30 October 2026
Meta researchers unveil a method to catch models trained on watermarked outputs
First reported
Latent Space ran this on .
- Meta's FAIR research group built WM-Basin, a method to detect when one AI model was trained on another model's watermarked outputs.
- The method works using only API access, so the suspect model's internal weights are not needed.
- It catches most models distilled from watermarked outputs while rarely flagging innocent ones (a low false-positive rate).
Our read
Only Latent Space of the 30 newsletters covered this story. It is the second item in our archive on watermarking and model distillation, after the September 3 report on manipulating language models into harmful outputs.
Earlier on this story
Written by Newskeryx from which newsletters covered this story and from our own archive.
How it was covered
Latent Spaceswyx & Alessio
Detection rates against baselines and robustness results