An online polygraph with AI has learned to separate truth from falsehood using only text prompts.

To analyze whether truth and lies can be distinguished from simple text communication, researchers

developed an online game in which playersthe roles of “saint” or “sinner” were randomly assigned. 40 subjects took part in 80 gaming sessions, during which a machine learning system looked for inconsistencies between their answers.

As a result, the neural network found that liarsrespond faster, using words like “always” and “never” to confirm a sense of confidence. Those who spoke the truth used words like “guess” and “maybe” and spent more time answering.

As a result, the neural network learned to recognize lies and truth with an accuracy of 85 to 100%, while a person is able to separate the truth only 50% of cases.

Earlier, NVIDIA at the GTC 2019 exhibition presented a neural network capable of turning any sketchy drawing into a photorealistic image in a few seconds.