Károly Zsolnai-Fehér explains statistical watermarking for AI-generated text using a simplified distinction between preferred and non-preferred tokens. A slight bias in token selection can leave a pattern that a detector evaluates across a passage without inserting visible marks or hidden characters.
Károly Zsolnai-Fehér discusses why copy-pasting or small edits may preserve a watermark, while extensive rewriting can remove the original statistical pattern. He distinguishes his simplified explanation from more elaborate research variants and discusses access to detection tools.
Károly Zsolnai-Fehér presents the discussion in response to what he describes as a Claude watermarking rollout and argues for locally run open-weight alternatives. That rollout and its implementation are the video's claims, not independently verified findings in this catalogue account. The underlying explanation is about probabilistic text provenance, not proof of an individual author's identity.
Watch on YouTube
