AI Detection & Watermarking
Can you prove a text or image came from an AI? After the fact, usually not: AI writing is statistically ordinary, so detectors that scan finished text guess from style and often accuse humans falsely. Watermarking works at generation time instead. At every step the model secretly splits its vocabulary into a green list and a grey list (keyed by the previous tokens) and slightly favours green. The text reads normally, but count the greens and the skew is overwhelming evidence — for anyone holding the key. Limits: paraphrasing dilutes it, and it only marks models that opt in.
Each square is one token. Toggle the watermark: with it on, green tokens dominate and the detector fires; without it, green sits near 50% and there is nothing to find. Hit new text to resample.
The generator secretly favours green-list tokens. The text reads normally, but the statistical skew is unmistakable to anyone with the key.
Check yourself
Why does a watermark detector almost never accuse a human by mistake, while a style-based detector often does?
Go deeper (free): Google DeepMind — SynthID watermarking ↗