Towards a Psychological Generalist AI: A Survey of Current Applications of Large Language Models and Future Prospects

1. INTRODUCTION

• Mental-health conditions are prevalent, but many health and social systems overlook them, leaving people to face challenges without requisite care and support.The introduction also notes that prompt intervention can save lives in critical situations such as potential suicides.Source: §1 · p. 1
• Expert-annotated deep learning can perform well, but its reliance on extensive annotations creates costs that can impede institutions and researchers.Source: §1 · p. 1
• LLMs’ prompt-based interaction can address multiple tasks without extensive expert-labeled data, and their conversational ability positions them as potential listeners and virtual companions.Source: §1 · p. 2
• The paper reviews LLM applications across domains, populations, and psychological constructs, and outlines future generalist AI prospects and the paper’s section structure.Source: §1 · p. 2

2. THE POTENTIAL OF GENERALIST AI MODELS IN PSYCHOLOGY

• GPAI is envisioned to handle broader psychological tasks across scenarios and demographic groups than current AI models, with minimal or no labeling in some tasks.Source: §2 · p. 2
• Ethical monitoring could support risk detection on online platforms and in psychiatric rehabilitation centers, with patient privacy remaining a priority.The section describes people expressing suicidal intentions online and potential anomalies or emergencies in rehabilitation settings.Source: §2 · pp. 2–3
• Multimodal GPAI is envisioned to examine images and text for social-media risks and process patient-clinician dialogue to inform tailored assessments or treatment.Source: §2 · p. 3
• With psychological and medical knowledge, GPAI could model illness progression and estimate future suicide rates from hotline interactions to aid expert interventions.Source: §2 · pp. 3–4

3. MODEL EVALUATION

• Task-specific evaluations include emotion recognition or awareness and mental-health classification, while exploratory assessments examine psychological reasoning and abilities such as emotional intelligence.The section also describes benchmarks for belief attribution, social-bot tasks, planning, explanation generation, and opinion-network interventions.Source: §3.1 · p. 4, §3.2 · p. 5

3.1 Task-specific evaluations in psychology

• GPT-3.5 reached a LEAS Z score of 4.26 in one evaluation, while reported classification F1 scores were 0.73 for stress, 0.86 for depression, and 0.37 for suicidality.The review also reports emotion-recognition instability across labels and datasets, improvement after fine-tuning, and differences between GPT-3.5 and GPT-4 in suicide-risk assessment.Source: §3.1 · p. 4

3.2 LLM performance in psychological contexts

• Exploratory evaluations cover intuitive-psychology reasoning, social-bot behavior, planning, explanation generation, emotional intelligence, and opinion-network interventions, though quantifiable criteria are often lacking.Source: §3.2 · p. 5

4.1.1 Sentiment detection and analysis

• A social-media analysis system reported best performance across full-data, few-shot, zero-shot, IID, and OOD settings, while LLM performance declined on more complex cognitive-distortion tasks.The review also describes GPT-3.5 as satisfactory for detecting constructs such as sentiment, discrete emotions, and offensiveness.Source: §4.1.1 · p. 5

4.1.2 Intelligent Q&A chatbots

• Mental-health Q&A chatbots can offer on-demand, non-judgmental support and patient-specific education, but face harmful-content filtering, communication consistency, information retention, overdependence, and stigma-related concerns.The reviewed work includes Replika user comments and systems for patient-care interactions and discharge-instruction questions.Source: §4.1.2 · pp. 5–6

4.1.3 Online psychological support tools

• Online psychological-support studies report higher preference and empathy ratings for chatbot responses than physicians’ responses, and evaluate LLM tools for counseling, recommendations, and coherent answers.One system evaluation involved ten professional counselors; other work describes 24/7 counseling and question-answering assessed with perplexity and human review.Source: §4.1.3 · p. 6

4.2.1 Mental interview, psychological counselor assistance

• LLMs are proposed for adjunct psychotherapy [73], psychiatrist–patient dialogue simulation [68], and online recommendations supporting non-professional counselors [32].Source: §4.2.1 · p. 6

4.2.2 Wearable devices and behavior monitoring

• Liu et al.74 found LLMs can analyze physiological and behavioral time series with minimal tuning, including cardiac signals, activity, metabolism, and stress.Source: §4.2.2 · p. 7

4.3.1 Focused on children and adolescents

• Among 23 adolescents aged 13–18, a depression and behavioral-activation chatbot drew moderate satisfaction, with over half giving positive feedback [76]; safety and professional guidance remain concerns [77,78].Source: §4.3.1 · p. 7

4.3.2 Educational settings and students

• GPT-3.5 chatbots may educate students about sexual violence [79]; among over 1,000 Replika users, students reported high social support despite greater loneliness than typical students [80].The passage says Replika could play a role in reducing loneliness and suicidal ideation.Source: §4.3.2 · p. 7

4.4.1 Simulating social networks

• LLM social-network studies report self-organized homophily [59] and promising simulation accuracy [81,82], but GPT-3 falls below humans on SocialIQa and ToMi [83].Chirper content was generally more generic than human content [59].Source: §4.4.1 · pp. 7–8

4.4.2 Modeling individual agents

• Turing experiments, human-response comparisons, and clinician or patient diagnostic conversations are used to assess population simulations; Chen et al.68 and Aher et al.84 report potential for accurate simulation.Source: §4.4.2 · p. 8

4.4.3 Simulating environments and contexts

• LLM agents have been used to simulate human interactions in communities and classrooms, while SocialAI school supports customizable socio-cognitive experiments, including role reversal [86–88].The reported findings point to a promising direction in guiding agents to attend to and adopt a partner’s role [88].Source: §4.4.3 · p. 8

4.4.4 Applications and innovations in psychological theory

• LLMs have modeled personality-linked responses [90] and human decision behavior, sometimes outperforming traditional cognitive models [91]; theory-of-mind work reports both successes and failures [58].Ullman et al.58 emphasize exceptional failure cases over average success rates when evaluating models.Source: §4.4.4 · p. 8

• GPT-4-generated semantic judgments positively correlated with human judgments and sometimes exceeded human inter-annotator agreement [92].Source: §4.5 · p. 9
• LLM-generated English psycho-linguistic judgments were explored as a way to create extensive datasets, addressing the time and economic constraints of collecting them [92].The datasets include lexical judgments such as concreteness and age of acquisition.Source: §4.5 · p. 9
• The study also examined systematic differences between psycho-linguistic norms generated by LLMs and humans [92].Source: §4.5 · p. 9

5.1 Potential application

• Reviews describe LLM applications in medical question answering, assessments, clinical conversations, and research, while noting task performance can remain below expert models [19,18,96,97].The literature also discusses potential roles in psychiatric measurement, diagnosis, personalized treatment, counseling, and treatment planning [18,96,67].Source: §5.1 · p. 9

5.2 Bias and fairness issues

• The literature identifies risks of gender, race, religion, and political bias, including possible underdiagnosis in marginalized communities; training-data perspectives and fairness measures are also discussed [15,18,19,21,67,93–99].Suggested responses include ethical nudges, ongoing audits, inclusive discussion, and scrutiny of who fine-tunes models and under what standards [18,21,99].Source: §5.2 · p. 9

5.3 Moderation approach

• Proposed moderation measures include antidiscrimination components, prompt engineering, instruction fine-tuning, expert oversight, adversarial testing, and RLHF; reward loops and system gaming may mislead training [93,18,19,15,97].The literature also notes that adversarial prompts can exploit model vulnerabilities and produce toxic outputs [19].Source: §5.3 · pp. 9–10

5.4 Privacy Issues

• Privacy measures discussed include de-identification and limiting collected information, but restricting identifiable patient data in prompts could hamper deployment opportunities [15,18,19,21,67,93,99].The literature also notes risks of sensitive-data exposure through training datasets and prompt attacks, and uncertainties about federated learning or synthetic data for LLMs [18,21,95].Source: §5.4 · p. 10

5.5 Monitoring and evaluation

• The literature calls for ongoing evaluation of accuracy, fairness, transparency, and ethics, while noting difficult clinical endpoints and benchmarks and the need for randomized controlled trials [19,21,49,67,93–99].Manual evaluation may be time-consuming and unscalable; model-based evaluation and unified systems are proposed, while current prompt- and dataset-based evaluations may be unreliable [15,49,94,97].Source: §5.5 · p. 10

6.1 Generalist AI in psychotherapy process

• AI could support psychological assessment, forecast risks from historical, genetic, and behavioral data, and provide rapid preliminary diagnoses to guide professional intervention.The section also anticipates that large-scale training may help identify medical conditions and reduce misdiagnosis and missed diagnosis.Source: §6.1 · p. 11

Treatment (cognitive behavioral therapy, communication dialogue, companionship)

• Proposed treatment uses include personalized CBT, round-the-clock dialogue, companionship, real-time behavior alerts, and monitoring treatment adherence and relapse.Monitoring could provide caregivers or medical professionals with timely insights for support and care adjustments.Source: Treatment (cognitive behavioral… · p. 12

6.2 Generalist AI across varied psychological scenarios

• Family and school applications range from home therapy, conflict and caregiver support to student distress monitoring, preliminary counseling, and mental-health education.The proposed school counseling addresses academic stress, peer issues, and personal challenges.Source: §6.2 · p. 13
• Across online and community settings, AI could check for distress, design targeted awareness campaigns, identify early signs, and connect people through support groups.Hospital uses include urgency-ranked intake and treatment recommendations based on medical history and psychological assessments.Source: §6.2 · p. 13

6.3 Tailoring psychological Generalist AI across age groups

• Age-tailored proposals include detecting children’s concerns through behavioral patterns, adolescent emotional-support platforms, and adult work-life-balance tools.The section also proposes interactive learning and parental guidance for children, and online therapy and peer-interaction analysis for adolescents.Source: §6.3 · p. 14

Elderly

• For older adults, proposed tools include cognitive exercises and memory aids, companionship, and monitoring that considers psychological and physical health.Source: Elderly · p. 14

6.4 Generalist AI in stages of psychological disease

• Across disease stages, generalist AI is proposed for prevention and early warning, crisis response, post-crisis recovery monitoring, and relapse detection.Examples include alerts for severe distress, daily recovery check-ins, and detecting behavioral patterns resembling prior distress episodes.Source: §6.4 · pp. 14–16

7.1 Leveraging open source pre-trained models

• Fine-tuning open-source pretrained models offers resource-constrained institutions an alternative to building expansive AI models from scratch.Training from scratch demands substantial time, computing power, and finances, which can be untenable for hospitals and smaller enterprises.Source: §7.1 · p. 16

7.2 The imperative of multi-domain evaluation

• Generalist AI efficacy must be verified across diverse domains and scenarios to establish its capabilities, boundaries, and real-world relevance.The section presents multi-domain and multi-scenario verification as necessary for pursuing broadly applicable AI.Source: §7.2 · p. 16

7.3 Navigating the ethical maze

• Fair service requires continuous monitoring, adjustment, and evaluation for fairness, transparency, bias, and cultural or regional nuance.The authors call for vigilance against inherent biases and transparent methodologies.Source: §7.3 · p. 16

7.4 Safeguarding permissions and privacy

• Practical AI systems must protect personal information and prioritize privacy, consent, and purposeful data use.The section identifies medical records and private social-media chats as examples of sensitive information.Source: §7.4 · p. 16

7. PERSPECTIVES IN PRACTICAL APPLICATION

• Practical adoption must address technical and ethical challenges while drawing on the AI community’s collective knowledge.The authors describe these steps as moving toward AI as a trusted companion in practical settings.Source: §7.4 · p. 16

8. CONCLUSION

• The paper reviews model evaluation and applications of foundation models, particularly LLMs, in psychology.It categorizes relevant literature by type and application direction and gathers viewpoints from related literature.Source: §8 · p. 17
• Future directions focus on psychotherapy processes, application scenarios, age groups, and disease processes.These are the four aspects identified for advancing generalist AI in psychology.Source: §8 · p. 17
• The paper discusses practical concerns including security, openness and transparency, multi-domain evaluation, ethics, and privacy to establish clear system boundaries.Its analysis aims to chart potential pathways toward psychological generalist AI.Source: §8 · p. 17

9. ABBREVIATION

• AI means Artificial Intelligence; NLP means natural language processing; and LLM means Large language model.Source: §9 · p. 17
• GPT means Generative Pretrained Transformer, and GPAI means Generalist Psychology Artificial Intelligence.Source: §9 · p. 17
• ToM means Theory of Mind; CBT means Cognitive Behavioral Therapy; and LEAS means Level of Emotional Awareness Scale.Source: §9 · p. 17
• WMGPT means Well-Mind ChatGPT; APA means American Psychiatric Association; CLEVA means Chinese Language Models EVAluation Platform; and GMAI means General Medical Artificial Intelligence.Source: §9 · p. 17

11. AUTHORS CONTRIBUTIONS

• Guanghui Fu, Bing Xiang Yang, and Jianqiang Li conceptualized the paper’s idea and structure.Source: §11 · p. 17
• Tianyu He, Fan Wang, Guanghui Fu, Yijing Yu, Changwei Song, and Hongzhi Qi prepared materials and collected literature; Tianyu He and Guanghui Fu wrote the initial draft.Source: §11 · p. 17
• Bing Xiang Yang, Jianqiang Li, Qing Zhao, Dan Luo, and Huijing Zou reviewed and edited the article, and all authors provided feedback and approved the final draft.Source: §11 · p. 17