A large meta-analysis by Bond and DePaulo (2006) pooling over 200 studies found that people's ability to detect lies from demeanor alone averages around 54% accuracy — barely above the 50% you'd get by guessing randomly, and this holds true even for groups assumed to be experts, like police officers and judges, in most studies. Popular myths — liars avoid eye contact, liars fidget, liars look up and to the left — have been directly tested and don't hold up as reliable indicators; some studies even found the opposite pattern, since people who know these myths sometimes overcompensate by maintaining exaggerated eye contact while lying.
What research shows works modestly better than demeanor-watching is analyzing the content and consistency of a story rather than the storyteller's body language: asking someone to recount events in reverse order (much harder to do consistently if fabricated), asking unexpected follow-up questions that a rehearsed story wouldn't anticipate, and looking for inconsistencies across repeated tellings over time rather than a single read of nervousness in the moment. None of these are close to reliable lie detectors either — they simply outperform demeanor-watching, which is close to useless.
Bond and DePaulo pooled results from every controlled lie-detection study they could find — 206 in total — where participants watched or listened to someone and judged whether they were lying. Average accuracy across the entire pooled dataset came out to about 54%, barely above the 50% a coin flip would produce. Critically, this held even for groups widely assumed to have real expertise: police officers, judges, customs officials, and psychiatrists performed no better on average than college undergraduates with no relevant training at all. The specific cues people were most confident in — avoided eye contact chief among them — showed essentially no reliable link to actual lying across the pooled studies, and some individual studies even found the opposite pattern.