Question for Michal Kosinski on Emergent Properties in Large Language Models

[muhua-h]

Emergent Properties in Large Language Models. Large Language Models (LLMs) trained to predict the next word in a sentence have surprised their creators by displaying emergent properties, ranging from a proclivity for biased behavior to an ability to write computer code and solve mathematical tasks. This talk discusses the results of several studies evaluating LLMs' performance on tasks typically used to study human mental processes. Findings indicate that as LLMs increase in size and linguistic dexterity, they can navigate false-belief scenarios, sidestep semantic illusions, and tackle cognitive reflection tasks. This talk will explore several possible interpretations of these findings, including the intriguing possibility that LLMs do not merely model language but also the psychological processes underlying how humans use language.

Reading List:

• Kosinski, M. (2024). Evaluating Large Language Models in Theory of Mind Tasks. Proceedings of the National Academy of Sciences (PNAS).
• Hagendorff, T., Fabi, S., Kosinski, M. (2023). Human-like intuitive behavior and reasoning biases emerged in large language models but disappeared in ChatGPT. Nature Computational Science.

[saniazeb8]

As language models take on roles in education, hiring, and decision making tasks we once thought required human understanding, how can we tell whether they truly understand or are just mimicking it? And what does that mean for how we value human judgment in high-stakes social and economic contexts?

[dario-marino]

It would be interesting to see where these models start to fail. I imagine that by increasing the chain of logic steps to understand the probability of getting the answer wrong increases. Can we quantify this relation?

[PENGRUISU]

How do the emergent properties of Large Language Models (LLMs), such as navigating false-belief scenarios and tackling cognitive reflection tasks, challenge or support the notion that LLMs can model not only language but also the psychological processes underlying human language use?

[Mengyang-Cao]

How can we know whether LLMs truly understand how humans use language, rather than simply mimicking it?

[Jessieliao2001]

Thanks for sharing! As large language models develop unexpected cognitive skills—from coding to reasoning—how might their deployment at scale reshape productivity growth, labor demand, and the diffusion of innovation across economies?

[yifan-lyf]

Given that LLMs can perform well on tasks traditionally used to study human cognition, how can we distinguish between genuine cognitive understanding and pattern-based statistical prediction in these models?

[JairusJia]

ChatGPT shows higher accuracy than previous models on questions containing "gotcha" prompts, and significantly reduces intuitive errors even without using chain-of-thought reasoning. Is it possible that these performances reflect the model's memory of specific task patterns or repeated exposure in the training set, rather than true improvement in reasoning ability?

[0theinfinite]

I am curious about when the model start to fail. Is it something similar to overfit?

[ralphluo11]

You highlights that LLMs demonstrate emergent behaviors such as navigating false-belief scenarios and solving cognitive tasks as they scale. Given that these capabilities resemble core aspects of human Theory of Mind, to what extent do you think these behaviors reflect genuine internal modeling versus pattern recognition based on statistical cues? In other words, how should we interpret these “psychological-like” behaviors epistemologically—are they mechanistic approximations or signs of deeper representational structures?

[yangyuwang]

Thanks for sharing! The research tested upon LLMs dating to gpt4. One question is about reasoning models, including gpt-o3/o4 and deepseek. Do you have any evidence that these reasoning models perform better in the tasks, and are they still follow the same logic as the prior models?

[fabrice401]

Could the observed ToM-like behavior in LLMs be functionally sufficient for social interaction, even if it lacks the underlying cognitive architecture of human mentalizing?

[jstanleyi]

Thanks for sharing! I am thinking, to what extent do the abilities of advanced LLMs, such as ChatGPT-4, to solve false-belief tasks indicate genuine theory of mind (ToM) capabilities, as opposed to sophisticated pattern recognition and language processing—and, furthermore, could the reasons behind early LLMs’ lack of ToM-like performance offer insight into why some autistic individuals are perceived as lacking ToM, particularly in terms of differences in learning processes, context understanding, or social cue interpretation?

[maxzhuyt]

I am very intrigued by your two papers. Both suggest LLMs are emerging cognitive agents and simulate human cognitive patterns. This perspective contrasts with the view in “Large AI models are cultural and social technologies” by Farrell, Gopnik, Shalizi, and Evans, which argues that LLMs should not be understood as intelligent agents, but as cultural and societal-level repertoire of knowledge beyond individual minds. Do you think these two opposing views can be reconciled, and how? Thank you!

[BaileyMeche]

Dr. Kosinski, given your reputation as the whistleblower on Cambridge Analytica and a leading critic of AI-driven privacy incursions, how do you reconcile championing emergent theory-of-mind in LLMs with the ethical safeguards needed to prevent the very biases and privacy intrusions your past work so starkly exposed?

[BaileyMeche]

As language models take on roles in education, hiring, and decision making tasks we once thought required human understanding, how can we tell whether they truly understand or are just mimicking it? And what does that mean for how we value human judgment in high-stakes social and economic contexts?

Distinguishing true comprehension from clever mimicry requires stress-testing LLMs with behavioral benchmarks drawn from developmental and cognitive psychology. For instance, the PNAS battery of 40 bespoke false-belief tasks embeds true-belief controls and reversed scenarios so that a model must track protagonists’ mental states rather than rely on surface cues . Even ChatGPT-4, which “solves” 75 % of these tasks on par with six-year-olds, still fails in one-quarter of cases and can be tripped up by trivial prompt alterations.

Similarly, tests of system 1 versus system 2 reasoning, like the Cognitive Reflection Test and semantic illusions, show that pre-ChatGPT models fall for human-like biases (conjunction and framing errors), whereas ChatGPT-3.5/4 largely avoids them, not through deeper insight, but by leveraging chain-of-thought prompting and reinforcement-learning-from-human-feedback fine-tuning. True understanding would demand robustness to novel, adversarial scenarios, transparent reasoning traces, and generalization beyond any single test suite.

In high-stakes domains, hiring, admissions, legal judgments, relying on pattern-matching risks amplifying hidden biases and eroding accountability. Human experts bring moral reflection, contextual nuance, and normative deliberation that go beyond statistical associations. LLMs can serve as powerful assistants, but final decisions should remain in human hands to ensure values, fairness, and responsibility are upheld.