Auto-Psych: Formalizing Machine Psychology

• Auto-Psych is a framework that encodes, infers, and manipulates human psychological structures in a machine-readable form for applications like compatibility matching and agent self-modeling.
• It spans multiple methods, from graph-based personal profiles and psychometric inference to psychology-aware control systems in automated counseling and driving.
• Evaluation strategies include statistical models, multimodal data analysis, and automated experimental designs that quantify latent psychological constructs and system self-improvement.

Auto-Psych can be understood as an umbrella theme for systems that formalize, infer, store, manipulate, or experimentally probe psychological structure in machine-readable form. Across the literature, this includes electronically encoded human psychological profiles for compatibility matching, automatic personality and conflict recognition from behavioral or interview data, psychology-aware control layers for counseling and automated driving, explicit architectures for an artificial psyche, psychometric characterization of AI systems themselves, and agentic systems that automate psychological theory discovery (Gluskin, 2011, Bogdan et al., 10 May 2026, Prystawski et al., 24 Jun 2026).

1. Scope, lineage, and conceptual boundaries

A useful historical axis begins with early attempts to encode psychology as a computable internal structure. "Toward Psycho-robots" (0709.2065) proposed a metric “mental space” in which ideas are points, unconscious processing is modeled by dynamical systems, and conflict, repression, and symptom formation are implemented through distances to databases of “interesting,” “forbidden,” and “repressed” ideas. "The 'psychological map of the brain', as a personal information card (file)" (Gluskin, 2011) proposed a simpler but socially oriented architecture: a personal electronic card/file containing a graph of “main brain centers” and weighted directed influences among them, to support matching for marriage, work, and other close relations. These early works did not present validated deployed systems; they are better read as conceptual blueprints for machine-readable psychology.

Later work diversified the agenda. Some papers target automated inference about human psychological variables from multimodal or behavioral data; some target automation around psychological care; some model the psyche as an internal architecture for artificial agents; some treat AI systems as psychometric subjects; and some automate the scientific study of mind itself (Kimura et al., 2023, Liu et al., 23 Apr 2025, Kolonin et al., 16 Mar 2026, Khadangi et al., 2 Dec 2025, Prystawski et al., 24 Jun 2026).

Strand	Representative papers	Core object
Machine-readable human profiles	(Gluskin, 2011, Pessanha et al., 2022, Hossain et al., 27 Mar 2025)	traits, conflicts, compatibility
Psychology-aware automation	(Liu et al., 23 Apr 2025, Shen et al., 2021, Bao et al., 13 Jun 2025)	counseling, trust, personalized control
Artificial psyche and self-modeling	(0709.2065, Kolonin et al., 16 Mar 2026, Li et al., 2024, Prado, 5 Mar 2026)	needs, identification, dynamic modulation
Psychometrics of AI	(Riva et al., 3 Nov 2025, Khadangi et al., 2 Dec 2025, Bogdan et al., 10 May 2026)	calibration, self-models, synthetic psychopathology
Automated science of mind	(Prystawski et al., 24 Jun 2026)	theory discovery and experimentation

A persistent boundary condition in this literature is that Auto-Psych is not identical to neuroscience. Gluskin’s “psychological map of the brain” is explicitly not a CT or MRI map, but a systems-style representation of behavioral “centers” and their directed couplings (Gluskin, 2011). A second boundary is that Auto-Psych is not necessarily autonomous psychotherapy. Systems such as PsyCounAssist are framed as therapist-support architectures, not clinician replacements (Liu et al., 23 Apr 2025).

2. Structured representations and inference targets

One major branch of Auto-Psych treats psychological description as a representational problem. In Gluskin’s proposal, a person is encoded as nodes such as love, anger, patience, and responsibility, with directed edge weights $a_{ik}$ or $d_{ik}$ representing the “probability of influence” of center $i$ on center $k$. The representation is therefore graph-like or matrix-like, and the intended computational use is automatic compatibility comparison between two personal informational cards (Gluskin, 2011). The worked toy example $a_{12}=0.4;\; a_{21}=0.1$, where “Love feelings” strongly influence “Anger,” illustrates how Auto-Psych can reduce a complex profile to a compact weighted interaction structure.

A second representational strategy uses observational or metadata features to infer psychometric variables. The ICCV 2021 ChaLearn "Automatic self-reported personality recognition track" (Pessanha et al., 2022) predicts Big Five / OCEAN self-reports by multitask regression from dyadic interaction data. Its main methodological result is a warning: a model trained solely on simple metadata features—age, gender, and number of sessions—achieved superior or similar performance to simple audio, linguistic, or visual systems, with test MSE $=0.825$, and on development data the metadata model outperformed the challenge baseline overall ($1.0405$ vs. $1.0632$ MSE). This directly exposes demographic and participation confounds as a central Auto-Psych problem.

A third strategy targets clinically structured latent constructs from long-form language. "AutoPsyC: Automatic Recognition of Psychodynamic Conflicts from Semi-structured Interviews with LLMs" (Hossain et al., 27 Mar 2025) processes full-length OPD interviews averaging about 90 minutes. Its final task covers four Axis III conflicts—self-dependency/dependency on others, dominance/submissiveness, self-sufficiency, and self-value/self-esteem—each scored on a five-point scale. The method combines summarization, RAG, LoRA-fine-tuned segment-specialized LLaMA 3.1 8B models, and weighted aggregation

$\hat{y} = \arg\max_c \sum_{i=1}^k w_i\, p_i(c),$

with best weighted F1 scores of $0.78$, $d_{ik}$0, $d_{ik}$1, and $d_{ik}$2 across the four retained conflicts. This suggests that Auto-Psych systems can target latent, partially unconscious constructs, but only with substantial architectural support for long context, domain knowledge, and temporal specialization.

At the broadest measurement-theoretic level, "Machine Psychometrics: A Mathematical Psychology of Artificial Intelligence" (Bogdan et al., 10 May 2026) proposes the Machine Mindprint: a multidimensional, domain-bounded, versioned profile spanning calibration, source integrity, suggestibility resistance, context stability, expressive alignment, tool integrity, drift monitoring, and distributional grounding. Its core formal exemplar is an IRT model,

$d_{ik}$3

used not merely for correctness, but for latent dispositions such as abstention, source monitoring, and suggestibility.

3. Counseling automation and longitudinal psychological support

A second branch of Auto-Psych automates parts of psychological care workflows. "PsyCounAssist: A Full-Cycle AI-Powered Psychological Counseling Assistant System" (Liu et al., 23 Apr 2025) is explicitly therapist-support, not therapist-replacement. It decomposes counseling support into three modules: Real-time Emotion Prediction (REP), Automated Structured Session Reporting (ASSR), and Personalized Follow-up Support (PFS). REP combines speech and PPG, updates affective estimates every 60 seconds, and reports with latency under 1 second. ASSR uses GPT-4 or DeepSeek to generate five-section structured reports from transcripts or manually annotated text. PFS generates short follow-up check-ins from session summaries, emotion trends, and therapeutic goals.

The most concrete quantitative evidence in PsyCounAssist comes from its PPG branch. In a pilot with 30 participants, Random Forest achieved validation accuracy $d_{ik}$4, validation F1 $d_{ik}$5, test accuracy $d_{ik}$6, and test F1 $d_{ik}$7 for binary sad-versus-relax classification from wrist-worn PPG-derived features (Liu et al., 23 Apr 2025). The paper itself notes important limitations: the speech and multimodal branches are not quantitatively validated in the deployed setting, the physiological online inference section mixes PPG with SCR/GSR terminology, and the binary training task does not match the deployed 3-class sad/neutral/positive scheme.

"PsychAgent: An Experience-Driven Lifelong Learning Agent for Self-Evolving Psychological Counselor" (Yang et al., 1 Apr 2026) pushes further toward autonomous improvement. Its architecture combines a Memory-Augmented Planning Engine, a Skill Evolution Engine, and a Reinforced Internalization Engine. Memory is formalized as $d_{ik}$8, where $d_{ik}$9 is an evolving client profile and $i$0 is episodic session history. Skill retrieval is stage- and context-aware, and high-reward trajectories are internalized by rejection fine-tuning: $i$1 On PsychEval, PsychAgent achieved counselor shared $i$2, counselor specific $i$3, client shared $i$4, and client specific $i$5, outperforming strong general and counseling-specific baselines across the reported aggregated dimensions (Yang et al., 1 Apr 2026).

Taken together, these systems define a clinically bounded Auto-Psych pattern: persistent memory, staged planning, multimodal state sensing, structured documentation, between-session continuity, and selective internalization of high-quality practice. A plausible implication is that counseling-oriented Auto-Psych now spans both workflow automation and agent self-improvement, but the literature still treats real-world safety, privacy, and crisis handling as unresolved.

4. Mobility, trust, and psychology-aware autonomous driving

Automated driving is the most developed human-factors application area in this corpus. One line of work studies the psychology of the human–AV relation as a personality-matching problem. "An Automated Vehicle (AV) like Me? The Impact of Personality Similarities and Differences between Humans and AVs" (Zhang et al., 2019) uses a sample of $i$6 U.S.-licensed drivers in a $i$7 within-subjects video study. Its central result is nuanced: similarity between human and AV personalities improved perceived AV safety only when both were high in certain traits, especially agreeableness, conscientiousness, and emotional stability, while dissimilarity improved safety when the AV was higher than the human on those traits. Extroversion and openness showed no reliable safety effects.

A second line inserts subjective feeling directly into control. "Evaluation and Control Model Design of Human Factors for Autonomous Driving Systems" (Deng et al., 2023) learns subjective confidence categories from vehicle dynamics using four classifiers; Mahalanobis distance performs best at $i$8 accuracy. The resulting confidence model and psychologically motivated obstacle/road potential fields are then integrated into MPC: $i$9 The paper also formalizes a psychologically last point to steer (PLPTS), with reported distances of $k$0, $k$1, $k$2, and $k$3 m at $k$4, $k$5, $k$6, and $k$7 km/h. This is a clear Auto-Psych move from physical safety to felt safety.

A third line focuses on driver modeling from behavior. "Estimating Driver Personality Traits from On-Road Driving Data" (Kimura et al., 2023) uses naturalistic on-road data from 23 older drivers and shows that road-type segmentation matters. The strongest cognitive-function results are $k$8 for TMT(B) and $k$9 for UFOV. The study is strongest for cognitive-function estimation and weaker for broader DSQ/WSQ self-reports, but it shows that psychologically relevant traits can be inferred from ordinary telemetry when context is modeled explicitly.

Trust calibration is addressed directly by "AutoPreview: A Framework for Autopilot Behavior Understanding" (Shen et al., 2021). AutoPreview introduces a target autopilot $a_{12}=0.4;\; a_{21}=0.1$0 and a delegate autopilot $a_{12}=0.4;\; a_{21}=0.1$1 that mirrors the target while outputting explainable actions $a_{12}=0.4;\; a_{21}=0.1$2. In a pilot with 10 participants, $a_{12}=0.4;\; a_{21}=0.1$3 reported that they rarely or never check release notes, while $a_{12}=0.4;\; a_{21}=0.1$4 said they would prefer a previewing tool prior to purchase or deployment. The treatment group using preview cues was more accurate at exact action-timing prediction, yet all participants in that group chose not to deploy the aggressive target autopilot, compared with $a_{12}=0.4;\; a_{21}=0.1$5 deployment preference in the comparison group. The result supports trust calibration rather than trust maximization.

The most explicit occupant-state architecture is "Your Ride, Your Rules: Psychology and Cognition Enabled Automated Driving Systems" (Bao et al., 13 Jun 2025). PACE-ADS introduces three foundation-model agents: Driver Agent for scene semantics, Psychologist Agent for passive psychological signals and active instructions, and Coordinator Agent for high-level behavioral decisions. The Psychologist Agent consumes EEG, heart rate, facial image, and verbal command; the Coordinator outputs

$a_{12}=0.4;\; a_{21}=0.1$6

In CARLA closed-loop simulations, PACE-ADS adapts braking, speed, lane-change gaps, pedestrian yield distances, and route replanning to states such as very impatient, impatient, relaxed, anxious, and very anxious. The system is positioned as a semantic planning layer above conventional AV modules, activated when occupant state changes or instructions are issued.

5. Artificial psyche, self-modeling, and behavior modulation in AI

Another major Auto-Psych trajectory attempts to engineer a psyche directly into artificial agents. In "Toward Psycho-robots" (0709.2065), ideas are points in a metric mental space, unconscious processing is iteration under maps $a_{12}=0.4;\; a_{21}=0.1$7, and motivational filters are distance-based. The key valuation functions are explicit: $a_{12}=0.4;\; a_{21}=0.1$8 and, in the “normal” case,

$a_{12}=0.4;\; a_{21}=0.1$9

Model 4 adds repression, a “domain of doubts,” repressed-idea collectors, and symptom formation. This is one of the clearest early attempts to give AI an internal psychodynamic architecture rather than a task policy.

"Computational Concept of the Psyche" (Kolonin et al., 16 Mar 2026) recasts the psyche as an operating system over a state space that includes needs, sensations, and actions. Long-term need priorities $=0.825$0 and momentary need actualization $=0.825$1 jointly determine motivation, and reinforcement is formalized as

$=0.825$2

The paper’s minimal implementation is a single-player ping-pong agent with a four-dimensional need space—Happy, Sad, Novelty, Expectedness—and reports that excessive weighting of negative reinforcement can impair learning. The result is small-scale, but it operationalizes a needs-based Auto-Psych architecture.

"Enabling self-identification in intelligent agent: insights from computational psychoanalysis" (Li et al., 2024) aligns Lacanian imaginary and symbolic identification with active inference. Its claim is that imaginary identification is characterized by an integrated body schema with minimal free energy, while symbolic identification is mediated by language and the Other. The proposed FreeAgent combines CNN-based body-image processing, MLP-based symbolic mapping, and ChatGPT as an avatar of the Other. The system is more conceptual than fully formalized, but it extends Auto-Psych from motivation to self-identification.

The same agenda appears in more behaviorally grounded form in "Ailed: A Psyche-Driven Chess Engine with Dynamic Emotional Modulation" (Prado, 5 Mar 2026). Ailed decomposes behavior into static personality and dynamic psyche, with

$=0.825$3

updated from five positional factors and then passed through an audio-inspired signal chain—noise gate, compressor/expander, five-band equalizer, saturation limiter—to reshape move probabilities. Across 12,414 games, both tested probability sources showed a monotonic top-move-agreement gradient of about 20–25 percentage points from stress to overconfidence, and for Maia2+Psyche the competitive score fell from $=0.825$4 under overconfidence to $=0.825$5 under stress. The paper explicitly disclaims human-subject validation, but it demonstrates a modular low-state Auto-Psych layer that changes how a policy acts without changing what it knows.

6. Psychometrics of AI and automated science of mind

A distinct but convergent line treats AI systems themselves as psychometric objects. "Automatic Minds: Cognitive Parallels Between Hypnotic States and LLM Processing" (Riva et al., 3 Nov 2025) organizes the analogy around automaticity, suppressed monitoring, and heightened contextual dependency. It argues that hypnotized cognition and LLMs both exhibit functional agency without subjective agency, and introduces the “observer-relative meaning gap”: coherent but ungrounded outputs whose meaning is completed by an external interpreter. This is a psychological model of LLM failure modes, not a claim that LLMs literally instantiate human hypnosis.

"When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models" (Khadangi et al., 2 Dec 2025) goes further by treating LLMs as psychotherapy clients. PsAIch has two stages: therapy-style elicitation over up to four weeks, then a psychometric battery covering syndromes, empathy, Big Five traits, dissociation, shame, and self-consciousness. Its central empirical claim is that item-by-item “therapy-style” administration can push a base model into multi-morbid synthetic psychopathology, while whole-questionnaire prompts often lead ChatGPT and Grok, though not Gemini, to recognize the instruments and produce strategically low-symptom answers. The paper is explicit that this does not establish subjective suffering; it establishes structured, model-specific self-descriptions and psychometric behavior.

"Machine Psychometrics" (Bogdan et al., 10 May 2026) generalizes this into a measurement science. Its philosophical intervention is to reject both Artificial Mind Blindness and Artificial Mind Projection in favor of Artificial Mind Discipline: measurement before judgment. The Trust Protocol links Mindprints to deployment through probe batteries, perturbation testing, reliability and validity analysis, runtime verification, and longitudinal monitoring. This yields a non-anthropomorphic but explicitly psychological language for calibration, source integrity, context stability, expressive alignment, and tool integrity.

Finally, "auto-psych: Automating the science of mind using agent-driven theory discovery and experimentation" (Prystawski et al., 24 Jun 2026) inverts the direction of analysis. Instead of profiling people or AI systems, it automates theory discovery in computational cognitive science through nested agent loops. The outer loop selects experiments by expected information gain,

$=0.825$6

launches them on Prolific, and pools the data; the inner loop fits and critiques probabilistic cognitive models using ELPD-LOO and posterior-predictive checks. On human subjective-randomness judgments, the best discovered model, “Minkowski typicality,” achieved $=0.825$7, RMSE $=0.825$8, and $=0.825$9, outperforming seeded literature-inspired theories and explaining about $1.0405$0 of the explainable variance. This is Auto-Psych in a strict methodological sense: automation of the science of mind itself.

Across these lines of work, several tensions recur. Many proposals are speculative, pedagogical, or proof-of-concept rather than clinically or operationally validated (0709.2065, Gluskin, 2011, Li et al., 2024). Several benchmark results are highly sensitive to confounds, administration protocol, or deployment wrapper (Pessanha et al., 2022, Khadangi et al., 2 Dec 2025). The strongest measurement-oriented papers explicitly avoid treating fluent behavior as evidence of consciousness (Riva et al., 3 Nov 2025, Bogdan et al., 10 May 2026). The field’s likely near-term trajectory is therefore not a single unified “machine psychology,” but a layered ecosystem: structured human profiling, psychology-aware control, artificial psyche architectures, psychometrics of AI, and automated cognitive science, linked by a common commitment to formalizing psychological structure as something that can be encoded, inferred, tested, and monitored.