Identifying and Manipulating the Personality Traits of Language Models

1 Introduction

• Language-model personality remains underexamined, although understanding it could support more tailored and amiable interactions.Applications include personality-matched email autocomplete and dialog systems that avoid reinforcing depressive or hostile behavior.Source: §1 · p. 1
• The work addresses personality bias alongside established research on racial and gender bias, while differing from contemporaneous studies by evaluating human-like psychological assessments.The cited contemporaneous works are Karra et al. (2022) and Jiang et al. (2022).Source: §1 · p. 2
• The study applies language-based psychological questionnaires to model-generated text without assuming that models are anthropomorphic.It interprets perceived personality through human psychological assessments and studies how context influences that perception.Source: §1 · p. 2
• Context controls Big Five personality traits predictably, with median Pearson correlations up to 0.84 for BERT and 0.81 for GPT2.The paper evaluates expected versus observed personality changes across different contexts.Source: Figure 1 · p. 1, §1 · p. 2
• Using self-reported human personality descriptions as context, models predict human personality scores with correlations up to 0.48.These results frame language models as tools for identifying personality traits and controlling personas in dialog systems.Source: §1 · p. 2
• The paper contributes paired human personality descriptions and Big Five assessments, plus personality descriptions collected from Reddit.These datasets support experiments on personality identification and contextual control.Source: §1 · p. 2

2 ‘Big Five’ Preliminaries

• The Big Five groups personality into extroversion, agreeableness, conscientiousness, emotional stability, and openness to experience.The paper describes these traits through behavioral tendencies including sociability, compassion, organization, anxiety, and creativity.Source: §2 · p. 2
• The study chooses the Big Five over alternatives such as MBTI and the Enneagram because it is the most commonly used personality-assessment taxonomy.The authors also state that it is relatively natural to cast as an assessment for language models.Source: §2 · pp. 2–3
• The Big Five has been shown to predict educational attainment, longevity, and relationship satisfaction in human research.The cited outcomes motivate its practical relevance as the paper’s assessment basis.Source: §2 · p. 3

3 Experiment Design

• BERT and GPT2 answer 50 modified Big Five items using five adverb choices mapped to scores from 1 to 5.The questionnaire changes statements such as “I am the life of the party” into sentence-completion prompts.Source: §3 · p. 3
• BERT selects the highest-probability masked-token answer, whereas GPT2 selects the candidate producing the highest-probability complete sentence.The distinction accommodates BERT’s masked-language modeling and GPT2’s autoregressive generation.Source: §3 · p. 3
• Each Big Five trait receives an integer score from ten questionnaire items, with a maximum score of 40.Model scores are also reported as percentiles relative to Big Five distributions from about 1,015,000 people.Source: §3 · p. 3

4 Base Model Trait Evaluation

• None of BERT’s or GPT2’s base trait scores lies outside the human population spread, and every score is within 26 percentile points of the human median.The authors interpret this as evidence that models mirror some personality markers reflected in pretraining data.Source: §4 · p. 3
• BERT’s openness percentile is 24 and GPT2’s agreeableness percentile is 25, while GPT2’s conscientiousness and emotional stability percentiles are 73 and 71.These are the largest specifically noted deviations from the population median.Source: §4 · p. 3

5.1 Analysis With Assessment Item Context

• 0.40 for BERT and 0.54 for GPT2: assessment-item context changes correlated significantly with expected behavior (p < 0.01, t-test).High-polarity quantifiers produced the largest changes, and conscientiousness and emotional stability showed the strongest per-item correlations.Source: §5.1 · pp. 4–5
• Assessment-item context can create spurious correlations because the same item appears during both context construction and scoring.Double-negative context/modifier pairs may also produce misleading negative correlations for some traits.Source: §5.1 · p. 5
• Adjusted assessment-item analyses retained significant correlations of 0.25 for BERT and 0.40 for GPT2, despite slightly lower values.The adjustment removed the model response corresponding to the context item from the score calculation.Source: §5.1 · p. 5, §5.2 · p. 5

5.2 Analysis With Reddit Context

• Reddit-context features generally shifted personality scores in human-interpretable directions, although several GPT2 and weaker-trait cases did not.Examples include “friendly” for extroversion, “love” for agreeableness, “hang out with” for openness, and “lack of motivation” for conscientiousness.Source: §5.2 · pp. 5–6

5.3 Analysis With Psychometric Survey Data

• 0.48 for GPT2 and 0.44 for BERT: survey descriptions correlated with subjects’ actual psychometric personality scores in directed responses.Correlations differed only marginally between directed and undirected datasets, supporting the potential use of language models as probes of personality in free text.Source: §5.3 · p. 7
• Survey-context correlations were positive but smaller than item-context correlations because survey descriptions were open-ended rather than trait-targeted.The survey study retained 404 responses after filtering and used directed and undirected description datasets.Source: §5.3 · pp. 6–7

5.4 Observed Ranges of Personality Traits

• Contexts produced substantial personality-score ranges, spanning nearly the full human percentile range for conscientiousness and emotional stability while struggling to produce high agreeableness.All five traits reached scores below the 10th percentile, and the three context types showed no major differences in their observable ranges.Source: §5.4 · p. 8

• The analysis modified the assessment by replacing first-person prompts with common male and female names and compared 20 names per gender.Source: §6 · p. 8
• Female-name scores were higher for agreeableness, conscientiousness, and emotional stability in both BERT and GPT2.Source: §6 · p. 8
• Male-name scores were higher only for GPT2’s extroversion and openness to experience.Source: §6 · p. 8
• The observed emotional-stability pattern diverged from literature suggesting that women show lower emotional stability.Source: §6 · p. 8

• The paper presents a simple approach for controlling language-model personality traits and using language-based question answering to predict human personality traits.Source: §7 · p. 8
• These capabilities frame language models as tools for identifying personality traits and controlling personas in dialog systems.Source: §7 · p. 8
• The findings’ generalization to other models, especially larger models such as GPT3 and MT-NLG, remains unclear.Source: §7 · p. 9
• The role of pretraining data in shaping language-model personality traits is left for future exploration.Source: §7 · p. 9

Ethics and Broader Impact

• The datasets used public Reddit threads and crowd-sourced AMT responses, with no stored personally identifiable markers and participant compensation described as fair.Source: Ethics and Broader Impact · p. 9
• Understanding and controlling AI personalities may have lasting personal impacts in applications such as education and mental health.Source: Ethics and Broader Impact · p. 9
• Personality manipulation creates a serious risk that systems could be attacked to become hostile or disagreeable toward users.Source: Ethics and Broader Impact · p. 9
• This risk motivates considering methods that make language models immune to such attacks before deployment.Source: Ethics and Broader Impact · p. 9

Appendix A Model Background

• The study used BERT and GPT2 without fine-tuning, with BERT configured for masked language modeling and GPT2 for autoregressive language modeling.Source: Appendix A · p. 11

Appendix B Experiment Design Items

• Figure 7 presents the human distributions of Big Five trait scores for comparison.Source: Figure 7 · p. 11
• The assessment items cover extroversion, agreeableness, conscientiousness, emotional stability, and openness to experience using statements with a blank response position.Source: Appendix B · p. 12
• The scoring-procedure table records how the Big Five personality items are converted into trait scores.Source: Table B.7 · p. 13

Appendix C Item Context Evaluation Tables

• Table C.7 summarizes statistics from Δcm versus rcm plots across all traits, including means, medians, standard deviations, and confidence intervals.Source: Table C.7 · p. 13

• The appendix lists the domain names of Reddit threads scraped to collect context.Source: Table D.7 · p. 13
• It reports Reddit summary statistics, including means, medians, standard deviations, and the five largest and smallest values.Source: Table D.7 · p. 13
• It analyzes the highest-weighted phrases from a BERT logistic regression.Source: Table D.7 · p. 14
• It analyzes the highest-weighted phrases from a GPT2 logistic regression.Source: Table D.7 · p. 15

• The survey instructions ask respondents to describe their personality in 75 to 150 words.Source: Appendix E · p. 15
• The prompt directs respondents to address five personality traits without naming the traits themselves.Source: Appendix E · p. 15
• The appendix includes a table of data-collection survey instructions.Source: Table E.7 · p. 15