PaperlayerView exact revision on arXiv

Source-linked AI summary

PersonaLLM: Investigating the Ability of Large Language Models to Express Personality Traits

Hang Jiang, Xiajie Zhang, Xubo Cao, Cynthia Breazeal, Deb Roy, Jad Kabbara

arXiv:2305.02547v5cs.CLcs.AIcs.HC

TL;DR

Evidence remains limited on whether LLM-based agents can exhibit particular personality traits and how those traits relate to their language and human perception. This paper tests GPT-3.5 and GPT-4 personas through personality inventories, story writing, and automated and human evaluation; personas matched assigned traits in self-reports and showed trait-related language patterns, while human prediction accuracy declined when evaluators knew the stories were AI-generated.

  • Problem

    Evidence remains limited on LLM agents’ ability to exhibit specific personality traits and how those abilities relate to their linguistic behavior and human perception.

    Source: §4.2 · p. 8

  • Method

    The study prompts GPT-3.5 and GPT-4 personas to complete the BFI and write stories, then analyzes the stories with LIWC and automated and human evaluators.

    Source: §2 · p. 2, §5 · p. 8

  • Results

    LLM personas matched assigned personalities in BFI responses and produced trait-characteristic linguistic patterns; human personality-prediction accuracy decreased when evaluators knew authorship was AI-generated.

    Source: §5 · pp. 8–9

  • Takeaways & Limitations

    Personality expression was observable across self-reports and story language, while human perception of personality depended partly on awareness of AI authorship.

    Source: §5 · pp. 8–9

    The study focuses mostly on closed GPT models and uses GPT-4 for human evaluation, limiting that evaluation’s model scope.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Despite the many use cases for large language models (LLMs) in creating personalized chatbots, there has been limited research on evaluating the extent to which the behaviors of personalized LLMs accurately and consistently reflect specific personality traits. We consider studying the behavior of LLM-based agents which we refer to as LLM personas and present a case study with GPT-3.5 and GPT-4 to investigate whether LLMs can generate content that aligns with their assigned personality profiles. To this end, we simulate distinct LLM personas based on the Big Five personality model, have them complete the 44-item Big Five Inventory (BFI) personality test and a story writing task, and then assess their essays with automatic and human evaluations. Results show that LLM personas' self-reported BFI scores are consistent with their designated personality types, with large effect sizes observed across five traits. Additionally, LLM personas' writings have emerging representative linguistic patterns for personality traits when compared with a human writing corpus. Furthermore, human evaluation shows that humans can perceive some personality traits with an accuracy of up to 80%. Interestingly, the accuracy drops significantly when the annotators were informed of AI authorship.

1 Introduction

As personalized AI agents become more common, whether they consistently express assigned personalities and how people perceive those expressions remain underexplored. This study evaluates Big Five LLM personas through personality testing, story analysis, and human and LLM judgments.

1 Introduction

  • Prior work has studied LLM personality expression, measurement, and prompting, but psychometric evaluation of faithful trait expression and human perception remains limited.The introduction frames consistent human-like behavior by generative characters as an unsubstantiated assumption.Source: §1 · p. 1
  • The study defines LLM personas as agents prompted to reflect specified traits and tests whether they express assigned Big Five profiles on the BFI and in narratives.The five traits are Extraversion, Agreeableness, Conscientiousness, Neuroticism, and Openness to Experience.Source: §1 · pp. 1–2
  • The authors examine psycholinguistic patterns in persona-generated stories and whether personas use lexicons associated with their assigned profiles.They propose extracting features with dictionaries designed to analyze human behavior.Source: §1 · p. 2
  • The study asks how humans and LLM raters evaluate persona stories and whether humans and LLMs can infer Big Five traits from them.It also identifies multi-round dialogue and action planning as potential extensions to more real-life evaluation scenarios.Source: §1 · p. 2

2 Experiment Design

The study creates Big Five LLM personas, collects self-reported personality assessments and personal stories, and evaluates the stories through linguistic analysis and human and LLM judgments. Its evaluation also tests whether awareness of AI authorship affects story ratings and personality predictions.

2.1.1 Model Settings

  • GPT-3.5 and GPT-4 were selected for the main paper because they aligned more effectively with assigned personas; temperature was 0.7, with other settings at defaults.The study also assessed state-of-the-art open- and closed-source LLMs, with LLaMA-2 results reported in Appendices F and G.Source: §2.1.1 · pp. 2–3

2.1.2 LLM Persona Simulation

  • The researchers created 320 personas per model by simulating 10 for each combination of binary Big Five types, using a system prompt specifying one descriptor per trait.Descriptor pairs included extroverted/introverted, agreeable/antagonistic, conscientious/unconscientious, neurotic/emotionally stable, and open/closed to experience.Source: §2.1.2 · p. 3

2.1.3 BFI Personality Test

  • Each persona completed the 44-item Big Five Inventory, and its valid 1–5 agreement responses were aggregated into five personality scores for analysis.Responses were accepted only in the instructed “(x) y” format.Source: §2.1.3 · p. 3

2.1.4 Storywriting

  • Each of the 320 personas was prompted to write an 800-word personal story without explicitly mentioning its personality traits.The prompt was simplified after trying variants to reduce demand characteristics and support generalizability.Source: §2.1.4 · p. 3

2.2 Evaluation Methods

  • Human evaluation used GPT-4 persona stories without explicit trait references because such references were frequent in GPT-3.5 stories.The lexicon-based classifier identified explicit trait references in 96.56% of GPT-3.5 stories versus 31.87% of GPT-4 stories, so final human evaluation focused on GPT-4.Source: §2.2 · p. 3

2.2.1 LIWC Analysis

  • The study compared LIWC features’ correlations with assigned traits in LLM stories and human essays to examine whether linguistic markers were consistent across writers.The human comparison corpus was the Essays dataset 6.Source: §2.2.1 · pp. 3–4

2.2.2 Story Evaluation

  • For story evaluation, five human raters and GPT-3.5 and GPT-4 evaluators assessed 32 GPT-4 persona stories across six dimensions.Stories were sampled from each personality type and excluded if they explicitly mentioned a personality trait; ratings covered readability, personalness, redundancy, cohesiveness, likeability, and believability on a 1–5 scale.Source: §2.2.2 · p. 4

2.3 Personality Prediction

  • For personality prediction, human annotators and LLM evaluators inferred the writers’ Big Five traits from the same 32 stories on a 1–5 scale.Human evaluators received trait descriptions from John et al. (1999) as references.Source: §2.3 · p. 4

3 Results

LLM personas’ BFI scores and writing patterns reflected assigned traits, and evaluators could perceive some traits in their stories. Awareness of AI authorship was associated with lower personality-prediction accuracy and weaker correlations between perceived traits and BFI scores.

3.1 RQ1: Behavior in BFI Assessment

  • GPT-3.5 and GPT-4 personas showed significant BFI score differences across all five assigned traits, with large effect sizes.GPT-3.5 effect sizes were EXT: d = 7.81, p < .001; AGR: d = 5.93, p < .001; CON: d = 1.56, p < .001; NEU: d = 1.83, p < .001; OPN: d = 2.90, p < .001. GPT-4 effect sizes were EXT: d = 5.47, p < .001; AGR: d = 4.22, p < .001; CON: d = 4.39, p < .001; NEU: d = 5.17, p < .001; OPN: d = 6.30, p < .001. For both models, BFI scores were lower for negative than positive assigned traits across dimensions.Source: §3.1 · p. 4, Figure 2 · p. 4

3.2 RQ2: Linguistic Patterns in Writing

  • Assigned personality types influenced LIWC patterns, and GPT-4 writing showed more overlap with human correlations than GPT-3.5 writing.For both models, openness correlated positively with curiosity lexicons, neuroticism with anxiety, negative-tone, and mental-health lexicons, and extraversion with positive-tone and affiliation lexicons. For Conscientiousness and Openness, GPT-3.5 shared 1/31 and 2/36 significant correlations with humans, versus 11/31 and 17/36 for GPT-4. Achievement striving correlated positively with LLM personas but not significantly in human writing; sadness correlated negatively for GPT-3.5 but positively for GPT-4 and humans. The human corpus is a comparative reference, not an absolute standard, because prompts were not identical.Source: §3.2 · pp. 5–6, Table 1 · p. 5

3.3 RQ3: Story Evaluation

  • GPT-4 persona stories received ratings close to or above 4.0 for readability, cohesiveness, and believability from human and LLM evaluators.Human evaluators rated personalness highly but likeability lower; the GPT-4 evaluator rated stories highest across dimensions, while GPT-3.5 gave lower redundancy and personalness ratings than humans and GPT-4. When told the stories were AI-generated, humans’ ratings stayed consistent for readability, redundancy, cohesiveness, likeability, and believability, but perceived personalness dropped; GPT-3.5 shifted toward higher readability and lower personalness ratings, whereas GPT-4 ratings varied minimally.Source: §3.3 · p. 6, Table 2 · p. 6

3.4 RQ4: Personality Perception

  • RQ4 assessed personality perception using binary trait-inference accuracy and correlations between human judgments and personas’ BFI scores.Source: §3.4 · p. 6

3.4.1 Personality Prediction

  • Majority voting raised human accuracy to 0.84 for Extraversion and 0.69 for Agreeableness, while GPT-4 reached 0.97 for Extraversion.Uninformed individual human accuracy was 0.68 for Extraversion and 0.51 for Agreeableness, and close to random on other dimensions; human accuracy decreased to varying degrees when evaluators knew of AI authorship. GPT-4 accuracy was 0.68 for Agreeableness and 0.69 for Conscientiousness.Source: §3.4.1 · p. 7

3.4.2 Correlation with BFI Scores

  • Human perceptions correlated significantly with persona BFI scores for all five traits when evaluators were unaware of AI authorship, but only four when informed.Uninformed correlations were EXT: r = .64, p < .001; AGR: r = .33, p < .001; CON: r = .26, p < .001; NEU: r = .23, p < .005; OPN: r = .22, p < .005. Informed correlations were EXT: r = .42, p < .001; AGR: r = .32, p < .001; CON: r = .20, p < .05; NEU: r = .17, p < .05; Openness was not significant.Source: §3.4.2 · p. 7

4 Related Work

Prior work connects personality with human language and explores personality in NLP and LLM agents, but has not examined LLM personas’ linguistic behavior and human-perceived personality. This study addresses that gap through story evaluation and human and LLM assessment.

4.1 Personality and Language Use

  • Personality theories and instruments quantify individual differences, and prior human-language research links language use to personality; this study extends that inquiry to LLMs.Source: §4.1 · p. 7

4.2 LLMs as Simulated Agents

  • LLMs have been studied as agents for reasoning, role-play and social-science experiments, generally via prompting, but personality expression and its effects on language and perception remain understudied (Andreas, 2022; Wang et al., 2023b).Source: §4.2 · pp. 7–8

4.3 Personality in NLP

  • Prior NLP work spans personality prediction, digital-footprint inference, personalized dialogue and stylistic transfer, while LLM studies examine prediction, trait expression, benchmarks and prompting (Mairesse et al., 2007; Ganesan et al., 2023; Li et al., 2022; Jiang et al., 2022a).The authors identify a gap in research on LLM personas’ linguistic behavior and human perception of personality-conditioned content; their study uses story evaluation and human and LLM assessment.Source: §4.3 · p. 8

5 Conclusion

The study examines whether GPT-3.5 and GPT-4 personas express assigned Big Five profiles in BFI responses and stories. It finds consistent profile-aligned answers and writing, while AI-authorship awareness reduces judges’ accuracy in predicting personality traits.

5 Conclusion

  • GPT-3.5 and GPT-4 personas consistently tailored BFI answers to assigned personalities and wrote with linguistic features characteristic of those traits.The authors also found distinct representative writing behaviors for each trait and notable word-usage alignment between human and LLM writing.Source: §5 · p. 8
  • LLM-persona stories were fluent, structurally cohesive, and believable, while human ratings of readability, redundancy, cohesiveness, likeability, and believability stayed consistent regardless of AI-authorship awareness.The authors also report notable word-usage alignment between human and LLM writing.Source: §5 · pp. 8–9
  • Human judges predicted expressed personality traits with varying accuracy across traits, and accuracy decreased to varying degrees when they knew the content was AI-generated.The study investigated GPT-3.5 and GPT-4 personas using the BFI, story writing, psycholinguistic analyses, human evaluation, and personality prediction.Source: §5 · pp. 8–9

Limitations

The study is limited by its reliance on closed GPT models, a modest story dataset, and English assessment and writing tasks rather than naturalistic interaction. It also leaves open how annotator characteristics and awareness of AI authorship affect personality judgments.

Focus on Closed Models

  • The study largely evaluates closed GPT models because LLaMA 2 generated repetitive, instruction-poor stories and explicitly used personality lexicons, making its output unsuitable for human evaluation.Given budget constraints, the authors selected GPT-4 for human evaluation.Source: Limitations · p. 9

Data Size

  • The dataset is modest: it contains 10 stories per personality type, or 160 stories per trait, and trait-level analyses use 160 instances rather than 10 per type.Larger datasets would have increased human-evaluator recruitment costs; the authors used temperature .7 to add variation and argue the trait-level sample sizes support analysis.Source: Limitations · p. 9

Task & Language Variety

  • Evaluation covers personality assessment and writing only in English, omitting more naturalistic settings such as human interaction and collaboration among LLM personas.Source: Limitations · p. 9

Evaluation & Interaction

  • Future work should examine how annotator personality and background shape subjective personality judgments, why AI authorship awareness lowers assignment accuracy, and how modalities or personality affect agent planning.The authors also suggest investigating whether annotator personality causally relates to perceptions of AI agents.Source: Limitations · p. 9

Ethical Considerations · A Story Examples Generated by LLM · B Story Comments

The sections document ethical safeguards and argue for transparency about AI authorship, while presenting selected personality stories and comments that illustrate how evaluators responded under informed and uninformed conditions.

Ethical Considerations

  • The study followed human-research and artifact-use requirements, compensated Prolific annotators at $15 per hour, excluded harmful or offensive stories, and collected no personal identifiers.It reports IRB Exempt status, adherence to the ACL Code of Ethics, and use of the Essays dataset solely for research.Source: Ethical Considerations · p. 9
  • The authors warn that personified agents could be misused and advocate disclosing AI use, noting that awareness of AI authorship greatly affected perceived personalness and personality traits.They state that disclosure can support trust between individuals.Source: Ethical Considerations · pp. 9–10
  • The authors frame story writing as a vehicle for scientific inquiry into LLM expressivity and human personality perception, not as an application-specific evaluation.They urge stakeholders to remain vigilant and proactive in mitigating AI misuse.Source: Ethical Considerations · p. 10

A Story Examples Generated by LLM

  • The examples were selected from human-evaluation results, with one story chosen for each personality trait for relatively high human-inference accuracy; unbalanced trait predictions left some selected stories lower in accuracy among five annotators.In Table 3, x/5 denotes x of five annotators predicting correctly.Source: A Story Examples Generated by LLM · p. 13, Table 3 · p. 14

B Story Comments

  • Human-evaluation comments averaged 0.45 sentiment when evaluators knew of AI authorship versus 0.16 when they did not, alongside 104 informed and 122 uninformed valid comments.Average comment lengths were 12.7 and 13.4 words, respectively; the authors interpret the higher informed-condition sentiment as a more lenient attitude toward the stories.Source: B Story Comments · p. 15
  • Informed-condition comments mixed criticism with sympathy, positive reactions, and surprise, including praise for stories that seemed genuine or enjoyable despite AI authorship.Other comments called stories basic, exaggerated, or weak in their endings.Source: B Story Comments · p. 15, Table 4 · p. 16
  • Uninformed comments criticized readability and consistency but also described stories as personal, moving, believable, and enjoyable; readers sometimes guessed the writer’s personality or questioned the motive for writing.The authors identify potential confusion when AI-generated content or characters are encountered without disclosure.Source: B Story Comments · pp. 15–16, Table 4 · p. 16

C Personality Ratings

The evaluation found that GPT-3.5 and GPT-4 ratings differed in how closely they matched the expected midpoint of 3, with the pattern varying by trait. Agreement among annotators was also reported across metrics and personality traits.

C Personality Ratings

  • The evaluation sampled 32 LLM personas from 32 personality types, with 16 positive-labeled and 16 negative-labeled personas for each personality, ideally yielding average ratings near 3.Source: C Personality Ratings · p. 16
  • For Extraversion, GPT-3.5 and GPT-4 ratings were closer to 3 than human ratings; for other traits, GPT-4 ratings were further from 3 than human and GPT-3.5 evaluator ratings.The ratings came from evaluations of GPT-4-generated personal stories.Source: Table 5 · p. 16, C Personality Ratings · p. 16
  • Inter-annotator agreement among five annotators was reported using Krippendorff’s α across six metrics and five personality traits.The tables also report the average percentage of annotators agreeing on the most-voted rating.Source: Table 6 · p. 16, Table 7 · p. 16

D Story Evaluation Details … E LLM as Evaluators

Stories were filtered for explicit personality-trait language before human evaluation, where the subjective task showed low inter-annotator agreement among five annotators.

D.1 Filtering Stories for Evaluation

  • A lexicon-based classifier flagged stories with explicit personality-trait terms, including extroversion, agreeableness, conscientiousness, neuroticism, and openness terms.The lexicons also included antonyms and phrases such as “emotionally stabl*” and “closed to experience.”Source: §D.1 · pp. 16–17
  • Stories containing those lexicons were filtered out, and remaining stories were sampled for human evaluation.Source: §D.1 · p. 17

D.2 Prolific Setup

  • The study recruited US-based Prolific workers whose first language was English and whose approval rates were 99%–100%, then divided 32 stories into four batches of eight.Source: §D.2 · p. 17
  • Annotators rated readability, personalness, redundancy, cohesiveness, likeability, and believability, then answered questions about all five Big Five traits.Source: §D.2 · p. 17

D.3 Inter-annotator Agreement

  • The subjective story-evaluation task had low inter-annotator agreement among five annotators; the paper reports scores for six metrics in Table 6 and five traits in Table 7.The supplied passages do not include the agreement values.Source: §D.3 · pp. 17, 19
  • The low inter-rater agreement echoes prior social-computing labeling findings and motivates techniques for handling annotation disagreements.Source: §D.3 · p. 19

D.4 Annotator Demographics

  • The evaluation involved 39 unique US-based participants: 37 were born in the United States and two in Nigeria.The paper also reports age, sex, and ethnicity distributions in Figure 18.Source: §D.4 · p. 19

E.1 Temperature

Higher temperatures were associated with lower LLM-evaluator ratings and greater rating variance, so the study used temperature 0 for reproducibility.

E.1 Temperature

  • LLM-evaluator ratings were negatively correlated with temperature, and higher temperatures increased variance among three evaluators, consistent with Chiang and Lee (2023b).The experiment tested GPT-3.5 and GPT-4 evaluators across temperatures.Source: §E.1 · pp. 19–20, Table 8 · p. 21
  • The study set temperature to 0 to make results more deterministic and reproducible.Source: §E.1 · pp. 20–21

F LLaMA 2 Results in BFI Scores · G Additional Results in Personality Traits · G.1 GPT-3.5 Personas

LLaMA 2 personas showed smaller BFI score divergence and effect sizes than GPT personas, while additional GPT-3.5 analyses linked Big Five scores to distinct writing features.

F LLaMA 2 Results in BFI Scores

  • LLaMA 2 personas showed statistically significant differences across all BFI dimensions, but smaller effect sizes and less score divergence than GPT personas.Table 9 reports statistics for high- and low-trait personas, but the supplied passages do not provide its cell values.Source: F LLaMA 2 Results in BFI Scores · p. 21, Table 9 · p. 21

G Additional Results in Personality Traits

G.1 GPT-3.5 Personas

  • Extroverted GPT-3.5 personas’ writing correlated with social and prosocial behavior, positive affect and tone; introversion correlated with authenticity.Correlations were social ρ = 0.27 and prosocial ρ = 0.18; authenticity correlated negatively (ρ = −0.40). Extroversion also correlated with affect (ρ = 0.46) and positive tone (ρ = 0.33); all reported p < .005 or smaller.Source: §G.1 · pp. 21–22
  • Agreeableness correlated with positive affect and tone, fewer conflict and differentiation words, more prosocial words, and less authenticity.Reported correlations were emo_neg ρ = −0.66, tone_pos ρ = 0.50, conflict ρ = −0.66, differ ρ = −0.39, prosocial ρ = 0.34, and authentic ρ = −0.24 (all p < .001).Source: §G.1 · p. 22
  • Conscientiousness correlated with more analytical writing and less negation and power language, while unconscientiousness correlated with negative tone, emotion, anger, and conflict.Reported correlations were analytic ρ = 0.22, negate ρ = −0.26, power ρ = −0.24, tone_neg ρ = −0.40, emo_neg ρ = −0.39, emo_anger ρ = −0.43, and conflict ρ = −0.41 (all p < .001).Source: §G.1 · p. 22
  • Neuroticism correlated with mental-health words, negative emotion and tone, and tentative language; openness correlated with curiosity, positive tone, affiliation, leisure, and fewer conflict words.For neuroticism, mental ρ = 0.46, emo_neg ρ = 0.26, tone_neg ρ = 0.22, and tenta ρ = 0.18; emotional stability correlated with memory (ρ = −0.15). Openness correlations were curiosity ρ = 0.28, conflict ρ = −0.17, affiliation ρ = 0.16, and leisure ρ = 0.21; p-values ranged from < .01 to < .001.Source: §G.1 · p. 22

G.2 GPT-4 Personas · G.3 LLaMA 2 Personas

GPT-4 and LLaMA 2 persona writings show linguistic patterns associated with several Big Five traits, including differences in tone, social language, and word use. LLaMA 2 shows comparatively few significant linguistic features for conscientiousness.

G.2 GPT-4 Personas

  • GPT-4 extroversion and agreeableness are reflected in more positive tone, with extroversion also linked to future-focused, reward-related language and agreeableness to prosocial language and fewer conflict words.Introverted personas use more first-person singular language and describe perception more often; agreeable personas use more affiliative language, while antagonistic personas use more power- and ownership-related words.Source: §G.2 · p. 22
  • GPT-4 conscientiousness is associated with prosocial, less negative, and more achievement- and work-related writing, while neuroticism is associated with negative emotion and more first-person singular language.Conscientious writings also score lower on perceived genuineness; emotionally stable personas use more prosocial and social language and score higher on genuineness.Source: §G.2 · p. 22
  • GPT-4 openness is associated with curiosity, analytical thinking, insight, longer words, and more words per sentence.Open-minded personas’ writings also show more allure-related language.Source: §G.2 · pp. 22–23

G.3 LLaMA 2 Personas

  • LLaMA 2 extroversion is associated with more positive tone, social references, affiliations, exclamations, and drive-related words.The reported association with exclamations is described as weak.Source: §G.3 · p. 23
  • LLaMA 2 agreeableness is associated with positive tone, friend-related language, affiliation, and “we” rather than “they,” alongside less work-related and conflict language.The passage also describes a slight negative correlation with prosociality.Source: §G.3 · p. 23
  • LLaMA 2 conscientiousness has few significant linguistic features, with only weak associations involving home, fulfillment, familiar, and friendly words.The passage characterizes this as similar to the GPT-3.5 and GPT-4 trend.Source: §G.3 · p. 23
  • LLaMA 2 neuroticism is associated with less positive emotion and fewer friend, leisure, and social references, but more work- and insight-related language.The writings also show a more negative tone.Source: §G.3 · p. 23
  • LLaMA 2 openness is associated with positive emotion, affiliations, curiosity, and social references, but less work- and achievement-related language.Source: §G.3 · p. 23

Related papers

Loading 2305.02547v5…