How Do Humans Perceive AI-Generated Action?

The recent developments in the field of generative Artificial Intelligence (AI) allow for generationof realistic visual content, virtually indistinguishable from genuine human-created videos andimages. While several studies focused on people’s trust or biases toward AI and differencesbetween human and AI-generated visuals, the methodology is usually qualitative or aims touncover explicit biases. Here, we introduced a method adapted from action perception researchthat measures implicit biases humans have about intentional agents’ behaviour. Participantswatched a series of videos showing a hand reaching for an object, sometimes around anobstacle, with either a straight or a high, arched trajectory. The actions were either efficient(avoiding an obstacle) or inefficient (heading straight for the obstacle) and concluded in thehand’s disappearance. It has been previously shown that participants misperceive the last locationof the hand towards the efficient trajectory, but only when observing intentional agents. Wetested whether this bias would change when informing participants that the videos show eithergenuine human action or are AI-generated, across two experiments. The results indicated thatgeneral perceptual bias was preserved, however, there is no difference based on the video origin,suggesting that AI-generated action is not viewed as less or more intentional than human one.On the other hand, some indications of perceivable differences between the two video typeswere found in the representational momentum effects. Limitations and possible improvementsof the presented research are outlined, such as additional modifications to stimuli design, to ensurerobustness and correct interpretation of the obtained data in the future.