenAI claims no significant performance difference, but independent researchers need to verify this. Watermarks might affect output quality in subtle ways that current tests haven't detected.
Could watermarks be abused for surveillance? Although OpenAI says watermarks do not identify users, tracking is technically possible given enough data and access. Human rights organizations are monitoring this issue closely.
Will other AI companies comply? If OpenAI and Anthropic comply while smaller companies or open-source models do not, detecting AI content will become an endless game of cat and mouse.
When Transparency Is Not Just an Option
I have watched the internet evolve over two decades, and every time a new technology emerges, we face the same question: How do we want it to change the world?
OpenAI's text watermark is not a flawless solution. It has flaws, limitations, and unresolved unknowns. But it is a step in the right direction—an acknowledgment that in the AI era, transparency is no longer an option, but a duty.
The important thing is that we—users, developers, policy makers—do not sit still. We need to track how this technology evolves. We need to question its limits. And we need to ensure transparency does not become a tool for control, but a means to build trust.
Because ultimately, the question is not what AI can do. It is what we want AI to become in our lives.
What do you think about AI companies watermarking all text they generate? Is it an inevitable step toward transparency and abuse prevention, or a privacy intrusion making users feel monitored? If you are a content creator, how would you feel knowing that all AI-assisted text can be detected—and would that make you hesitate to use AI in your work? Share your views below, because in the debate over AI's future, every voice matters!
0 Comment
Add your comment to this article