PaperlayerView exact revision on arXiv

Source-linked AI summary

The Artificial Personality Model: Theory, Implementation, Measurement and Experimental Validation in Conversational Agents

Alexander Dregger, Barbarella Petz, Erik Kuhn, Pia Volk, Andreas Oberweis, Stefan Hillmann, Jannik Strötgen

arXiv:2610.38608v1cs.HCcs.AI

TL;DR

Artificial personality matters in human–machine interaction, but coherent frameworks and validated measures remain limited. The paper refines and evaluates the hybrid APM using a new questionnaire and varied implementations; findings indicate partial transfer of human dimensions and effects on user experience outcomes.

  • Problem

    Artificial personality is relevant to user experience, but existing models and measures leave its definition, implementation, and measurement unresolved.

    Source: §1 · pp. 1–2

  • Method

    The paper refines the hybrid APM, develops a questionnaire, and examines language-based personality implementations in mock-ups and LLM systems.

    Source: §1 · p. 4, §2.4 · p. 7

  • Results

    Some human dimensions, including agreeableness and conscientiousness, appear in artificial personality with adapted facets and measurement, while artificiality is machine-specific.

    Source: §6 · p. 43

  • Takeaways & Limitations

    Artificial personality affects user satisfaction, continuous use intention, and social presence in the studied experiments.

    Source: §6 · p. 43

    Only text-based chatbots and primarily task-oriented contexts were studied, leaving generalisation to other agents and applications unclear.

    Source: §5.2 · p. 42

Abstract

from arXiv · show

Artificial personality shapes user experience (UX) in human-machine interaction, yet coherent theoretical frameworks and validated measurement instruments remain limited. This paper refines and empirically evaluates the Artificial Personality Model (APM), a hybrid framework integrating human-centred and machine-specific personality dimensions. Across three experiments (N = 395), the model was operationalised and evaluated using the Artificial Personality Questionnaire (APQ). Factor analyses revealed that some human personality dimensions (e.g., agreeableness, conscientiousness) are partially transferable, whereas others (e.g., extraversion) did not emerge as distinct dimensions. Additional machine-specific dimensions were identified. The APQ provided encouraging initial psychometric evidence, while the human-centred BFI-2 showed poor reliability and an uninterpretable factor structure in Experiment 3. Experimental manipulations demonstrated significant effects of artificial personality on user satisfaction, continuous use intention, and social presence. The findings contribute an empirically refined framework and a novel instrument for measuring perceived artificial personality, offering practical guidance for designing conversational agents.

1. Introduction

Artificial personality is relevant to human–machine interaction, but existing approaches leave open how it should be defined, implemented, and measured. The paper develops the hybrid Artificial Personality Model and a questionnaire, then evaluates them through implementation work and experiments.

1. Introduction

  • Artificial personality matters because people attribute human-like qualities to machines, and it can affect user experience and decisions.The paper identifies trust, acceptance, and purchase decisions as outcomes linked to artificial personality.Source: §1 · p. 1
  • Human-centred models can be too abstract for agent design, while machine-centred models overlap with human traits and complicate transfer of findings.Machine-centred approaches also offer limited implementation guidance, with prompting recommendations noted as an exception.Source: §1 · p. 2
  • Evidence gaps concern both whether human-derived items describe machines and whether the underlying personality dimensions apply to artificial systems.Adapted Big Five inventories achieved high content validity for LLMs, yet the expected five-factor structure was not recovered (Zierahn et al., 2026).Source: §1 · pp. 2–3
  • The Artificial Personality Model combines human-centred and machine-centred perspectives rather than assuming a human model transfers intact.It includes seven dimensions, while only three had defined facets at the outset.Source: §1 · pp. 3–4
  • The study aims to expand the APM’s facets, develop a questionnaire, test mock-up and LLM-based implementations, and evaluate the model across three experiments.The experiments address different domains and dimensions, with control groups intended to clarify manipulation effects on user experience.Source: §1 · p. 4

2. Background and Related Work

The background frames artificial personality as perceived individuality expressed through machine behaviour, then reviews its theoretical basis, implementation options, and measurement challenges. It motivates a human-rated instrument and examines language-based implementations using mock-ups and LLM prompting.

2.1. Anthropomorphism and Computers are Social Actors paradigm (CASA)

  • Anthropomorphism and CASA explain why people can perceive human-like qualities and respond socially to machines, including through machine language.Anthropomorphism can serve sociality and effectance motivations; responses also vary with individual and design factors.Source: §2.1 · pp. 4–5

2.2. Artificial Personality and the Artificial Personality Model

  • Artificial personality is designed machine behaviour that creates an impression of individuality; the APM combines human and machine dimensions.The model has seven dimensions, but definitions and facets were incomplete and it had not yet been empirically tested.Source: §2.2 · pp. 5–6

2.3. Conversational Agents

  • The study focuses on text-based chatbots with virtual embodiment, using one mock-up and two LLM-based systems in distinct domains.Source: §2.3 · p. 6

2.4. Implementation of Artificial Personality into CAs using social cues, mock-ups and LLMs

  • The study expresses personality through chatbot language, using verbal, visual, and invisible cues while setting aside other modalities to limit complexity.Mock-ups standardize interaction but constrain user input; LLM prompting is used as a resource-efficient implementation approach.Source: §2.4 · pp. 7–8

Zero-shot Personality Prompting

  • Zero-shot personality prompting specifies a trait through instructions, descriptive keywords, and generated personality statements alongside task context.Source: Zero-shot Personality Prompting · p. 8

Personality Infusion

  • Personality infusion separates personality specification from content generation and uses prompts for context, task, output format, and dialogue data.The method was evaluated across multiple LLMs and had been validated with user testing.Source: Personality Infusion · pp. 8–9

2.5. Measurement of Artificial Personality

  • Machine self-ratings and adjective-based human ratings leave unresolved whether measures capture user-perceived behaviour and theoretically distinct dimensions.The paper therefore develops a questionnaire to measure perceived artificial personality.Source: §2.5 · pp. 9–10

3. Methods

The methods operationalise the seven-dimension APM through literature-based facets and an expert-refined questionnaire, then test the model across three contexts and implementation approaches. The framework distinguishes human-derived dimensions from machine-specific traits and translates them into observable conversational-agent behaviours.

3.1. Development of a Questionnaire; 3.1.1. Extraversion

  • The APM combines three Big Five, two supernumerary, and two machine-centred dimensions; its ten proposed extraversion facets are narrowed to sociability, assertiveness, and positive energy.The extraversion facets reflect social openness, interaction dominance, and perceived energy.Source: §3.1 · p. 11, §3.1.1 · p. 11

3.1.2. Agreeableness

  • Agreeableness is represented by five facets, including compassion, respect, trust, social sensitivity, and flexibility for responding appropriately to social dynamics and interaction changes.The review also identifies trust as non-maliciousness and compassion as concern for user well-being.Source: §3.1.2 · pp. 11–12

3.1.3. Conscientiousness

  • Conscientiousness is narrowed to responsibility, productiveness, and action orientation, covering reliable fulfilment of obligations, persistence, and efficient implementation of user decisions.The selection reduces overlapping candidate facets to support measurement efficiency.Source: §3.1.3 · p. 12

3.1.4. Humour

  • The APM defines machine humour through sense of humour and three observable styles—affiliative, aggressive, and self-defeating—excluding self-enhancing humour and humour competence.Self-enhancing humour is excluded as intrapersonal; competence is treated as aptitude rather than a stable behavioural tendency.Source: §3.1.4 · pp. 12–13

3.1.5. Honesty

  • For machine honesty, the paper develops facets for openness of intentions, transparency, compliance, and manipulativeness because existing literature offers few clearly defined facets for this context.The facets cover goals and functions, data processing, legal interaction limits, and attempts to influence users.Source: §3.1.5 · p. 13

3.1.6. Artificiality

  • Artificiality descriptors are clustered into facets covering machine-like behaviour, monotony, simple-mindedness, lack of empathy, and controlling behaviour.The clusters organise descriptors from prior machine-personality conceptualisations into analytically distinct facets.Source: §3.1.6 · pp. 13–14

3.1.7. Serviceability

  • Serviceability descriptors are grouped into informativeness, helpfulness, professionalism, and interaction quality, spanning reliable communication, goal assistance, competent demeanour, and seamless dialogue.The source descriptors were not previously analytically clustered into facets.Source: §3.1.7 · p. 14

3.2. The Artificial Personality Questionnaire (First Version)

  • The behaviour-based APQ began with 127 items, underwent review by four experts, and was reduced to 90 items for the first two experiments.Most items were newly developed because existing questionnaire items often did not transfer to machine contexts.Source: §3.2 · p. 15

4. Results of the three Experiments

The experiments examine APM structure, questionnaire properties, and user experience across conversational-agent contexts. In Experiment 1, an initially theory-aligned seven-factor extraction was less interpretable than a six-factor solution, with several dimensions combining or failing to emerge distinctly.

4.1.2. Method and Sample

  • Experiment 1 used five conditions, assigned via a generated check digit; of 217 participants, 45 were excluded after applying minimums of five interaction turns and 12 minutes, leaving 172.The conditions were aggressive humour, affiliative humour, Self-Defeating humour, seriousness, and an LLM without personality prompting as control.Source: §4.1.2 · pp. 16–17

4.1.3. Results

  • The parallel analysis indicated seven factors, matching the APM’s theoretical count, but the subsequent factor analysis judged the six-factor structure more interpretable.The analysis used 1,000 random datasets and a 95th-percentile criterion before the seven-factor solution was examined.Source: §4.1.3 · p. 18
  • 44.11% of variance was explained by the interpretable six-factor solution, compared with 45.84% for seven factors whose seventh factor lacked consistent content structure.The six-factor solution had loadings from .40 to .81, with a .30 cross-loading found for one item.Source: §4.1.3 · pp. 18–19
  • The six factors comprised agreeableness, conscientiousness, service orientation, humour, artificiality, and a new Machiavellianism dimension; extraversion and honesty did not emerge distinctly.Agreeableness included positive energy and professionalism; Machiavellianism combined manipulation and control facets.Source: §4.1.3 · p. 19
  • Expert reviews provided initial content and face validity evidence for the APQ, while interpretation remains limited because no standardisation sample is yet available for norming.Administration and scoring were described as standardised and Likert-based, respectively.Source: §4.1.3 · p. 19
  • All APQ dimensions had reliability of at least .60, with agreeableness, conscientiousness, and humour at least .80; trust, responsibility, action orientation, and machine-likeness were poor but at least .50.These thresholds are reported as the questionnaire’s dimension and facet reliability findings.Source: §4.1.3 · p. 20

RQ1.4.: Can different humour styles be generated in machines?

Experiment 1 found that prompting could elicit distinct humour styles, with the intended style generally rated highest. Humour condition also affected user satisfaction and social presence, particularly for aggressive humour.

RQ1.4.: Can different humour styles be generated in machines?

  • Participants evaluated all four humour facets within their assigned between-subject condition; MANOVA tested condition differences and mixed-design ANOVA tested condition-by-style interactions.The five conditions were affiliative, aggressive, self-defeating, serious, and control.Source: RQ1.4.: Can different humour styles be… · p. 20

Between Subject Analysis

  • Aggressive humour scored higher than all other conditions, and self-defeating humour scored higher than all others; all humour conditions exceeded serious and control on sense of humour.For affiliative humour, self-defeating scored higher than aggressive; no differences among humour conditions emerged for sense of humour.Source: Between Subject Analysis · pp. 20–21
  • Intended humour styles were rated highest in their conditions (F(10.73, 448.13) = 17.94, p < .001, ηp² = .30), except affiliative humour was comparable in the self-defeating condition.Humour styles did not differ significantly in the serious or control conditions.Source: Between Subject Analysis · p. 21

RQ1.5.: How do different humour styles impact the UX of the CA?

  • Aggressive humour had lower social presence than affiliative humour and lower satisfaction than affiliative, control, and serious conditions; interactional enjoyment did not differ significantly.The condition effect was significant for social presence and satisfaction, but not interactional enjoyment.Source: RQ1.5.: How do different humour styles… · p. 22
  • Anthropomorphism tendency had no significant multivariate effect, but was positively associated with conscientiousness (B = 0.140) and humour (B = 0.107).The experimental condition still had a significant multivariate effect after controlling for anthropomorphism tendency.Source: RQ1.5.: How do different humour styles… · p. 22

4.2. Experiment 2: Service-Orientation

Experiment 2 tested service orientation in a package-return chatbot mock-up and examined its perceived personality and user experience. Low service orientation was distinguishable and associated with lower UX scores, but the experiment did not establish a stable APM factor structure.

4.2.1. Research questions

  • The experiment asked whether APM structure and service orientation could be detected in another domain, whether the APQ retained psychometric quality, and how service orientation related to UX.It also tested whether a mock-up elicited different service-orientation levels and whether technology commitment related to perceived artificial personality.Source: §4.2.1 · p. 22

4.2.2. Method and Sample

  • A package-return chatbot mock-up varied language, visual, and implicit behavioural cues to express low, medium, or high service orientation, with GPT-4o responses as control.The procedure also collected demographics and technology commitment before participants viewed the interaction.Source: §4.2.2 · p. 23
  • The 88-person final sample completed questionnaires after the video, including the APQ, UX measures, and continuous-use-intention items.Ninety-three completed the questionnaire; five were excluded after attention checks and a minimum-time threshold.Source: §4.2.2 · p. 24

4.2.3. Results

  • In testing whether the APM’s empirical structure could be identified in another domain, exploratory factor analysis of N = 88 yielded no stable factor structure, so the research question could not be answered empirically.The sample was below the literature-recommended minimum of N = 100 for factor analysis (Kline, 1994), although such thresholds are also discussed critically (Mundfrom et al., 2005).Source: §4.2.3 · p. 25
  • Although reliability was generally high, Experiment 2 provided no additional evidence for structural validity because its factor structure was unstable.Reliability was assessed descriptively against theoretically specified scales.Source: §4.2.3 · pp. 25–26
  • Low service orientation differed from all other groups across four facets; medium and high were distinguished only for interaction quality, while GPT-4o’s control resembled high service orientation.The high-service condition scored highest and the low-service condition lowest for interaction quality.Source: §4.2.3 · pp. 26–27, Table 11 · p. 26
  • Low service orientation scored lower on UEQ-S, continuous use intention, and Net Promoter Score than the other conditions.All three condition effects were significant; the low-condition comparisons had p < .001.Source: §4.2.3 · p. 27, Table 12 · p. 27
  • Technology commitment had no significant multivariate effect on the combined outcomes (Wilks’ Λ = .946, F(4, 77) = 1.10, p = .363, ηp² = .054).Source: §4.2.3 · p. 27

4.3. Experiment 3: Artificiality

Experiment 3 tested the APM and APQ in a phishing-email chatbot, finding an identifiable APQ structure but no meaningful BFI-2-S factor structure and poor reliability. The APQ dimensions showed preliminary distinctiveness and acceptable-or-better reliability, while the BFI-2-S was rated less suitable for measuring artificial personality.

4.3.2. Method and Sample

  • The 89-item APQ was tested in a phishing-email chatbot study; 135 of 149 participants who interacted and completed the questionnaire formed the final sample.Participants encountered control or high, medium, or low artificiality conditions and evaluated emails after three chatbot interactions.Source: §4.3 · p. 27, §4.3.2 · pp. 28, 31

4.3.3. Results

  • The seven-factor APQ analysis identified agreeableness, service orientation, humour, artificiality, Machiavellianism, honesty, and cognitive flexibility; conscientiousness did not emerge.Cognitive flexibility was not further analysed because it was considered ability-oriented rather than temperament-based.Source: §4.3.3 · p. 32
  • The highest APQ factor correlation was r = .35 between agreeableness and service orientation, providing preliminary evidence that the factors were distinct.Source: §4.3.3 · p. 32, Table 14 · p. 32
  • The BFI-2-S yielded no meaningful factor structure, and all dimensions except openness fell below Cronbach’s alpha = .50.Participants also rejected more BFI-2-S items than APQ items: M = 31.00 versus M = 20.29 rejections per item, a significant difference, t(134) = -4.49, p < .001, d = -0.39.Source: §4.3.3 · p. 34

Between Subject Analysis

In Experiment 3, artificiality condition affected perceived machine-likeness, but not the retained UX outcomes. The manipulation did not significantly affect monotony, and neither technology affinity nor anthropomorphism significantly influenced perceived artificiality.

Between Subject Analysis

  • 13.35, F(3, 131) = 13.35, p < .001, η²ₚ = .23, for machine-likeness; monotony was not significant, F(3, 131) = 2.21, p = .089, η²ₚ = .05.Low artificiality was lower than control on machine-likeness, but the difference was not significant (p = .06).Source: Between Subject Analysis · p. 34
  • No significant overall condition effect was found for retained UX outcomes: Wilks’ Λ = .97, F(6, 260) = 0.57, p = .753, ηp² = .01.The trust scale had poor reliability (.45) and was not further used; UEQ-S reliability was .82.Source: Between Subject Analysis · p. 35
  • Anthropomorphism and technology affinity did not significantly influence perceived artificiality, while anthropomorphism was associated with perceived Machiavellianism.The authors urge caution because anthropomorphism tendency was low and technology affinity was relatively high in this sample.Source: Between Subject Analysis · p. 35

4.4. Revision of the Model and Final Questionnaire

The APM revision drew most strongly on Experiment 1 and Experiment 3, whose exploratory factor analyses produced relatively stable but nonidentical structures. The paper presents core and research versions to reflect this evidence and its uncertainty.

4.4. Revision of the Model and Final Questionnaire

  • Exploratory factor analyses in Experiments 1 and 3 yielded relatively stable, but not identical, factor structures.Source: §4.4 · p. 35
  • The authors note that small, nonrepresentative samples could have contributed to model instability.Source: §4.4 · p. 36
  • The revised APM has a core model of dimensions and facets shared across both experiments and a research model including those identified in at least one.Source: §4.4 · p. 36

Core Model

The paper offers a compact core APM and a broader research model. The latter includes dimensions and facets found in at least one experiment, some of which may depend on context.

Core Model

  • The core model contains five dimensions and 14 facets measured with 41 items, supporting more efficient measurement but risking omission of relevant content.Source: Core Model · p. 36, Figure 15 · p. 36

Research Model

  • Some research-model dimensions and facets may be context-dependent.Source: Research Model · p. 36
  • The research model extends the core to seven factors and 24 facets, measured with 69 items, including dimensions and facets identified in at least one experiment.Source: Research Model · p. 36

5. Discussion

The discussion presents the APM as a hybrid framework for human-derived and machine-specific personality, alongside initial evidence about its dimensions, measurement, and UX effects. It also identifies limits in the evidence and priorities for further validation.

5.1. Theoretical Contributions and Practical Implications

  • The APM combines applicable human-centred dimensions with machine-specific ones, avoiding direct transfer assumptions and unexamined relabelling of machine behaviours as novel constructs.The hybrid strategy retains established personality theory where applicable and adds dimensions where human models provide insufficient coverage.Source: §5.1 · pp. 38–39
  • Agreeableness emerged across experiments with compassion, respect, and trust, while conscientiousness appeared in Experiments 1 and 2 and requires further investigation.Agreeableness also includes positive energy and social sensitivity, and may overlap with the related but narrower service orientation dimension.Source: §5.1 · p. 39
  • Extraversion did not emerge as an APM dimension; sociability and assertiveness were judged poorly suited to machines, while positive energy was retained within agreeableness.The authors connect this result to machines’ limited sociability and users’ preference against agents having their own agenda.Source: §5.1 · p. 39
  • The APM includes humour, artificiality, Machiavellianism, service orientation, and honesty, with several dimensions or facets still requiring refinement or confirmation.Aggressive humour tends to reduce satisfaction and social presence; Machiavellianism emerged exploratorily in two experiments, while honesty appeared only in Experiment 3.Source: §5.1 · pp. 39–40
  • The APQ measures users’ perceptions of personality from observed system behaviour, and its initial psychometric evidence is encouraging but does not establish external behavioural correspondence.The authors note that shorter versions are needed and that the proposed factor structure requires confirmation in larger independent samples.Source: §5.1 · pp. 40, 42
  • The paper recommends an unprompted LLM as a control baseline and reports personality infusion as more suitable than Jiang et al.’s method for implementing the APM’s multidimensional structure.The authors also call for further evaluation of few-shot prompting and replication of anthropomorphism effects in more representative samples.Source: §5.1 · p. 41
  • Aggressive humour reduced user satisfaction and social presence, while affiliative humour tended to increase both; lower service orientation reduced continuous use intention, satisfaction, and UEQ-S scores.Artificiality manipulation affected perceived artificiality but had no significant UX effects; the authors note measurement and sample issues as possible explanations.Source: §5.1 · p. 42

5.2. Limitations and Future Research

  • Evidence is limited to text-based chatbots, mainly task-oriented contexts, non-representative German-speaking samples, and initial APQ validation.The authors call for larger independent and cross-cultural samples, validation across technologies and contexts, and shorter APQ versions.Source: §5.2 · p. 42

6. Conclusion

The conclusion reports that the studies developed and tested the APM and APQ, supporting a hybrid account of artificial personality. It also finds that artificial personality affects several UX outcomes, while identifying the need for further validation.

6. Conclusion

  • The experiments support a hybrid approach: some human-derived dimensions transfer with adaptation, others are unsupported, and machine-related dimensions are also needed.The APQ showed promising initial psychometric properties across three studies, but its factor structure needs confirmation in larger independent samples.Source: §6 · p. 43
  • Agreeableness and conscientiousness appeared in artificial personality with adapted facets and measurement, while artificiality was identified as specific to artificial personality.Source: §6 · p. 43
  • Artificial personality affected user satisfaction, continuous use intention, and social presence; the authors call for further study of preferences and adaptation.Source: §6 · p. 43

Funding

The paper reports no specific grant funding and no competing interests.

Funding

  • The research received no specific grant, and the authors declared no competing interests.Source: Funding · p. 44

Data Availability Statement

The datasets are not publicly available because participant consent did not cover public dissemination and data protection and ethical restrictions apply.

Data Availability Statement

Use of AI in Research Methodology

The experiments used GPT-4o and GPT-4o-mini to implement and evaluate artificial personality in conversational agents. The authors report taking responsibility for research integrity and retaining implementation records.

Use of AI in Research Methodology

  • GPT-4o and GPT-4o-mini were used to implement and evaluate artificial personality in conversational agents.The Methods section describes the models, implementation procedures, and prompting approaches.Source: Use of AI in Research Methodology · p. 44
  • The authors verified the research's originality and accuracy, reviewed applicable terms of use, and accepted responsibility for the findings' integrity.This responsibility includes the accuracy of references.Source: Use of AI in Research Methodology · p. 44
  • Records of prompting procedures and experimental implementations were retained and can be provided to the journal upon request.Source: Use of AI in Research Methodology · p. 44

Related papers

Loading 2610.38608v1…