OpenAI to Watermark ChatGPT Text in the EU Under AI Act Rules

OpenAI said on 5 October 2026 that it will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act’s transparency rules, which took effect on 2 August. TechCrunch reported the rollout will reach eligible ChatGPT and Codex users on all plans in the EU over coming weeks, while API developers anywhere can enable watermarking for select models — off by default globally.

The method, called textGrain in a technical report, subtly shapes word choices into a detectable pattern without a visible symbol. OpenAI said it saw no meaningful performance change and that the watermark does not identify the user.

Limits OpenAI acknowledged

• Editing can weaken detection — replacing about 10% of words with synonyms cut detection from roughly 92% to 66% in one test
• Short passages, maths answers, and translated text are harder to detect
• A missing watermark does not prove human authorship
• Initial detector access is limited to approved researchers and expert organisations

OpenAI cautioned that watermarks can indicate an OpenAI system generated or processed part of a passage, but not how much human judgment went into it.

What this means for Malaysian readers

Malaysia is outside the EU default rollout, yet Malaysian exporters, agencies, and students dealing with EU clients may meet buyers who ask how AI drafts were produced. Local SMEs using ChatGPT for proposals should keep human editing logs and disclose AI assistance where contracts or tenders require transparency. Policymakers drafting Malaysia’s own AI rules will watch how watermarking works in practice — including how easily it can be washed out.

Malaysia’s compliance ripple

Even without a local watermark mandate yet, multinational clients may import EU-style disclosure expectations into Malaysian vendor contracts. Marketing agencies and software houses should decide internal policies on AI-assisted copy before a client audit forces the issue.

Anthropic’s earlier global watermark choice, referenced in the TechCrunch piece, shows product differentiation on transparency. Users who dislike watermarks may switch tools; enterprises that need audit trails may prefer them.

Educators should remind students that undetectable or watermark-free text still requires citation honesty — detection tools are imperfect, and academic rules are not limited to OpenAI’s detector access list.

Context for the week ahead

Malaysian readers following business and technology headlines this week will notice how often AI investment, Budget 2027 preparation, and SME cash-flow stress appear in the same news cycle. That overlap is not accidental: public spending choices, private capital expenditure, and household purchasing power are moving together. When you evaluate any single story, place it beside your own order book, job search, or household budget rather than treating it as abstract drama.

Reliable habits help: read primary statements where possible, distinguish confirmed figures from analyst estimates, and avoid resharing unverified social posts about budgets or corporate deals. For SMEs, the actionable layer is usually operational — cash timing, skills, contracts — not predicting the next index move. For workers, skills that complement automation and compliance literacy travel well across employers.

Stay sceptical of certainty. Forecasts get revised, deals can fail regulatory tests, and surveys sample only those who respond. Use this summary as a briefing, then click through to the credited source for the full report before making financial or operational decisions.

Source: TechCrunch