Discriminating real and posed smiles: Human and avatar smiles

This study addresses the question whether untrained and unaided observers can discriminate between real and posed smiles from different sets of smiling videos. Observers were shown smiling face videos, either in single or paired, to rate each of them on a number of scales intended to assess perceived real and posed smiles. We implemented four experiments where single smiles were shown in most cases, and paired smiles were shown in the third experiment. We found that observers are more accurate in response to paired videos (72.9%) compared to single videos (61.7% and 60.2%). Performance is increased when voting is introduced, but we need 11 to 13 observers in single smiles and 13 to 15 in paired smile discrimination. We found that female observers are more accurate compared to male observers. On testing with 8 'smiling' virtual avatars we found only one was rated a real smile. This work will have significant impact on judgement of genuineness of smiles from avatars / automated virtual assistants, in that untrained individuals using normal human abilities can estimate how real the virtual smiles will seem for the human users to whom they are directed.