At the core of what defines us as humans is the concept of theory of mind: the ability to track other people’s mental states. The recent development of large language models (LLMs) such as ChatGPT has led to intense debate about the possibility that these models exhibit behaviour that is indistinguishable from human behaviour in theory of mind tasks. Here we compare human and LLM performance on a comprehensive battery of measurements that aim to measure different theory of mind abilities, from understanding false beliefs to interpreting indirect requests and recognizing irony and faux pas. We tested two families of LLMs (GPT and LLaMA2) repeatedly against these measures and compared their performance with those from a sample of 1,907 human participants. Across the battery of theory of mind tests, we found that GPT-4 models performed at, or even sometimes above, human levels at identifying indirect requests, false beliefs and misdirection, but struggled with detecting faux pas. Faux pas, however, was the only test where LLaMA2 outperformed humans. Follow-up manipulations of the belief likelihood revealed that the superiority of LLaMA2 was illusory, possibly reflecting a bias towards attributing ignorance. By contrast, the poor performance of GPT originated from a hyperconservative approach towards committing to conclusions rather than from a genuine failure of inference. These findings not only demonstrate that LLMs exhibit behaviour that is consistent with the outputs of mentalistic inference in humans but also highlight the importance of systematic testing to ensure a non-superficial comparison between human and artificial intelligences.

Newborns come into the world wired to socially interact. Is a propensity to socially oriented action already present before birth? Twin pregnancies provide a unique opportunity to investigate the social pre-wiring hypothesis. Although various types of inter-twins contact have been demonstrated starting from the 11^th week of gestation, no study has so far investigated the critical question whether intra-pair contact is the result of motor planning rather then the accidental outcome of spatial proximity. Kinematic profiles of movements in five pairs of twin foetuses were studied by using four-dimensional ultrasonography during two separate recording sessions carried out at the 14^th and 18^th week of gestation. We demonstrate that by the 14th week of gestation twin foetuses do not only display movements directed towards the uterine wall and self-directed movements, but also movements specifically aimed at the co-twin, the proportion of which increases between the 14^th and 18^th gestational week. Kinematic analysis revealed that movement duration was longer and deceleration time was prolonged for other-directed movements compared to movements directed towards the uterine wall. Similar kinematic profiles were observed for movements directed towards the co-twin and self-directed movements aimed at the eye-region, i.e. the most delicate region of the body. We conclude that performance of movements towards the co-twin is not accidental: already starting from the 14th week of gestation twin foetuses execute movements specifically aimed at the co-twin.

Humans spend most of their time interacting with other people. It is the motor organization subtending these social interactions that forms the main theme of this article. We review recent experimental studies testing whether it is possible to differentiate the kinematics of an action performed by an agent acting in isolation from the kinematics of the very same action performed within a social context. The results indicate that social context shapes action planning and that in the context of a social interaction, flexible online adjustments take place between partners. These observations provide novel insights on the social dimension of motor planning and control.

How do we understand the intentions of other people? There has been a longstanding controversy over whether it is possible to understand others’ intentions by simply observing their movements. Here, we show that indeed movement kinematics can form the basis for intention detection. By combining kinematics and psychophysical methods with classification and regression tree (CART) modeling, we found that observers utilized a subset of discriminant kinematic features over the total kinematic pattern in order to detect intention from observation of simple motor acts. Intention discriminability covaried with movement kinematics on a trial-by-trial basis, and was directly related to the expression of discriminative features in the observed movements. These findings demonstrate a definable and measurable relationship between the specific features of observed movements and the ability to discriminate intention, providing quantitative evidence of the significance of movement kinematics for anticipating others’ intentional actions.