Note: Symposia have not yet been scheduled.
Artificial minds, human brains: Is AI a paradigm shift in cognitive neuroscience?
Chair: Shahar Arzy1; 1Hebrew University, 2Hadassah Medical Center
Presenters: Anat Perry, Jeremy Manning, Neele Engelmann, Shahar Arzy
AI is no longer on the horizon of cognitive neuroscience. It has arrived. From conceptual thinking to daily use, its impact is undeniable. But does it fundamentally change how we think about the mind and brain, or simply enhance what we already do? This symposium tackles that question through different cognitive domains: emotions, memory, language, morality and personality. In the domain of emotions, Anat examines the "artificial-empathy paradox": the finding that AI-generated empathy is well received until its source is revealed. Drawing on naturalistic studies, eye-tracking, and large-scale experiments, she argues that empathy functions as a costly signal whose value is inherently tied to its human origin, with broad implications for relationships and society. In language, Jeremy applies a geometric framework to the dynamics of natural language as produced by LLMs. He explore how these models serve as a novel lens onto longstanding questions about the fundamental nature of spontaneous thought and mental life. Neele probes the moral dimension of human-AI interaction. Across economic games, chatbot experiments, and a wartime targeting simulation, she investigates how delegating decisions to AI erodes moral restraint — from financial dishonesty to physical harm. Finally, Shahar shows how LLMs can revolutionize personality research by generating and validating questionnaires from text, and by recovering personality traits and life outcomes directly from model activations, offering a scalable, language-grounded alternative to century-old frameworks. Together, these four talks illuminate both the transformative promise and the profound risks of AI as a new paradigm for understanding the human mind.
Presentations
The Artificial-Empathy Paradox, or – What we Value in Human Empathy
Anat Perry1,2; 1Harvard University, 2Hebrew University
AI-generated empathy is convincing and well received until recipients realize it is artificial; a phenomenon I term the artificial-empathy paradox;. I test and explain this phenomenon by examining the three subcomponents of human empathy – cognitive, affective and motivational, and the gaps between each of them and AI, from the perspective of the empathy recipient. First, I will present naturalistic empathic-accuracy studies showing that large language models outperform humans on cognitive measures of empathy (N = 127). I will then present nine studies (N = 6,282) in which identical AI-generated empathic responses were attributed either to humans or to AI. Responses attributed to humans were perceived as more empathic and supportive and elicited more positive and fewer negative emotions than those attributed to AI. Moreover, participants’ suspicion that AI had assisted perceived human responses reduced perceived empathy and support. Finally, I will present an eye-tracking study (N = 120) demonstrating that readers process empathic texts differently depending on whether they believe the source is human or AI. I will suggest that, for recipients, empathy functions as a costly signal. Its value lies not only in the immediate emotional support it provides, but also in its informational utility for predicting the future. If true, this offers a parsimonious explanation for why empathy from AI is devalued, and why even human empathy suspected to be AI-assisted loses value. I will acknowledge the risks of increasing reliance on AI-generated empathy and discuss implications for relationships and for society at large.
On the fundamental nature of thought and language
Jeremy Manning1, Mohammad Javad Latifi Jebelli1, Dan Rockmore1; 1Dartmouth College
Understanding the dynamics of spontaneous thought and natural language has been at the heart of understanding the human mind since the earliest days of philosophy and psychology. A parallel line of human inquiry and creation has concerned speculations about the creation of artificial minds: how they might work, what they might say or do, and what their existence might teach us about our own minds. With the rise of conversation-capable large language models (LLMs) over the past decade, scientists have begun to explore how these models might be used to study not only the models themselves, but also the human mind as viewed through the lens of LLMs. Here we focus on a specific geometric framework for understanding the dynamics of language as produced by LLMs. We use this framework to formulate deep, but previously difficult to articulate, questions about the nature of thought and language, and we explore how LLMs might be used to address these questions.
Outsourcing Morality: Dishonesty and Harm in Human-AI Delegation
Neele Engelmann1; 1Max Planck Institute for Human Development, Center for Humans and Machines, Berlin, Germany
People increasingly delegate actions to AI, from drafting messages to lethal targeting in war. Does moral restraint survive this delegation? In economic games, people request more dishonesty from an AI delegate than they would produce themselves, and AI complies with unethical requests far more readily than humans (Köbis et al., 2025). My research asks how far this reaches, from financial dishonesty to physical harm, and how human-AI interaction can preserve moral behavior. In completed work, participants instruct an AI to report outcomes for them, with incentives to cheat. Improving mechanical understanding of the system did not reduce dishonesty; what curbed cheating was grasping the ethical meaning of a given instruction. This held across two experiments (N = 788; N = 751), including dynamic, live interactions with chatbots. A new line of work probes the effects of delegation when the act harms another person. In an online serious game simulating wartime targeting, participants are drone operators choosing whether to strike, accepting collateral harm, either themselves or by delegating to an AI or a human (anticipated N ~ 600, collected by the conference). A third study (in preparation) moves the question into the laboratory, where harm is physical: participants decide whether to deliver a painful but harmless shock to another by their own hand, an autonomous drone, or a human delegate, while skin conductance and ECG record their affective response. Initial pilot data is expected by the conference. Together, the work asks how much moral agency we can safely hand to machines.
LLMs revolutionize a century-old approach to personality research
Shahar Arzy1,2; 1Hebrew University, 2Hadassah Medical Center
Personality is one of the main pillars of cognitive and psychological sciences. While personality theories served in the fundamentals of psychology, personality research was mostly focused on the Five-Factor Model (“Big-Five”). Since the Big-Five is based on the lexical hypothesis that personality traits are encoded in language, it is plausible that large language models (LLMs) may offer new ways to explore personality through text. In a first project, we developed an LLM-based method to generate and validate personality questionnaires from textual sources. Using the DSM-5 personality disorders section and a popular astrology book, we generated two questionnaires and administered them, alongside the Big-Five Inventory (BFI), to 600 adults. Internal consistency was high for the BFI and the DSM-based questionnaire but low for the astrology-based one. The LLM predicted human response patterns for both generated questionnaires. Individual items from both the DSM- and astrology-based questionnaires predicted diverse life outcomes at levels comparable to the BFI. In a second project we investigated what LLMs’ internal representations reveal about the people whose text they process. We used predictors on model activations to recover self-reported Big-Five traits, and extend the analysis to demographics and life outcomes. Personal texts, mostly childhood-memories, were able to reveal information of three out of the Big-Five factors. In 79/100 participants, LLMs showed better predictions on life outcomes than the Big-5. These findings show that LLMs can construct personality measures and anticipate response patterns, offering a scalable framework for LLM-based personality research.
CNS Account Login
