Breaking01 / AI
Hybrid AI lacks evidence for 13.6% of police video events
Researchers reported speaker attribution errors and poorer performance across all tested systems on edited footage, long recordings and noisy audio.

Researchers reported on 10 Nov that 13.6% of events described by a hybrid AI system lacked support in the footage in an evaluation of publicly released police video.
The paper's abstract identifies the evaluation set as 27 publicly released UK police body-worn camera recordings, totalling 118 minutes.
The researchers published results for their hybrid processing system, NarrativeBridge-PAL-GPT-5. In these researcher-published benchmark figures, the system attributed 27.7% of utterances to the wrong speaker.
The researchers reported that every tested system performed worse on edited footage, long recordings and noisy audio. They also proposed readiness levels that distinguish tasks where AI can assist from those requiring verification of every output.
The paper describes police body-worn footage as still reviewed by people, with existing AI tools using only its audio. The evaluation covers transcription, speaker attribution, event detection and whether events described by AI have support in the recording.
The abstract does not disclose individual error rates for the multimodal models used for comparison—models that process more than one type of input, such as video and audio.