People detected political speech deepfakes substantially better with audio or audiovisual information than from transcripts, but text-to-speech fakes remained far from perfectly recognized.

What this does not establish

Political speech is not brand advertising; human discernment is not automated detector performance or a measured provenance-label effect.

Counterevidence & uncertainty

Modality helped, but detection remained stimulus- and base-rate-sensitive with material residual error.

What would change the reading

Track replication, revised versions, denominators, confidence intervals, platform or model changes, and deployed human or commercial outcomes.

Primary routes

1. SSCO-S008https://doi.org/10.1038/s41467-024-51998-zOpen source record →