A hidden mark in the paragraph does not make it true

On 5 October 2026, OpenAI described how it will mark some of its text. The EU AI Act, OpenAI says, requires generative AI providers to make generated text identifiable in a machine-readable way. The company's method is called textGrain. It is an invisible statistical signal in the model's word choices. It is not a hidden character, an invisible space, or odd punctuation. A reader cannot see it. A detector looks for the pattern.

The rollout is narrow, and the limits are the story.

Over the coming weeks, eligible ChatGPT and Codex users in the European Union, on all plans, get the mark. OpenAI is not making text watermarking a global default at launch. Starting 5 October, API customers anywhere can opt in for select models. In the API it stays off unless they turn it on. The detector is not a public website. Access starts with approved researchers and expert organizations. OpenAI says false alarms and missed marks are why.

Their own tests show how fragile the signal is. At a target false-positive rate of 1 percent, the detector found the mark in about 80 percent of 200-token psychology passages, and about 95 percent of 400-token ones. Mathematics, where the wording has less room to move, was substantially harder. On 400-token passages, swapping 10 percent of the words for synonyms dropped detection from about 92 percent to 66 percent. Swapping 25 percent dropped it to 17 percent. A light edit, the kind a person does before publishing, can wipe the trail.

OpenAI also lists what a detection does not mean. It does not measure how much a human edited. It does not decide who owns the text. It does not name the user. It does not say the passage is true. And a missing mark does not prove a person wrote it. The text may be short, edited, translated, from an older model, or from another company's tool.

OpenAI's quality table, for a model it calls Astra, shows small moves in both directions. GPQA Diamond is 94.44 percent without the mark and 93.94 percent with it. They say they do not see a meaningful performance difference. A watermark is not a writing course. It does not add the fact your buyer needs.

That is the part a company site should take seriously. A paragraph can leave ChatGPT, lose its mark in an edit, and still be wrong. It can keep its mark and still be wrong. Google's own guidance on generative AI says the model predicts words. It tells site owners to fact-check what they publish, including the title, the meta description, the structured data, and the image alt text, because those can appear in search. A marked sentence and an unchecked title are both ways to put a false line where a buyer will see it.

So the useful question is not "can a detector catch us." The useful question is whether a person, who knows the company, agreed that the page is accurate.

When we write a company website, the written plan names the pages, the price, and who owns the code before we start. You agree that plan. The pages have to match it. We do not invent a client list, a rank, or a load time for every country. After handover, we fix faults in the agreed work for at least 3 months. A new page you think of later is a new line in a plan, not a silent edit.

Search work has the same discipline. Titles and descriptions should match the page. A sitemap should list pages that exist. We do not promise a rank. If an assistant cites you later, that citation is only as good as the page. Cloudflare and OpenAI's discovery note, from 15 October, is about finding pages. It is not a fact-checker. We do not claim a citation.

The same check applies if the job is an app rather than a page. A mobile app's screens, a desktop app's labels, an AppSheet form built on your sheets: the words on them are part of the plan. A watermark on a draft in a chat does not review a button that charges the wrong amount. Staff attendance and Kasir Kita, if you want to see a product before a custom job, have demos. A custom build still gets its own written scope.

If your team has been pasting assistant text onto the site, do not hunt for an invisible mark. Read the page as a buyer would. Does the title match? Does the price, if you state one, match the plan? Does the author exist? Is the hour you advertise an hour you answer? WhatsApp and email here are answered Monday to Friday, 08:00 to 17:00 Western Indonesia Time.

Tell us what you need. We reply by the next business day. You see the written plan, and you agree it, before we start.

Source: OpenAI, "Our approach to EU text provenance rules," 5 October 2026. https://openai.com/index/eu-text-provenance/