Text watermarking is real but works nothing like an image watermark. Systems like SynthID-Text subtly bias which tokens a model chooses, creating a statistical pattern a matching detector can score later. There is no pixel to edit — the “signal” is the distribution of word choices across the passage.
Two different things people mean
- Watermarks — a deliberate statistical bias inserted at generation time (e.g. SynthID-Text).
- Detectors — third-party classifiers guessing “AI or human” from stylistic tells, with no watermark required.
They call for different responses. A watermark is a distribution to disturb; a detector is a stylistic profile to move away from. Rewriting affects both, but neither is a guaranteed clear.
What humanization can and can't do
Rule-based humanization rewrites stock AI openers and high-frequency tell-vocabulary and varies sentence rhythm. Optional neural paraphrase (a T5-class model, when weights are present) restructures more deeply. Both reduce statistical tells — but the more you preserve the original wording, the more of any watermark you leave intact. It's a genuine trade-off, not a solved problem.
Best-effort, multi-layer, and reported honestly — originals always preserved.