Moshe Glickman’s research while affiliated with University College London and other places

Artificial intelligence (AI) technologies are rapidly advancing, enhancing human capabilities across various fields spanning from finance to medicine. Despite their numerous advantages, AI systems can exhibit biased judgements in domains ranging from perception to emotion. Here, in a series of experiments (n = 1,401 participants), we reveal a feedback loop where human–AI interactions alter processes underlying human perceptual, emotional and social judgements, subsequently amplifying biases in humans. This amplification is significantly greater than that observed in interactions between humans, due to both the tendency of AI systems to amplify biases and the way humans perceive AI systems. Participants are often unaware of the extent of the AI’s influence, rendering them more susceptible to it. These findings uncover a mechanism wherein AI systems amplify biases, which are further internalized by humans, triggering a snowball effect where small errors in judgement escalate into much larger ones.

Knowledge is distributed over many individuals. Thus, humans are tasked with informing one another for the betterment of all. But as information can alter people’s action, affect and cognition in both positive and negative ways, deciding whether to share information can be a particularly difficult problem. Here, we examine how people integrate potentially conflicting consequences of knowledge, to decide whether to inform others. We show that participants (Exp1: N = 114, Pre-registered replication: N = 102) use their own information-seeking preferences to solve complex information-sharing decisions. In particular, when deciding whether to inform others, participants consider the usefulness of information in directing action, its valence and the receiver’s uncertainty level, and integrate these assessments into a calculation of the value of information that explains information sharing decisions. A cluster analysis revealed that participants were clustered into groups based on the different weights they assign to these three factors. Within individuals, the relative influence of each of these factors was stable across information-seeking and information-sharing decisions. These results suggest that people put themselves in a receiver position to determine whether to inform others and can help predict when people will share information.

Models of multi-attribute decision making vary on whether all or only part of the information available is being processed, as well as on whether preference formation is based on within-option or within-attribute processing. Here we carry out a combined empirical and computational study in which we rely on lottery-options with varying task complexities. We monitor eye-gaze during the decision formation to determine which decision-relevant information participants attend, and when. We then compare a large set of models, of different levels of complexity in their ability to account for the choices made by individual participants, and we find that two models outperform all others. The first is the two-layer 1 leaky-competing accumulator based on Prospect Theory (LCA-PT), which predicts human choices on simple tasks better than any other model. For complex tasks a new model based on older work in operations research performs best in the complex task, with both its performance as well as that of the second-ranked LCA-PT model significantly exceeding that of all other models. Both of these models use the sequence of observed eye movements for each participant to capture the allocation of attention to specific options and attributes during the decision process, but make different assumptions about the effect of attention on decision making. Our results suggest that, when faced with complex choice problems, people form preferences primarily based on attention-guided pairwise, within-attribute value-comparisons.

Noise is intuitively thought to interfere with perceptual learning; However, human and machine learning studies suggest that, in certain contexts, variability may reduce overfitting and improve generalizability. Whereas previous studies have examined the effects of variability in learned stimuli or tasks, it is hitherto unknown what are the effects of variability in the temporal environment. Here, we examined this question in two groups of adult participants (N=40) presented with visual targets at either random or fixed temporal routines and then tested on the same type of targets at a new nearly-fixed temporal routine. Findings reveal that participants of the random group performed better and adapted quicker following a change in the timing routine, relative to participants of the fixed group. Corroborated with eye tracking and computational modeling, these findings suggest that prior exposure to temporal randomness promotes the formation of new temporal expectations and enhances generalizability in a dynamic environment. We conclude that noise plays an important role in promotion perceptual learning in the temporal domain: rather than interfering with the formation of temporal expectations, noise enhances them. This counterintuitive effect is hypothesized to be achieved through eliminating overfitting and promoting generalizability.

Individual differences in cognitive processing have been the subject of intensive research. One important type of such individual differences is the tendency for global versus local processing, which was shown to correlate with a wide range of processing differences in fields such as decision making, social judgments and creativity. Yet, whether these global/local processing tendencies are correlated within a subject across different domains is still an open question. To address this question, we develop and test a novel method to quantify global/local processing tendencies, in which we directly set in opposition the local and global information instead of instructing subjects to specifically attend to one processing level. We apply our novel method to two different domains: (1) a numerical cognition task, and (2) a preference task. Using computational modeling, we accounted for classical effects in choice and numerical-cognition. Global/local tendencies in both tasks were quantified using a salience parameter. Critically, the salience parameters extracted from the numerical cognition and preference tasks were highly correlated, providing support for robust perceptual organization tendencies within an individual.

Human decision-making biases are pervasive, influencing areas such as finance and medicine. Reliance on Artificial Intelligence (AI) systems has been offered as a solution to reduce such systematic errors. However, in judgments ranging from perception to emotion, AI systems can also exhibit biases. Here, across multiple experiments (N = 1,022) we reveal a feedback loop where human-AI interactions alter processes underlying human perceptual, emotion and social judgements that subsequently amplify biases in humans. This amplification is significantly greater than observed in human-human feedback loops. Participants are unaware of the extent of the AI’s influence, which can leave them more susceptible to it. The findings uncover a mechanism by which AI-human interaction creates a snowball effect: small human errors in judgement escalate into much larger ones.