8 October 2026
ChatGPT has joined the watermarking club. I suspect workarounds will follow
Invisible watermarks promise transparency, but raise questions about detection, human contribution and how users will respond.
On 5 October, OpenAI announced invisible watermarks for eligible ChatGPT and Codex text in the EU, rolling out over the coming weeks. Google already does this with Gemini; Anthropic applies it to supported Claude models worldwide.
The AI Act doesn’t require an “AI-written” label on every post.
Providers and those using their systems have different duties. AI providers must mark generated content so software can detect it, with exceptions including standard editing. Those using AI to generate or manipulate text published to inform the public on matters of public interest must disclose that involvement, unless there is substantive human review or editorial control and a person or organisation takes editorial responsibility for publication.
These rules apply from 2 August 2026. Systems already on the market before then have until 2 December to meet the marking requirement.
OpenAI’s API watermarking is opt-in and off by default. Mmm. It says this gives customers control. But a product setting isn’t a legal exemption. If an EU-facing API customer has a marking obligation, how do they meet it with watermarking switched off?
Access to text watermark detectors is restricted for now. OpenAI and Anthropic plan to expand access. Whether LinkedIn, Meta and X will turn these signals into labels remains an open question.
Some people may translate English into another language and back again to change the wording. OpenAI says rewriting and translation can weaken detection, although an AI tool may watermark the new version too.
Others may swap words for synonyms. In OpenAI’s test of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. By the way, we might learn some new words—but similar meanings don’t make words interchangeable. The writing could end up sounding awkward.
And someone will probably make money selling tools that promise to remove AI watermarks---a nice, unexpected business opportunity.
Therefore, my concern is how people might rewrite posts to avoid a flag.
If platforms turn watermarks into “AI-written” labels, they could confuse AI assistance with AI authorship. Help with English might get a label, while heavily rewritten AI text might go unlabelled. A watermark cannot tell readers how much the person contributed.
Would that deliver the transparency EU lawmakers were aiming for?
I doubt we’ll see a uniform approach worldwide. EU rules also reach overseas providers serving its market. Still, could uneven implementation create extra friction for European users and give competitors elsewhere an advantage?
Therefore, have we really thought through how people will respond?