Meta Paper Shows Watermarks Can Expose Distilled AI Models
A new detection method developed by researchers at Meta identifies AI models distilled from watermarked outputs far more effectively than previous techniques.
Meta Paper Shows Watermarks Can Expose Distilled AI Models
A new detection method developed by researchers at Meta identifies AI models distilled from watermarked outputs far more effectively than previous techniques.
Detecting Distilled Models With Meta AI
Copying a large language model traditionally required stealing its underlying weights. Today, bad actors can achieve a similar result by querying a model extensively and training a smaller student model on the collected responses. To combat this practice, researchers at Meta have introduced a novel detection approach.
A study published on October 27, 2026, by Meta's FAIR and Superintelligence Labs details a black-box technique known as WM-Basin. According to the research paper titled "Watermark Basins: Black-Box Detection of Distilled LLMs," the method identifies models that have been distilled from a watermarked teacher model.
Performance and Test Results
In rigorous testing, the WM-Basin technique significantly outperformed prior state-of-the-art token-level detectors. The study highlights a 79% true-positive rate at a false-positive rate of just 0.1% using n-gram=1.
By comparison, previous black-box token-level detectors managed a true-positive rate of only 29% under identical testing conditions. This substantial improvement provides developers and creators with a more reliable way to flag unauthorized student models trained on watermarked outputs.
About The Author
