OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act. The Act's transparency rules took effect on 2 August and require AI companies to mark generated content in a way other systems can identify. The watermark will roll out over the coming weeks to eligible users on all plans, but only in the EU. OpenAI says it is not making it a global default.

Like Claude, the watermark is not a symbol or a label. It works by subtly steering the model's word choices with a secret key, leaving a statistical pattern a reader cannot see, but a detector can find. Because it lives in the words, it survives copying and pasting.

OpenAI says it does not identify the user and caused no meaningful change in the model's performance. The method, called textGrain, was published in a technical report co-written with researchers from the University of Pennsylvania and Yale. It has clear limits. In OpenAI's own tests, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and short passages, maths answers and translated text are harder to detect.

For that reason, the detector will initially be available only to approved researchers and expert organisations, not the public. OpenAI is also careful to say a missing watermark does not prove a person wrote something, since the text may be too short, heavily edited or from another company's AI. Developers using OpenAI's API anywhere in the world can switch the watermark on for selected models from today, though it is off by default.

Anthropic announced two months ago that it would watermark text from its Claude models worldwide, which drew a backlash from some users.

TechCrunch's Aditya Mehta reported the rollout and noted the Wall Street Journal's 2024 reporting that OpenAI had built a text watermark earlier but held it back, partly over concerns that users would move to rivals that did not watermark. OpenAI's own announcement says a watermark can show an OpenAI system generated or processed part of a passage, but not how much human judgement, editing or creativity went into it. The European Commission's guidelines set out the AI Act transparency obligations the watermark is meant to meet.

For South Africa, the direct effect is small for now. ChatGPT text here will not carry the watermark by default, because the rule is an EU requirement, though Claude's watermark already applies worldwide.

The more useful point is what the technology can and cannot do. It can show that an AI was involved. It cannot show how much, and it cannot prove a person wrote something. So no school, employer or client should treat a watermark result as proof of cheating or of honesty, which matters for anyone tempted to screen essays or job applications with it, and fits the questions raised in the research on how students use AI.

The commercial lesson is in the 2024 report: OpenAI held its watermark back because it feared losing users to rivals, and released it only when the law required it. For a business producing work with AI, the practical step is simply to be open with clients about it, which is already what sensible workplace AI policies recommend. Being found out later is the bigger risk.

You might also like our piece on OpenAI holding back GPT-6.1 Astra, the EU's proposed under-13 social media ban, and the FSCA holding back on AI rules.

Get more SA tech and business news and subscribe to The Open Letter.