Should we build an AI watermark remover?
We're measuring real demand before writing a line of code. Your vote decides whether this gets built.
Here's where people stand
0 votes so far
Thanks for voting
Your response has been recorded.
Get notified when the tool drops — leave your email address here.
What is an invisible AI watermark?
When an AI model makes an image or a paragraph of text, many providers now hide a signal inside it — imperceptible to you, but readable by a matching detector. It marks the content as machine-generated. There's no visible logo to crop out; the mark lives in the content itself.
It's woven into the content
The mark is spread across an entire image or the word choices of a whole passage — not stored in a corner or a metadata field you could strip in one click.
Only the right key can read it
A person sees nothing unusual. A detector holding the secret key can confirm, with high confidence, that a model produced the content.
How image watermarking works
Techniques like Google DeepMind's SynthID add the watermark while the image is being generated. Instead of drawing a logo on top, they nudge the values of pixels across the whole frame into a subtle, structured pattern. The strength of each nudge is kept below what the human eye can notice, so the picture looks identical to an un-watermarked one.
The pattern is designed to be robust: it's meant to survive common edits like resizing, cropping, adding filters, or saving as a compressed JPEG. Because the signal is repeated and spread out, a detector can still recover it from a modified copy — which is exactly what makes these marks hard to remove without degrading the image.
How text watermarking works
Text has no pixels to nudge, so watermarking works on word choice instead. At each step where the model picks the next token, a secret key splits the vocabulary into a favored “green” list and everything else. The model is then biased to prefer green-list words whenever a natural choice exists.
In any single sentence this is invisible — plenty of ordinary sentences happen to use green words. But across a few hundred words, genuinely AI-generated text leans toward the green list far more than chance would explain. A detector counts that lean and reports how likely it is the passage came from a model. Paraphrasing or heavy editing weakens the signal, which is the tension every text watermark has to manage.
Read the full explainer: how AI text watermarking works — with an interactive demo →
Frequently asked questions
What is an invisible AI watermark?
A signal deliberately hidden inside AI-generated content — an image or a block of text — so it can later be identified as machine-made. Unlike a visible logo, it's imperceptible to a person but detectable by a matching algorithm.
How are AI images watermarked invisibly?
Watermarks such as SynthID embed a pattern across the pixels during generation. The change is spread over the whole image and tuned to stay below the threshold of human vision, so the picture looks unchanged while a detector can still read the mark — often even after resizing, cropping, or compression.
How is AI-generated text watermarked?
By biasing the model's word choices. At each step the vocabulary is split into a favored “green” list and the rest using a secret key, and the model is nudged toward green words. Any single sentence looks normal, but across a long passage the statistical lean is measurable.
Why do AI companies add invisible watermarks?
For provenance and transparency: to label synthetic media, fight misinformation and non-consensual imagery, and comply with emerging rules such as the EU AI Act that require AI output to be marked as such.
Can invisible AI watermarks be removed?
Metadata labels strip in one click, but marks woven into the content itself — like SynthID in pixels or green-list bias in text — are built to survive editing. Removing them reliably means degrading the image or rewriting the text. No trustworthy one-click remover exists today, which is why this site polls whether one should be built at all.
Is it legal to remove an AI watermark?
It depends on where you are and why you're doing it. Rules like the EU AI Act place the marking duty on AI providers, not on users — but removing a mark to pass AI content off as human-made can breach platform policies, consumer-protection rules, or fraud laws. This isn't legal advice; the legal and ethical debate is part of what our poll is measuring.