Hannah Fry asks Pushmeet Kohli how watermarking differs from trying to classify an output after it has been generated. Pushmeet Kohli describes deliberately introducing imperceptible signals that can survive common transformations, while balancing robustness, detection accuracy and deployment scale. Text watermarking uses statistical choices during generation, but short or tightly constrained answers leave little room for a reliable signal.
Hannah Fry asks Jeremy Ratcliff how the same provenance principle applies to AI-generated biological designs. Jeremy Ratcliff describes separate approaches for protein structures and sequences and reports promising tests on selected protein binders. The research is a proof of concept, not evidence of universal compatibility or an already deployed industry-wide system.
Hannah Fry, Pushmeet Kohli and Jeremy Ratcliff discuss adoption, shared detection tools and the limits of watermarks. Provenance can complement other screening and biosecurity measures, but does not by itself identify a hazard or certify safety. The conversation also considers metadata, adversarial changes and coordination across providers rather than presenting one technique as a complete solution.
Watch on YouTube
