Note: Symposia have not yet been scheduled.
Mind and machine: neural and behavioral dynamics of AI involvement
Chairs: Youngbin Kwak1, Roya Mohammadsadegh1; 1University of Massachusetts Amherst
Presenters: Roya Mohammadsadegh, Tobias Feldmann-Wüstefeld, Wen Wen, Moshe Glickman
As artificial intelligence (AI) increasingly integrates into daily life, a central question for cognitive neuroscience is how interacting with artificial systems reshapes human cognition and our sense of agency—the subjective experience of being the author of our own actions. This symposium examines the mechanisms through which AI influences the core of human cognitive experience. Using behavioral measures, computational approaches, and neural signals, this symposium examines how individuals maintain a sense of control when delegating tasks to non-human agents; how the brain signals a shift from trust to cognitive offloading before it reaches conscious awareness; how AI amplifies or reshapes existing cognitive biases; and how individual differences illuminate the fundamental nature of agency. By uniting agency, decision-making, trust calibration, and bias within a unified framework, this session investigates the redistribution of cognition between humans and machines. Moving beyond simple performance metrics, its originality lies in uncovering how reliance on artificial partners alters human cognition and subjective experience across behavioral, computational and neural levels. This investigation is uniquely timely as AI becomes embedded in everyday decision making, education, and work. This symposium will highly interest students, faculty, and CNS members across cognitive neuroscience, decision science, memory, metacognition, and human-technology interaction. Attendees will gain cutting-edge theoretical frameworks, novel experimental paradigms, and candidate neural markers for studying the co-evolution of human cognition and artificial intelligence.
Presentations
Who's in charge? Behavioral and neural signatures of agency when AI takes over stages of goal-directed action
Roya Mohammadsadegh1, Youngbin Kwak1; 1University of Massachusetts Amherst
Artificial intelligence (AI) is increasingly becoming a common partner in human decision making, guiding, or executing decisions. Yet it remains unclear how different forms of AI involvement shape the human experience of control, particularly the sense of agency, and how these effects are reflected at both behavioral and neural levels. We designed a two-stage human–AI decision task that dissociates information sampling from final placement of target. AI could enter at either stage, sampling (AI Sampler), placing (AI Placer), doing both (Full AI), or be completely absent (Baseline). Across three behavioral experiments (N = 220, 120, and 100; between-subjects), sense of agency dropped sharply when AI controlled information sampling but was enhanced or preserved when participants sampled information themselves and AI completed placement. An EEG study using the same task (N = 19; 64-channel; within-subject) tested how these behavioral effects were reflected in neural signals. Compared to baseline, mediofrontal theta power (4–8 Hz) was reduced during sampling in AI Sampler and Full AI but remained unchanged in AI placer (P < .001). During placing, centroparietal alpha power (8–14 Hz) increased selectively in the AI Placer condition (P < .001) and predicted trial-level agency, moderated by condition (P < .001). Together, these findings show that AI involvement does not simply reduce human engagement. Losing control over sampling of information carries a greater agency cost than losing control over execution based on sampled information, with dissociable oscillatory signatures that may serve as candidate markers of agency in AI-assisted decisions.
Neural signatures of trust calibration in human-AI collaboration
Tobias Feldmann-Wüstefeld1; 1Technische Universität Berlin
Effective human-AI collaboration requires people to continuously decide how much to offload to an artificial partner, a decision governed by trust. EEG has the potential to overcome drawbacks of traditional trust measures, assessing trust objectively, dynamically, and without interrupting the task. Here, we show that event-related potentials of the EEG signal used in visual attention and working memory research can be used to track trust-driven cognitive offloading during human-AI interaction. In a collaborative visual search task, participants either searched for a target among distractors themselves or monitored an AI performing the search. The N2pc, a marker of attentional selection, was reduced when the AI was highly reliable compared to an error-prone AI, reflecting reduced attentional engagement and thus greater offloading. In a collaborative change-detection task, participants had to store arrays of colored items in working memory for a short period. The contralateral delay activity (CDA), a marker of working memory storage, tracked how many memory items participants offloaded to an AI partner. Offloading increased with AI reliability, but only when spare working memory capacity was available. Trust was further enhanced when AI reliability information was transparently communicated. The CDA also tracked the stages of trust development: it increased during trust formation, fell during trust violation, and returned to pre-violation levels once reliability was restored. Together, these findings establish attentional and working-memory ERP components as objective, continuous, non-disruptive indices of trust calibration, offering mechanistic insight into how humans dynamically allocate cognitive resources and retain agency when collaborating with artificial systems.
People learn AI biases while judging AI with their own
Moshe Glickman1, Tali Sharot2,3; 1Bar-Ilan University, 2University College London (UCL), 3Massachusetts Institute of Technology (MIT)
People have always made decisions with others; increasingly, the other is artificial intelligence (AI). However, AI partners differ from humans: their responses are more consistent and they are often perceived as more accurate. Here, we examined whether people learn biases from AI, and whether they evaluate AI with the biases they apply to humans. In the first series of experiments (N=1,401), participants classified arrays of emotional faces as "more happy" or "more sad" and, after each response, saw a partner's response. While the original human judgments showed only a slight bias toward "more sad", an AI trained on these judgments amplified the bias. Participants interacting with the biased AI became more biased themselves (p<.001), whereas no such increase was observed after interacting with other humans (p=.48). A consistent effect was observed in a motion-estimation task: interaction with a biased AI increased participants' bias (p<.001), whereas interaction with an accurate AI improved accuracy (p<.001). Participants reported being influenced mostly by the accurate AI, although the actual influence was comparable. In the second series of experiments (N=379), we examined whether a self-evaluative bias (better-than-average effect) extends to AI partners. Although participants were less accurate than the AI (p<.001) and as accurate as their human partners (p=.90), they judged themselves as more accurate than both (all p<.009). Together, these findings suggest that interaction with AI engages existing cognitive biases rather than bypassing them. Some biases are shaped by properties of the AI itself, whereas others generalize from human social cognition to AI partners.
Sensitivity, criterion, and individual differences in the sense of agency
Wen Wen1; 1Rikkyo University
The sense of agency refers to the feeling of controlling one’s actions and their consequences. It varies substantially across individuals, yet the mechanisms underlying these differences remain unclear. This talk will introduce a framework that separates two components of agency judgments: sensitivity, reflecting the ability to discriminate changes in action–outcome relationships, and criterion, reflecting an individual’s general tendency to attribute outcomes to oneself. In the first behavioral study, conducted online, we measured participants’ sensitivity and criterion in attributing external events to themselves. Specifically, participants freely moved a dot on the screen using a mouse and made binary self-attribution judgments. For each participant, the slope and point of subjective equality (PSE) of the judgment curve were calculated and used to quantify sensitivity and criterion, respectively. In addition, we measured sensitivity in self–other discrimination by asking participants to identify, among three dots, the one that they were able to control. The results showed that these three components—sensitivity in self-attribution, criterion in self-attribution, and sensitivity in self–other discrimination—can vary independently and are differentially related to contextual beliefs about control, which were measured using questionnaires. Finally, a Bayesian integration model was used to characterize individual differences in sensitivity to different cues and in the criterion for self-attribution. This framework may provide a more precise way to characterize both typical variation and altered experiences of agency, and may help bridge research on agency across perception, action, and human–machine interaction.
CNS Account Login
