Mike Pound explains to Sean Riley how a language model can intentionally embed a watermark during text generation through its choice of tokens. Early approaches divide possible next tokens into favored and unfavored groups, which can change the output distribution. Mike Pound introduces a tournament method designed to retain the model's underlying probabilities while leaving a statistical pattern that a detector with the matching secret key can test.
Mike Pound begins the worked tournament by sampling candidate tokens according to the language model's probabilities. A keyed hash of recent context, each token and the tournament round assigns a binary score; candidates compete in rounds, with random tie-breaking. His explanation uses a four-token context so that a small edit affects a local part of the pattern. When the model has essentially one valid next token, the tournament cannot create much watermark signal without changing the intended output.
Mike Pound explains that the detector revisits the resulting text and recomputes the keyed scores from each token and its context. Ordinary unwatermarked text is expected to average around one half, while the tournament tends to favor higher-scoring tokens. Aggregating that small bias over a sufficiently long passage provides statistical evidence of the particular watermark. Mike Pound explicitly qualifies this evidence and does not establish definite authorship from a single detection result.
Mike Pound demonstrates Python code that wraps Qwen 3 through Ollama, retrieves candidate tokens and their likelihoods, and applies the tournament before choosing each output token. He compares unwatermarked and watermarked essays about Macbeth, then tries a more open-ended story. In his reported results, the unwatermarked example scores close to the baseline, while the more varied story produces a stronger signal despite being shorter. These are Mike Pound's demonstration results, not an independently reproduced evaluation.
Sean Riley asks how editing, unwatermarked models and code generation affect detection. Mike Pound explains why short passages and structured outputs such as code or JSON provide fewer unconstrained token choices. His code example produces a weaker signal after surrounding explanatory text is removed. Editing, paraphrasing and changes to the context used by the watermark can reduce its strength, while models that do not apply the method may leave no such mark. The method offers conditional evidence about marked content rather than a universal test for AI writing.
Watch on YouTube
