LLMs produce content that seems human, yet the underlying dynamics differ. #
Posted July 20, 2026 [ Reviewed by Michelle Quirk
](/us/docs/editorial-process)
Key points
- While LLMs appear to reason, their "thought" processes fundamentally differ from ours.
- Human reasoning is a kaleidoscope of processes, integrating memory, context, culture, and more.
- While LLMs can be improved, the underlying dynamics will always be different, according to some.
A team of psychologists, computer scientists, and physicists from Italy, Slovenia, and South Korea has proposed seven “fault lines” between human and artificial intelligence (AI). The researchers question the extent to which AI is yielding intelligence, at least in the same form as human intelligence. The authors note the critical shift that occurred between statistical natural language processing (NLP), where AI retrieves and ranks existing information, with the (human) user then able to judge between them, and generative AI, which instead presents one fluent, “authoritative-seeming” answer. In short, while large language models (LLMs) produce content that seems cognitively informed and deliberated, the processes underlying how that content is produced are fundamentally different from human cognition.
One of the fundamental divergences between human and artificial intelligence highlighted by the authors relates to how they use language. AI does not understand language the way a human mind processes it and derives meaning; AI instead recognizes statistical patterns, garnered from human-produced text: “[LLMs] do not track truth conditions or causal structure; they track patterns of co-occurrence, association, and continuation in text.” The authors argue that by supplementing this ability with a number of additional mechanisms such as retrieval-augmented generation, they make LLM outputs look more reliable and more convincing without becoming more knowledgeable.
Examining the differences between human and artificial intelligence, the authors propose seven epistemic fault lines that highlight these differences. These correspond to seven sequential stages of human judgement:
The grounding fault: The authors argue that humans arrive at judgments based on layers of information that are often not accessible to LLMs, such as facial expressions, tone of voice, and social cues. Accordingly, LLMs often underdetect expressions that are sarcastic or ironic, or other human-specific expressions. Despite that, LLMs appear to be getting better at detecting these, though that is related to more effective pretraining and conditioning rather than understanding.The parsing fault: Once humans receive sensory and social information, they begin processing it through simultaneous consideration of meaning and potential responses, while also integrating cultural and social knowledge into determining what is most relevant. LLMs instead tokenize textual information and focus on statistical relationships between tokenized text.The experience fault: Humans and LLMs draw from prior knowledge but in a fundamentally different way. Relying on both physics and psychology, humans draw from episodicmemoryto recall details that allow us to address new situations. In terms of physics, we can determine (to varying degrees) object permanence, solidity, gravity, and causal forces, while also integrating categories, scripts, and social norms through socialization and culturalization, but also through past experience. LLMs instead rely on statistical patterns or linguistic co-occurrences. Crucially, they do not have episodic or autobiographical memory and have neither intuitive physics nor psychology.The Humans tend to be driven bymotivationfault:emotion, motivation, and goals, whereas LLMs are driven by optimization rooted in mechanistic processes. Reinforcement learning with human feedback can adjust the equation for LLMs, but it does not imbue them with the same emotional and goal orientation as humans.The causality fault: Based on these previous processes, humans then reason, by which they infer, integrate, consider counterfactuals, and form causal explanations that can be cognizant of long-term plans. LLMs instead integrate textual context that is devoid of (human) reasoning, relying especially on correlations.The metacognitive fault: Further refining their reasoning, humans evaluate metacognitively, as they consider uncertainties, detect errors, estimateconfidence, and consider withholding judgment. This can include integration of social-epistemic norms, taking into account social context and potential reputational costs. According to the authors, LLMs lackmetacognitionentirely, and this is one of the factors that contributes to hallucinations.The value fault: Human reasoning is thus a kaleidoscope of processes that can integrate personal values, cultural norms, reputational concerns, and goals, while beingmindfulof consequences for relationships and self-identity. LLMs produce probabilistic, text-based judgments driven by the statistical structure of training data.
The authors’ term—epistemia—describes the phenomenon whereby linguistic plausibility takes precedence over epistemic evaluation. This divergence is consequential, especially as humans increasingly turn to AI for answers to complex problems, as well as intimate psychological problems. The similarity between LLM outputs and human writing is one of the reasons why humans have grown comfortable "talking to" LLMs. Yet, what is going on under the hood is profoundly different. Even as LLMs become more fluent and less prone to obvious errors, this does not, according to this study, mean they are reasoning in the way that humans do.
References
Quattrociocchi, W., Capraro, V., & Perc, M. (2025). Epistemological fault lines between human and artificial intelligence. arXiv. [https://doi.org/10.48550/arXiv.2512.19466]