A watermark on AI text sounds like proof. Marcus Laporte reads the fine print for Canadian businesses.
- Attributed claim: OpenAI says it will add an invisible watermark called textGrain to ChatGPT and Codex text in the EU over the coming weeks, per TechCrunch.
- Verified fact: the same report says API developers worldwide can opt in and that Canadian ChatGPT users are not covered.
- Analysis: a missing watermark proves nothing, so it cannot be used to accuse a supplier or an employee.
An OpenAI text watermark in the EU is a compliance step and not a Canadian rule, and it will not tell a Canadian employer who wrote a document. TechCrunch reported on 5 October that OpenAI will mark ChatGPT and Codex text in the EU with an invisible pattern called textGrain, tied to EU AI Act transparency requirements. Regulators and vendors still disagree in places on when those duties bite, so the date should be checked before anyone relies on it.
TL;DR
- Attributed claim: textGrain shapes word choices so the text carries a hidden signal that survives copy and paste.
- Attributed claim: OpenAI’s own testing found detection fell from about 92% to 66% when 10% of words were swapped for synonyms.
- Opinion: do not use watermark detection, present or absent, as evidence of anything in a hiring, grading or contract dispute.
What is textGrain and who gets it?
According to TechCrunch, it is applied to ChatGPT and Codex output for EU users over the coming weeks. API developers anywhere can switch it on, but it is off by default. Detection is limited to approved researchers and expert organizations, so an ordinary company cannot check a document. Canadian ChatGPT users are outside the EU rollout, so their text is not marked under this programme.
Does this change anything for a Canadian company?
Directly, no. Canada has no equivalent duty, and our earlier look at Bill C-36 explains why federal AI rules remain thin. Indirectly, yes in two cases. If you sell into the EU, you may be asked how your AI generated content is labelled. And if a supplier builds on the API, ask whether it switched marking on, since the default is off.
Fine Print
The limits are the story. OpenAI told reporters that a missing watermark does not prove human authorship, because short passages, maths answers, translations, edited text and other vendors’ models all weaken or avoid the signal. That cuts both ways. A marked document proves an OpenAI model was involved somewhere, and an unmarked one proves nothing. Anyone building a policy that treats detection as proof will punish the wrong people.
What should you do this month?
- Write a one-page rule on disclosing AI use to customers, based on your own contracts, not on a detector.
- Ask any vendor using the OpenAI API whether watermarking is on and who can verify it.
- If you sell in the EU, ask counsel which transparency duties reach your content and from what date.
- Remove detector results from any HR or grading procedure.
What does the sceptic say?
The sceptic says watermarks are a first step toward trust in AI content, and that Canada will copy the EU. Both may be right over time. But a first step with a 66% detection rate after light editing is not yet evidence, and no Canadian bill proposes it. What would prove me wrong is independent testing showing high detection after realistic edits.
What to watch
Watch for the first independent tests of textGrain, whether other vendors follow, and whether Ottawa mentions marking in a future AI bill. Also watch the EU timetable, since reports differ on when each obligation applies.
Frequently asked questions
Does ChatGPT watermark text in Canada?
Not under this programme. TechCrunch reports the textGrain rollout covers ChatGPT and Codex in the EU, and Canadian users are not included.
Can I tell if a document was written by ChatGPT?
Not reliably. OpenAI says a missing watermark does not prove human authorship, and detection is limited to approved researchers.
Does the EU AI Act apply to Canadian businesses?
It can, if you place AI systems or their output on the EU market. Ask counsel which duties reach your case and from what date.
The decision in one line
Write your own AI disclosure rule now, and never treat a watermark or its absence as proof of who wrote something.
Written by Marcus Laporte, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Archive entry dated 6 October 2026, written and fact-checked on 7 October 2026. Sources are linked on the claims they support.
