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.