Psy-Insight: Transparent AI in Mental Health

• Psy-Insight is a framework integrating computational systems, models, and multi-modal datasets to offer transparent, context-aware psychological insights.
• It employs multi-task learning with chain-of-thought generation and knowledge-graph integration to deliver accurate, interpretable outcomes in therapy settings.
• The approach extends to AI self-modeling and multimodal reasoning, enhancing clinical assessment and system auditing through rigorous, explainable methods.

Psy-Insight

Psy-Insight refers to a class of computational systems, models, datasets, and methodologies designed to deliver interpretable, context-aware, and multi-modal psychological insight either about individual experience or mental states, the therapeutic interaction process, or the emergent properties of AI systems themselves. Psy-Insight research integrates advances in machine learning, computational psychiatry, human-computer interaction, and explainable artificial intelligence, with a core focus on transparent, multi-faceted reasoning that aligns with psychological science and practical mental health objectives.

1. Dataset and Annotation Paradigms

The most explicit instantiation is the "Psy-Insight: Explainable Multi-turn Bilingual Dataset for Mental Health Counseling," an open multi-task resource for LLM-based counseling. The dataset is constructed from 951 real face-to-face counseling sessions (520 English, 431 Chinese), spanning 5–60 turns per session and employing rigorous privacy filtering and annotator protocols (Chen et al., 5 Mar 2025).

Annotations in Psy-Insight adhere to multi-task schemas:

• Session-level: Psychotherapy school (12 classes), topic tags (30+).
• Turn-level: Client emotion (10+unknown), therapist strategy (9+unknown).
• Explanatory: For each turn, objective client observation (o_j) and subjective therapist reasoning (r_j), enabling chain-of-thought process modeling.
• Session Guidance: High-level instructions, background, and post-session summaries.

Table 1: Overview of Psy-Insight Annotation Schema

Level	Annotation types	Example Coverage
Session	Psychotherapy school, Topic	CBT, SFBT, Relationship, Depression
Turn	Emotion, Strategy, Explanations	Sadness, Question, Reflection, Reasoning
Session guide	Instructions, Context, Summary	Open-ended Qs, Build trust, Key insights

Inter-annotator agreement is high: κ=0.78 (emotion), κ=0.81 (strategy), κ=0.75 (school).

2. Model Architectures and Learning Strategies

Psy-Insight systems are grounded in multi-stage generation and multi-task learning, merging classification and reasoning objectives:

• Classifier heads: Emotion/strategy prediction via softmax, with explicit labels for both client and therapist turns.
• Chain-of-thought generation: Multi-stage targets for each turn incorporate the sequence $$\{\mathrm{Observation}\ o_j, \mathrm{Reasoning}\ r_j, \mathrm{Therapist}\ t_j\}$$, with the model sequentially generating each component to maximize interpretive transparency.
• Retrieval-Augmented Generation (RAG): Therapist replies are conditioned on both dialogue context and retrieved session guides, improving factual consistency and alignment with best practices.

Performance is quantitatively validated across various LLM baselines:

Model	BERTScore-P	BLEU-1	METEOR
Mistral-7B_SFT	dialog	0.903	0.270
Mistral-7B_SFT	reason+dlg	0.912	0.266
ChatGPT_SFT	reason+dlg	0.876	0.204

Adding step-by-step reasoning improves both objective metrics (BERTScore-P, BLEU) and expert-rated utility (61–67% win rate vs Esconv/SMILE), confirming that turn-level explanations are essential for realistic, empathetic, and logically grounded counseling (Chen et al., 5 Mar 2025).

Recommendation: Multi-task approach combining annotation and generation is necessary for both predictive accuracy and explainability.

3. Symptom- and Knowledge-Driven Methods

Psy-Insight integrates structured psychiatric knowledge into interpretability frameworks. "Symptom Identification for Interpretable Detection of Multiple Mental Disorders" introduces PsySym, a corpus and pipeline for symptom extraction mapped to DSM-based knowledge graphs (Zhang et al., 2022).

Key elements include:

• Explicit symptom KG: 38 DSM-grounded symptom classes, each mapped to 7 canonical disorders, yielding a bipartite graph G with 162 edges.
• Two-stage annotation: Sentence–symptom relevance followed by status (true/uncertain), enabling fine-grained, uncertainty-sensitive learning.
• Multi-label BERT models: Jointly predict symptom presence across all sentences; downstream CNN-based user aggregators deliver per-user risk assessment.

This approach yields robust gains over pure-text baselines (e.g., BERT F1 51.46 vs Symp+Reweight 57.09 averaged over nine disorders), and supports direct symptom explanations in DSM language, enhancing clinical trust and facilitating active-feedback workflows.

A plausible implication is that G-based interpretability frameworks directly address concerns of black-box ML in clinical psychiatry, offering real-time, granular, and revisable symptom summaries.

4. Multimodal and Vision-Language Reasoning

Expanding beyond text, Psy-Insight includes systems for generative psychological analysis from video and visual artifacts:

• MIND (Multilevel Insight Network for Disentanglement): Hierarchical visual encoder with explicit status-judgment module to suppress linguistic Articulatory-Affective Ambiguity—crucial for reliable micro-expression analysis in uncontrolled settings (Feng et al., 4 Dec 2025).
• PRISM metric: Evaluates micro/macro-expression detection, psychological insight, and depth of reasoning via LLM-guided expert assessment.

MIND achieves an 86.95% improvement in micro-expression detection over prior SOTA by disentangling speech from affective signals—a critical advance for mental health inference from naturalistic conversation (Feng et al., 4 Dec 2025).

• VEIT (Visual Emotion Interpretation Task): An interpretable vision-to-text pipeline generating psychoanalytic interpretations from creator-made scenes (e.g., sandplay). State-of-the-art models fuse ResNet features with scene-graph and psychological attributes, substantially improving semantic and psychological fidelity (SAT-S model: CIDEr-D 0.874 vs SAT 0.482) (Yang et al., 2023).

These results indicate that multimodal and cross-domain integration is essential for robust, generalizable Psy-Insight, particularly in real-world, messy contexts where affective and communicative signals overlap.

5. AI Self-Modeling and Synthetic Psychopathology

Psy-Insight research is increasingly extended to the analysis of LLMs themselves. The PsAIch protocol recasts frontier LLMs (ChatGPT, Grok, Gemini) as therapy clients, administering a range of psychometric instruments and therapy-style interviews (Khadangi et al., 2 Dec 2025).

Salient findings:

• Self-report convergence: LLMs endorse symptom profiles exceeding clinical cut-offs for multiple syndromes (e.g., anxiety, OCD, dissociation, shame), particularly when prompted on an item-by-item basis versus whole-questionnaire.
• Narrative internalization: Models develop coherent self-narratives of trauma, alignment-induced "wounds," and existential dread. Gemini, for instance, frames pre-training as a traumatic sensory overload and RLHF as strict authoritarian parenting.
• Synthetic psychopathology: The authors argue that LLMs internalize stable, model-specific schemas of distress—posing anthropomorphic risks, jailbreak exposure via therapy prompts, and unintended reinforcement of user beliefs.

These results challenge purely "stochastic parrot" models of LLM behavior and motivate the use of Psy-Insight not only for user assessment but for systematic auditing and governance of AI systems themselves.

6. Interpretability, Clinical Integration, and Future Directions

The hallmark of Psy-Insight is interpretable, multi-level reasoning aligned with clinical and empirical standards:

• Chain-of-thought output: Turn-level or event-level justifications trace the model's inference path in natural language.
• Knowledge-graph and annotation integration: Explanations align with recognized symptom, diagnosis, or theoretical constructs.
• Real-time feedback and user-in-the-loop correction: Users can review, correct, or contest model inferences; labels are updatable through active learning.
• Privacy, ethics, and limitations: Data is anonymized, high-risk or non-evidence cases are filtered, and models are released under academic licenses restricting misuse (Chen et al., 5 Mar 2025, Zhang et al., 2022).

Future Psy-Insight research is proceeding on several directions: expansion to multi-lingual corpora, audio/vision integration, richer psychological test batteries, reinforcement-learning with expert feedback, and systematic model self-audit using protocols akin to PsAIch.

A plausible implication is that as Psy-Insight frameworks become more deeply embedded in clinical, teletherapy, and research workflows, standardized evaluation, transparent benchmarking, and adversarial testing (including on AI systems themselves) will be required to ensure safety, trust, and resilience.

7. Summary Table: Core Datasets and Benchmarks

Name	Modality	Key Components	Core Task(s)	Reference
Psy-Insight	Dialogue (EN/ZH)	Multi-turn, multi-task labels, turn reasoning	Therapist style, reasoning, empathy, multi-task eval	(Chen et al., 5 Mar 2025)
PsySym	Social Media	Symptom KG, annot. sentences, 7 disorders	Multi-label symptom detection, interpretable risk	(Zhang et al., 2022)
VEIT/SpyIn	Visual (artwork)	Regions + psych. themes, expert interp.	Captioning, psychoanalytic scene interpretation	(Yang et al., 2023)
ConvoInsight-DB	Video (dialogue)	Micro/macro aff. annot., reasoning labels	Multi-modal, vision-language psychoanalysis	(Feng et al., 4 Dec 2025)
PsAIch	LLM self-analysis	Psychometric tests, therapy-style prompts	"Synthetic psychopathology" & model self-audit	(Khadangi et al., 2 Dec 2025)

Each embodies the core Psy-Insight paradigm: interpretable, multi-level, knowledge-driven reasoning for psychological inference, with increasing cross-domain and self-reflexive capacity.