[gxigzigxigxi]

I wonder if a machine-learning approach really would learn how to detect lies. There is already a huge corpus to train on: police interrogation videos. It would take significant effort to label deception that was later confirmed. But I’ll bet we could teach a computer to see through almost any lie if we decided it was an important thing to spend time on.

[stedalus]

Julia Hirschberg gave a keynote at this year's EMNLP on this topic. Short answer was that ML can do better than trained humans in some cases, but very far from "almost any".

I don't think EMNLP has posted talk videos yet but these slides look very similar: http://www.cs.columbia.edu/~julia/talks/ICASSP2018-latest.pd...

[everdev]

Maybe if you're looking at non-syntax indicators like pace of speech, posture adjustments, facial expressions, etc. But NLP is still a major unsolved problem. AI simply isn't strong enough at understanding meaning especially in broken English or slang (ex. "Nah, I ain't done it.").

[gxigzigxigxi]

I’m not suggesting that the model would infer the truth from the input. Just that it would render a reasonably reliable estimate that the subject is dissembling. I don’t think the actual words used would be the principle component of such a determination.