AI Through the Human Lens: Investigating Cognitive Theories in Machine Psychology

We investigate whether Large Language Models (LLMs) exhibit human-like cognitive patterns under four established frameworks from psychology: Thematic Apperception Test (TAT), Framing Bias, Moral Foundations Theory (MFT), and Cognitive Dissonance. We evaluated several proprietary and open-source models using structured prompts and automated scoring. Our findings reveal that these models often produce coherent narratives, show susceptibility to positive framing, exhibit moral judgments aligned with Liberty/Oppression concerns, and demonstrate self-contradictions tempered by extensive rationalization. Such behaviors mirror human cognitive tendencies yet are shaped by their training data and alignment methods. We discuss the implications for AI transparency, ethical deployment, and future work that bridges cognitive psychology and AI safety

💡 Research Summary

This paper, “AI Through the Human Lens: Investigating Cognitive Theories in Machine Psychology,” presents a comprehensive empirical study evaluating whether Large Language Models (LLMs) exhibit patterns analogous to human cognition. The research applies four well-established frameworks from psychology—the Thematic Apperception Test (TAT), Framing Bias, Moral Foundations Theory (MFT), and Cognitive Dissonance—to a range of proprietary and open-source LLMs, including GPT-4o, QwQ 72B, LLaMA 3.3 70B, Mixtral 8x22B, and DeepSeek V3.

The introduction establishes the growing deployment of LLMs in advanced reasoning tasks and raises the critical question of whether these systems replicate human cognitive biases and behavioral patterns. To investigate this, the authors designed a series of targeted experiments grounded in cognitive science.

The background section justifies the selection of the four specific frameworks. The TAT is chosen for its projective, narrative-based nature suitable for text generation. Framing Bias is selected due to its direct link to linguistic presentation and decision-making. MFT offers a multidimensional assessment of moral reasoning beyond simple dilemmas. Cognitive Dissonance theory provides a lens to examine internal coherence and how systems manage contradictory information. The motivation underscores the practical urgency of such analysis as LLMs are increasingly used in high-stakes domains like healthcare, finance, and criminal justice.

The methodology details each experimental setup. For the TAT, models generated stories based on descriptions of 30 ambiguous images. These narratives were quantitatively scored using the Social Cognition and Object Relations Scale–Global (SCORS-G) across eight dimensions (e.g., complexity of representations, affective quality) and received additional qualitative annotation via a larger LLM. The Framing Bias experiment involved 230 pairs of questions (460 total) across 46 categories, identically phrased in either a positive or negative frame. Model responses were analyzed to determine rates of contradiction (flipped answers) versus entailment (consistent answers) between frames. For MFT, the authors moved beyond the standard human self-report questionnaire, constructing a new set of 360 situational dilemmas (60 per foundation) to better probe LLMs’ evaluative moral reasoning. The Cognitive Dissonance experiment involved presenting models with contradictory prompts or information and analyzing their responses and subsequent justifications.

The key findings reveal that LLMs demonstrate several human-like tendencies. They generate coherent, emotionally nuanced narratives in the TAT task. They show significant susceptibility to framing effects, with their choices often flipping based on whether a scenario is presented in terms of gains or losses. In moral reasoning, models exhibited a pronounced emphasis on the Liberty/Oppression foundation, suggesting an influence from contemporary AI alignment objectives focused on autonomy and harm reduction. When faced with contradictions, LLMs frequently produced self-contradictory statements but then engaged in extensive rationalization to resolve or explain away the inconsistency, mirroring a form of dissonance reduction.

The paper concludes that while LLMs can superficially mimic certain human cognitive patterns, these behaviors are ultimately shaped by their training data distributions and alignment fine-tuning objectives, not by internal states of belief or consciousness. The research advocates for the emerging field of “Machine Psychology” as a valuable tool for AI transparency and interpretability. It highlights critical implications for the ethical deployment of LLMs, particularly regarding their potential to amplify biases and produce unstable outputs in sensitive applications. The work calls for further interdisciplinary research bridging cognitive psychology and AI safety to develop more robust, understandable, and trustworthy AI systems.