Designing Personality-Adaptive Conversational Agents for Mental Health Care

Research question

How should personality-adaptive conversational agents be designed to improve interaction with users in mental health care?

Method

Iterative design science project with two spaces. In the problem space, the authors combine the five-factor model, the Computers Are Social Actors paradigm, domain problems and the literature, and an open qualitative questionnaire. Sixty participants viewed an explanation and a video of a simulated dialogue with Botsociety and produced user stories; the authors coded independently and resolved disagreements by consensus. In the solution space, seven meta-requirements were transformed into design principles following the anatomy of Gregor et al. Six mental health professionals evaluated them through semi-structured interviews; the recordings were transcribed and coded with MaxQDA 2020. The feedback refined six final principles, which were then represented in a conceptual architecture diagram. There was no functional prototype, therapeutic intervention, or trial with patients.

Sample: Questionnaire: 60 people (32 men and 28 women), aged 23 to 71 years, mean of 36, recruited from the authors' private network and Mechanical Turk in December 2020; the response required approximately 25–40 minutes and some people had no experience with mental health therapy. Evaluation: six experts from the authors' personal network, aged 28 to 38 years and with 2–9 years of experience: four psychiatrists with psychotherapeutic training, one psychologist, and one social worker/therapist. The interviews were conducted between March and April 2021 and lasted 50–80 minutes. There was no clinical sample exposed to a functional PACA.

Findings

• The process converts domain problems, problems described by the literature, and questionnaire responses into 28 user stories, seven meta-requirements, and six final design principles.
• The user stories are grouped into support, safety, and behavior; they include continuous availability, therapeutic techniques, human supervision, detection of warning signs, data protection, and heterogeneous preferences for style and personality.
• DP1, proactive support, recommends 24/7 availability and periodic contacts at a frequency mutually agreed upon with the user.
• DP2, competence, requires specific domain knowledge and therapeutic techniques so that the user perceives understanding and capability.
• DP3, transparency, requires communicating safety and privacy issues to foster trust when sharing sensitive data.
• DP4, social role, allows choosing between roles such as friend or therapist and changing role when this favors therapeutic progress.
• DP5, anthropomorphism, lets the user choose between chatbot, voice assistant, or embodied agent and, thereby, the degree of human appearance.
• DP6, personality adaptation, proposes linguistic signals linked to personality dimensions to adapt the agent to the user's preferred personality and increase interaction quality.
• The first three principles are presented as general prerequisites of a mental health agent; role, anthropomorphism, and personality provide personalization, and the authors consider DP6 the necessary feature to speak of a PACA.
• The experts judged personality adaptation as the most important principle, but asked that the agent be able to change its personality, bring the user out of their reserve, and not become someone who always pleases.
• For people with schizophrenia or paranoia, the experts warned that unsolicited proactive contact can feel intrusive; therefore, the periodicity should be agreed upon with the user or therapist.
• Four experts expressed doubts that a PACA can recreate the richness of a conversation with a human therapist.
• Transparency about privacy and security was considered fundamental to generate trust with mental health data.
• The experts observed that a friend role can avoid difficult questions and that certain roles or genders can trigger fear or aggressiveness; the appropriate role depends on the person and situation.
• Anthropomorphization can improve the interaction of some users, but also increase dependence or rejection, so the article proposes that the user control the format.
• The panel insisted on involving psychotherapy professionals in the design and on maintaining human supervision for severe cases; the PACA is conceived as a complementary tool.
• The experts themselves indicated that they could not fully judge the principles without experiencing a real system.
• The expository instantiation separates conversational interface, bot service, application logic, database, personality inference, and professional interface; it is a transferable representation, not an implementation.
• The diagram recommends text, voice, and possible virtual body, in addition to a professional interface for supervision, and mentions encryption and multifactor authentication for sensitive data.
• The claim that materializing the principles could improve interaction is a design expectation derived from the qualitative process, not a measured result in users of a PACA.

Limitations

• The study does not build the described PACA: the only materialization is a conversation mock-up and an expository instantiation in the form of a diagram.
• No Big Five inventory is administered, nor is the personality of participants, experts, or agents measured.
• No personality inference system from language is implemented or validated; LIWC, embeddings, and commercial services are only examples of architecture.
• It does not specify how to translate a Big Five estimate into linguistic signals, therapeutic decisions, or reproducible role changes.
• It does not define minimum accuracy, uncertainty, amount of text needed, user correction, voluntary exclusion, temporal drift, or management of contradictory profiles.
• There is no evaluation with patients using a PACA, measures of symptoms, interaction quality, therapeutic alliance, adherence, harm, crisis, or clinical outcomes.
• There is no comparative condition, randomization, longitudinal follow-up, quantitative analysis, or effect estimation.
• The 60 participants evaluated a hypothetical idea after an extensive explanation and a video designed by the authors; that preparation can induce expectations and does not equate to user experience.
• Recruitment through private network and Mechanical Turk is of convenience; the text does not characterize country, culture, education, diagnosis, clinical history, or other representativeness factors.
• Some people had no experience with therapy, a choice coherent with the intended non-clinical population but that limits the user stories about real care.
• The article does not report questionnaire compensation, dropout rate, exclusions, theoretical saturation, or distribution between private network and Mechanical Turk.
• Although the authors say they coded independently and resolved discrepancies by consensus, they do not report number of coders, codebook, complete examples, intercoder reliability, or external audit.
• The 28 user stories are aggregations of mentions; the article does not offer complete frequencies or raw data to assess prevalence or alternative interpretations.
• The panel contains only six experts recruited from the authors' personal network, all aged 28–38 years and with less than ten years of experience.
• The panel does not include patients, caregivers, privacy or cybersecurity specialists, ethics, health regulation, accessibility, suicide prevention, or AI evaluation.
• The experts evaluated descriptions of principles after an explanation of the concept; they did not observe functioning, errors, personality inferences, or real clinical conversations.
• The study does not report ethics committee approval, informed consent, or rationale for exemption for questionnaires, video, and interviews; the declarations section only reports funding and absence of conflict.
• No data repository, transcripts, complete questionnaire, interview guide, analysis code, or executable artifact is offered for independent reproduction.
• DP1 does not develop limits for proactive contacts, schedules, pauses, revocable consent, or specific protocols for paranoia, mania, violence, or suicidal risk.
• DP2 requires competence, but does not define clinical sources, content validation, updating, traceability, professional responsibility, or management of unsafe responses.
• DP3 is limited to transparency, encryption, and multifactor authentication; it does not specify minimization, purpose, retention, deletion, access, audit, breaches, third parties, data residency, or granular consent.
• DP4 includes the therapist role without developing safeguards against role deception, overconfidence, attribution of clinical authority, or conflict between preference and benefit.
• DP5 recognizes dependence and uncanny valley, but does not propose measurable limits for anthropomorphism, emotional attachment, or substitution of human relationships.
• DP6 can reinforce biases or stereotypes if it associates traits with language in a simplistic way; cultural equity, multilingualism, disability, age, or demographic differences are not evaluated.
• The architecture suggests that the database learns from interactions of all users, but does not specify isolation, training consent, anonymization, contamination between users, or right of opposition.
• The professional interface is drawn without defining responsibilities, supervision burden, alerts, escalation, availability, or what happens when no professional is connected.
• The six principles mix general requirements of any mental health agent with personality adaptation; only DP6 directly articulates that adaptation.
• The named technologies are examples from 2022 and do not constitute a selected, versioned, or evaluated stack; the design depends on third-party services without analyzing availability or changes.
• The work maintains that the advantages can outweigh the risks if the system is designed properly, but does not provide a safety validation that allows making that balance empirically.
• The qualitative evaluation supports perceived relevance of principles, not the causal claim that materializing them will increase interaction or improve care.

What the study does not establish

• It does not demonstrate that a functional PACA with the described capabilities exists.
• It does not demonstrate that an agent can validly or reliably infer a patient's personality from their messages.
• It does not validate correspondence between linguistic signals, Big Five dimensions, and individual therapeutic needs.
• It does not prove that adapting personality increases interaction quality, trust, adherence, therapeutic alliance, or satisfaction.
• It does not prove reduction of anxiety, depression, loneliness, crisis, waiting times, or any other mental health outcome.
• It does not demonstrate clinical safety, ability to detect warning signs, or adequate response to suicide, psychosis, mania, abuse, or emergency.
• It does not establish that PACAs can replace psychologists, psychiatrists, psychotherapists, or human relationships; the text presents them as complementary support.
• It does not demonstrate that 24/7 availability or proactive contacts are beneficial for all people; the experts identify cases in which they could be harmful.
• It does not demonstrate that adopting the user's preferred role is therapeutically correct or that a therapist role is safe.
• It does not prove that greater anthropomorphism is better or that dependence, isolation, or emotional attachment can be avoided.
• It does not demonstrate that encryption and multifactor authentication suffice for privacy, regulatory compliance, or responsible governance of health data.
• It does not validate any of the platforms, libraries, or techniques cited as a clinical component of a PACA.
• It offers no evidence of generalization to cultures, languages, ages, diagnoses, health systems, or populations other than the convenience sample.
• It does not quantify how much each principle contributes or whether the six together are necessary or sufficient.
• It does not causally prove that personality is the reason for the limitations of contemporary conversational agents.
• It does not establish a regulatory, ethical, computer security, or medical product evaluation.
• It does not convert the favorable perception of six experts about principles into evidence of efficacy of a system.
• It does not allow interpreting the phrase in the abstract about a promising source of support as a result of a clinical trial or of real-world use.

Traceability

Scope: Full text

Version: Information Systems Frontiers 24:923–943; published online 2 March 2022; PMC8889396.1

Consulted source: https://pmc-oa-opendata.s3.amazonaws.com/PMC8889396.1/PMC8889396.1.pdf

Review: Codex full-text, bilingual-fidelity, visual, metadata, design-science, qualitative-method, clinical-claim, personality-adaptivity, architecture, privacy, safety, ethics and evidence-level audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

• Five Factor Model / Big Five (kernel theory, not administered as a measure)
• Computers Are Social Actors paradigm (kernel theory)
• Botsociety mock-up of Raffi and Jules (stimulus only)
• Technology-independent expository PACA architecture (conceptual blueprint only)
• LIWC, word2vec, GloVe and IBM Watson Personality Insights (illustrative personality-mining options, not implemented or evaluated)
• Google Dialogflow, IBM Watson Assistant, Amazon Lex and Wit.ai (illustrative bot services, not implemented or evaluated)

Instruments and metrics

• Open qualitative questionnaire on support, safety/privacy, behaviour and communication preferences
• Predefined Botsociety conversation video and mock-up featuring Raffi and Jules
• Qualitative content analysis of user stories
• Semi-structured expert interview guide
• MaxQDA 2020 for coding expert interview transcripts
• Gregor et al. design-principle anatomy
• Expository instantiation and mapping diagrams

Data used

• More than 6,865 words of open-questionnaire responses from 60 participants
• Twenty-eight aggregated user stories across support, safety and behaviour
• Six audio-recorded and transcribed expert interviews
• Application-domain issues, literature issues, seven meta-requirements and six design principles
• No public raw-data or code repository reported in the article

Evidence and location

• Bibliographic identity, authorship, date, DOI, and complete abstract: Full text p. 1, title page and abstract; Information Systems Frontiers 24:923–943, DOI 10.1007/s10796-022-10254-9
• Research question and foundational theories: Full text pp. 2–3, Introduction and section 2
• Design science process and study steps: Full text pp. 5–6, section 4 and Figure 1
• Questionnaire, Botsociety, recruitment, n=60 and 6,865 words: Full text pp. 7–8, Step 1 qualitative study, Table 1 and Figure 2
• Independent coding and consensus: Full text p. 7, qualitative content analysis procedure
• Expert selection, composition, dates, and duration: Full text pp. 8–9, expert evaluation and Table 2
• Semi-structured guide, transcription, and MaxQDA 2020: Full text p. 8, expert interview procedure
• User stories, problems, and seven meta-requirements: Full text pp. 9–11, section 5.1 and Tables 3–4
• Evaluation and review of proactive support, competence, and transparency: Full text pp. 11–12, evaluation of DP1–DP3
• Evaluation and review of social role, anthropomorphism, and personality adaptation: Full text pp. 12–14, evaluation of DP4–DP6
• Final drafting of the six principles: Full text p. 14, Table 5
• Expository architecture and technology-independent character: Full text pp. 14–16, section 5.2 and Figure 3
• Bot services and inference techniques cited as possibilities: Full text pp. 15–16, expository instantiation
• Encryption, multifactor authentication, and professional interface: Full text pp. 15–16, Figure 3 and database description
• Classification of principles and complementary support with human supervision: Full text p. 16, section 6.1
• Recognized limitations of participants and expert panel: Full text p. 17, section 6.2
• Risks of isolation, attachment, third parties, and provider dependence: Full text p. 17, Limitations and Ethical Considerations
• Real scope: refined principles and diagram, without a tested system: Full text p. 17, Conclusion
• Funding and conflict of interest; absence of specific ethics declaration: Full text p. 18, Funding and Declarations
• Visual inspection: All 21 PDF pages rendered and visually inspected, including Figures 1–4 and Tables 1–5; checked 15 Jul 2026