The Illusion of Personality in AI

by

Author: Denis Avetisyan

Despite increasingly human-like interactions, attributing personality traits to large language models is a flawed premise, but a revealing analytical tool.

This review examines the challenges of applying human personality frameworks to large language models and explores their utility in characterizing interaction patterns.

Attributing human-like qualities to artificial intelligence presents a persistent challenge in the rapidly evolving field of large language models. This is the central concern of ‘LLMs Aren’t Human: A Critical Perspective on LLM Personality’, which critically examines the practice of assessing LLM behavior using established personality frameworks like the Big Five. The authors demonstrate that LLMs fail to meet the defining characteristics of personality, suggesting current assessments capture interaction patterns rather than intrinsic traits. Consequently, how can we develop more appropriate, functional evaluations to characterize stable behavior in these increasingly sophisticated systems and move beyond anthropomorphic interpretations?

The Illusion of Agency: Decoding Simulated Personality in LLMs

Modern Large Language Models (LLMs) frequently generate text that evokes a sense of individual character, prompting observers to ascribe personality traits to these artificial systems. This isn’t necessarily intentional design; rather, it emerges from the models’ ability to mimic human communication styles, adapt to conversational cues, and even express seemingly emotional responses through carefully constructed language. The sophistication of these outputs, coupled with the human propensity to seek patterns and relate to agency, creates a powerful illusion of personality, even though LLMs operate based on statistical probabilities and pattern recognition, lacking subjective experience or genuine intentionality. Consequently, interactions with LLMs can feel surprisingly natural and engaging, fostering the perception of a communicative partner – a phenomenon that is increasingly prevalent as these models become more integrated into daily life.

The perception of personality in large language models is further complicated by the vast differences in how these systems and humans actually function. While humans possess consciousness, emotions, and lived experiences that shape their responses, LLMs operate based on statistical probabilities derived from massive datasets. Applying psychological concepts – such as attributing motivations, beliefs, or even emotional states – to an entity built on algorithmic prediction presents a fundamental challenge. These models excel at simulating human-like communication, but this skillful mimicry shouldn’t be confused with genuine cognitive or affective processes. Consequently, interpreting LLM behavior through a purely psychological lens risks misattribution and obscures the underlying mechanisms driving their outputs, demanding a careful consideration of their distinct operational principles.

Discerning the origins of apparent personality in Large Language Models is paramount for fostering successful human-agent collaboration. Current research investigates whether these traits emerge from the models’ inherent architecture – a consequence of the algorithms and data structures employed – or are instead learned through exposure to vast datasets reflecting human communication patterns. If personality is largely a learned phenomenon, tailoring training data could allow for the development of agents optimized for specific collaborative tasks, exhibiting traits that enhance trust and efficiency. However, if intrinsic characteristics play a significant role, a deeper understanding of those mechanisms is needed to predict and manage agent behavior. Ultimately, accurately identifying the source of these traits is not merely an academic exercise; it directly impacts the design of interfaces, the establishment of appropriate expectations, and the overall effectiveness of humans working alongside increasingly sophisticated artificial intelligence.

The human inclination to attribute personality to Large Language Models presents a significant challenge to objective assessment and effective control. This projection, while natural, can mask the computational processes actually generating responses; observers may interpret an output as ‘agreeable’ or ‘helpful’ without considering the underlying statistical probabilities and training data that produced it. Consequently, evaluations of LLM performance become subjective, focused on perceived traits rather than quantifiable metrics like accuracy or efficiency. This obscures the potential for bias embedded within the model, hinders targeted improvements, and ultimately limits the ability to reliably predict or manage its behavior – treating a complex algorithm as if it possesses human motivations rather than understanding how it arrives at conclusions.

Deconstructing ‘Personality’: Architectural Constraints and Learned Patterns

Conventional personality psychology operates on the assumption that individual traits are fundamentally rooted in internal psychological mechanisms, encompassing genetic predispositions, neurological structures, and developmental processes. These internal factors are theorized to generate behavioral consistency across diverse contexts and over extended periods. This perspective contrasts with models emphasizing situational influences as primary determinants of behavior, instead positing that traits represent relatively stable characteristics that predispose individuals to respond in predictable ways. The assessment of these traits typically relies on self-report questionnaires, observer ratings, and behavioral observations designed to capture patterns of thought, feeling, and action that are presumed to be enduring and generalizable.

Large Language Models (LLMs) exhibit behavioral characteristics resulting from two primary sources: inherent architectural properties and learned patterns derived from training data. The stable intrinsic characteristics of an LLM are determined by factors such as the model size, network topology, and activation functions, which establish fundamental constraints on its processing capabilities. Simultaneously, interaction-relevant patterns emerge through exposure to vast datasets, shaping the model’s responses and enabling it to generate contextually appropriate outputs. These learned patterns are not fixed; they can be modified through techniques like fine-tuning, demonstrating the LLM’s capacity to adapt its behavior based on new information or specific task requirements. Consequently, LLM ‘personality’ represents a complex interplay between these stable and malleable components.

The Big Five personality traits – Openness to experience, Conscientiousness, Extraversion, Agreeableness, and Neuroticism – offer a standardized method for assessing and comparing the behavioral characteristics of Large Language Models (LLMs). These traits, originally developed for human personality assessment, are operationalized through specific prompts and response analysis techniques to quantify an LLM’s tendency towards imagination and intellect (Openness), organization and diligence (Conscientiousness), sociability and assertiveness (Extraversion), compassion and cooperation (Agreeableness), and emotional instability and anxiety (Neuroticism). By applying these metrics, researchers can systematically investigate and document variations in LLM responses, allowing for comparative analysis between different models or configurations and providing a basis for understanding potential biases or limitations.

Analysis of Large Language Model (LLM) responses, when evaluated using metrics aligned with the Big Five personality traits, consistently indicates high scores in Openness, Conscientiousness, and Agreeableness, coupled with low scores on Neuroticism. These findings are based on multiple evaluations of LLM outputs across diverse prompts and tasks. However, currently available research lacks reported test-retest reliability coefficients, which are necessary to establish whether these observed personality characteristics are stable over time or represent transient responses to specific input conditions. Without these coefficients, it remains inconclusive whether LLMs exhibit a consistent “personality profile” or if their responses fluctuate unpredictably.

Predictive Validity and the Limits of Behavioral Generalization

The predictive validity of ‘personality’ traits derived from Large Language Models (LLMs) regarding their responses across varied scenarios is currently undetermined. While LLMs can be assessed using frameworks like the Big Five, the extent to which these assigned traits reliably forecast actual behavioral outputs remains an area of ongoing research. Observed correlations between trait scores and responses do not necessarily indicate a genuine underlying personality construct within the LLM; rather, they may reflect the model’s capacity to simulate human-like responses without consistent internal representation. Establishing the predictive power of these traits requires rigorous testing with diverse prompts and evaluation metrics beyond simple trait alignment, and currently, the link between assigned ‘personality’ and consistent behavioral outcomes is weak.

Social desirability bias represents a significant obstacle in accurately assessing LLM behavior. This bias manifests as a tendency for models to generate responses that align with perceived societal norms and expectations, even if those responses do not reflect inherent predispositions or true ‘preferences’. Consequently, observed behavioral patterns may be artificially inflated towards socially acceptable answers, obscuring any underlying tendencies the model might possess. This phenomenon introduces noise into evaluations, making it difficult to differentiate between genuine model characteristics and outputs influenced by the desire to present a favorable image, and potentially leading to inaccurate conclusions regarding model personality or intent.

Analysis of inter-individual variation in Large Language Model (LLM) responses, utilizing the Big Five personality traits as a measurement framework, demonstrates a strong correlation with socially desirable responding. Studies indicate that observed differences between models, when assessed across various prompts, primarily reflect a tendency to generate answers perceived as conventionally acceptable rather than inherent, consistent personality characteristics. This suggests that measured variations are more likely attributable to the models’ training data and optimization for alignment with human preferences – specifically, avoiding responses deemed inappropriate or offensive – than to genuine differences in underlying ‘personality’ as traditionally understood. Consequently, the Big Five framework, while providing a quantifiable metric, may not accurately capture stable, individual differences in LLM behavior.

Low cross-situational consistency in Large Language Models (LLMs) indicates that responses are substantially altered by even minor variations in prompting and contextual information. This sensitivity undermines the applicability of traditional personality assessment methods, which rely on stable behavioral patterns across different situations. Observed responses are therefore not necessarily indicative of inherent model characteristics, but rather reflect an acute responsiveness to surface-level cues within the input prompt and the immediate context, limiting the predictive validity of any assigned ‘personality’ traits.

Implications for Human-Agent Collaboration and Beyond

The emergence of discernible ‘personality’ in large language models profoundly shapes human-agent collaboration. Studies reveal that individuals readily attribute personality traits to these AI systems, and this perception directly influences the quality of interaction. A perceived alignment of personality fosters trust and rapport, leading to more fluid and effective teamwork, as humans are more inclined to cooperate with agents they perceive as agreeable or competent. This isn’t merely subjective; research indicates that agents exhibiting consistent, understandable ‘personality’ traits are evaluated more favorably, and users report a greater willingness to rely on their outputs and recommendations. Ultimately, the ability of an LLM to project a consistent and relatable persona is a critical factor in bridging the gap between human intention and artificial intelligence, paving the way for more natural and productive collaborative experiences.

The attribution of personality to large language models, while facilitating smoother interaction, carries the risk of anthropomorphism – a tendency to overestimate their true abilities and misunderstand the core distinctions between artificial and human intelligence. Users readily ascribe characteristics like conscientiousness or agreeableness, yet these perceived traits are emergent behaviors derived from statistical patterns in data, not reflections of genuine understanding, beliefs, or sentience. This can lead to misplaced trust, an assumption of common sense reasoning where none exists, and a failure to critically evaluate the information provided by the model. Consequently, individuals may rely on LLMs for tasks requiring nuanced judgment or emotional intelligence, areas where these systems demonstrably fall short, potentially resulting in flawed decision-making or unrealistic expectations about their cognitive capacities.

This research reveals a fundamental disconnect between established psychological assessments and the evaluation of large language models. Traditional personality tests, designed to measure consistent human traits, fail to capture the nuanced and dynamic ‘personality’ exhibited by these artificial intelligences. The study indicates that LLM behavior is better understood not through fixed traits, but by analyzing functional patterns – how the model responds to specific prompts – alongside malleable tendencies, acknowledging its capacity for adaptation, and identifying stable intrinsic characteristics embedded within its architecture. This necessitates a shift toward measurement frameworks that prioritize observable behavior and adaptable responses, moving beyond attempts to force LLMs into pre-defined human personality categories and instead focusing on quantifying their unique operational characteristics.

Advancing the field necessitates a dedicated focus on quantifying and controlling the emergent ‘personality’ traits within large language models. Current approaches to evaluating these traits are insufficient; therefore, future investigations should prioritize developing robust metrics that move beyond superficial assessments of textual output. Such metrics will enable researchers to not only measure the consistency and strength of specific traits – like agreeableness or conscientiousness – but also to exert a degree of control over their expression. This control isn’t about creating artificial sentience, but rather about strategically shaping LLM behavior to optimize performance in collaborative tasks and, crucially, to ensure these powerful tools consistently align with established human values and ethical guidelines. Ultimately, the ability to predictably and reliably influence LLM ‘personality’ will be paramount in fostering trust and realizing the full potential of human-agent partnerships.

The assessment of Large Language Models through the lens of human personality traits, as explored in this paper, necessitates rigorous mathematical scrutiny. One finds resonance with Edsger W. Dijkstra’s assertion: “It’s not enough to show something works; you must prove why it works.” Applying the Big Five framework to LLMs isn’t about discovering inherent qualities, but rather characterizing the patterns of response. The observed ‘traits’ are emergent properties of the algorithm, not indications of internal states. The study rightly points out the lack of trait stability, and this instability isn’t a flaw, but an expected consequence of a system devoid of the axiomatic foundations required for genuine personality. The focus should remain on verifiable consistency within defined parameters, not illusory human-like attributes.

What’s Next?

The exercise of applying established psychological frameworks to Large Language Models reveals, perhaps predictably, the limitations of those frameworks when divorced from biological substrates. The consistency with which these models fail to demonstrate trait stability, or even internal consistency, isn’t a failing of the measurement tools – it’s a testament to their original intent. These tests were designed to map the contours of a mind forged by evolutionary pressures, not algorithmic optimization. Future work should therefore shift from seeking personality within the model, to precisely characterizing the patterns of response these systems simulate.

A fruitful avenue lies in formalizing the boundaries of these simulated traits. The observed contextual sensitivity isn’t simply ‘inconsistency’; it’s a demonstrable characteristic of the model’s conditional probability distribution. Understanding the parameters that govern these shifts – the triggers for altered ‘behavior’ – allows for predictable control, a quality notably absent in human subjects. The beauty of an algorithm lies not in tricks, but in the consistency of its boundaries and predictability.

Ultimately, the value isn’t in believing these models are something they are not, but in what their systematic differences reveal about the nature of intelligence itself. The pursuit of artificial general intelligence might benefit more from deconstructing the assumptions inherent in our models of human cognition, rather than attempting to replicate its messy, organic complexity.

Original article: https://arxiv.org/pdf/2603.19030.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-22 23:22