August 13, 2026
Marked for deletion
Text AI watermarks will always be trivial to remove
EU wants AI labels, but commenters say wiping them off will be child’s play
TLDR: The EU is about to require hidden labels on AI-written text, but the article argues those labels can be easily removed by simple rewriting. In the comments, people split between calling the rule useful as a deterrent and mocking it as pointless busywork that won’t fool anyone for long.
Europe is about to demand that artificial intelligence text come with a hidden label saying, basically, yes, a bot wrote this. The catch? The article’s big claim is that text watermarks are much easier to strip away than image watermarks, because writing has far less room to hide tiny changes. In plain English: if a chatbot’s answer gets lightly rewritten, the “secret signature” may vanish fast. And the comment section immediately turned into a full-on regulation cage match.
One camp was deeply unimpressed, with people asking why companies can’t just keep a searchable record of what their bots already said and offer a matching tool instead. Another group basically shrugged and said, so what if it’s breakable? One commenter compared it to a cheap lock: easy to smash, sure, but smashing it proves bad intent. That sparked the classic internet split between “better than nothing” and “this is just cookie pop-ups for AI”. Yes, the dreaded cookie-law comparison showed up, which is basically the online version of throwing a chair.
The funniest energy came from the practical cynics. Several readers pointed out the obvious loophole: just feed the marked text into another smaller bot and ask it to rewrite the whole thing. Poof, label gone. So the real drama isn’t whether lawmakers want AI text identified — it’s whether the rule becomes a meaningful speed bump or just another compliance ritual everyone learns to dodge by Friday. Community verdict: nice idea, messy reality.
Key Points
- •The article says the EU AI Act will be enforceable in August 2026 and that Article 50 requires AI outputs to be detectable as artificially generated.
- •The article argues that text watermarking is difficult because text lacks the imperceptible noise available in images, so even small alterations are usually noticeable.
- •It characterizes text watermarking as a text steganography problem in which hidden signals can degrade output quality.
- •The article says using model-likelihood checks instead of watermarks would lead to too many false positives and would be too expensive for broad public verification services.
- •The article identifies Google’s SynthID as the only publicly disclosed text watermarking system it is aware of and begins explaining its token-based approach.