RRI MENTAL AI RESEARCH FRAMEWORK · VERSION 2.0 · 10 NOVEMBER 2026
Physical AI studies intelligent systems acting in the physical world. Mental AI studies artificial models of mental processes, AI's effects on human mental life, and the dynamics through which both change in interaction.
Research lead: Ryan SangBaek Kim, Founder and Principal Investigator, Ryan Research Institute (RRI), Paris Published: 10 November 2026 · An integrated framework defining three research domains, shared evaluation principles, and directions for international collaboration. Scope: Artificial mental functions · AI-mediated human mental life · coupled human–AI dynamics
Mental AI | 인공지능과 정신적 세계의 통합 연구 · Research
Field Definition
Mental AI brings together the study of artificial mental functions, the effects of AI on human mental life, and the reciprocal dynamics through which humans and AI shape each other's judgments and behaviour. RRI's unified framework connects these three domains through shared research questions and methods of evaluation.
The field studies functional models of mental processes, AI's consequences for human mental life, and the interaction between these processes. Questions of machine consciousness are examined on their own evidentiary and philosophical terms. The field includes beneficial forms of assistance, creativity, learning and therapeutic support as well as harmful forms of manipulation, exclusion, interpretive displacement and dependence. RRI studies these outcomes through defined measures, comparative evidence and reproducible tests.
Mental AI connects cognitive science, neuroscience, philosophy of mind, affective computing, computational psychiatry, human–computer interaction, social psychology, AI evaluation and governance. RRI brings these fields into dialogue through a unified research agenda and testable questions.
The Three Domains
I. Artificial Mental Functions
How do artificial systems represent, simulate, coordinate, and functionally implement capacities associated with mental life?
The domain examines memory, attention, reasoning about belief and intention, self-modelling, metacognitive monitoring, goal representation, affective modelling, and adaptation. It asks what observable functional claims are warranted and how they can be tested under distribution shift, intervention, contradictory evidence and changes in context.
Research in this domain tests observable computational functions. Subjective experience, sentience, personhood and moral standing remain distinct empirical and philosophical questions requiring their own evidence.
Examples: testing whether a model's stated uncertainty predicts its errors; whether a self-model supports correction; whether theory-of-mind performance generalises across contexts; and whether a synthetic affect model supports interpretable control across controlled changes in prompts and system conditions.
II. AI-Mediated Human Mental Life
What does the presence and use of AI do to human cognition, emotion, self-understanding, relationships, well-being, learning and agency?
Here the object of study is the person. AI can offer prompts for reflection, lower barriers to support, improve access to information and help people learn. It can also create misplaced confidence, narrow expressive possibilities, induce dependence or change perceptions of expertise and authority. Direction, magnitude and persistence are empirical questions.
Examples: controlled studies of emotion labels and later self-interpretation; longitudinal research on attachment to conversational agents; changes in independent decision-making after AI assistance; and the effects of AI-mediated support on cognitive, emotional and behavioural outcomes.
III. Coupled Human–AI Mental Dynamics
What happens when people and AI systems continuously respond to, adapt to, or evaluate one another?
This domain addresses feedback, interactive alignment, reciprocal modelling, personalised memory, relational continuity, performative prediction, criterion formation and institutional decisions. Research designs separately measure changes in human behaviour, AI system behaviour and institutional decision rules, then examine the causal links among them.
Examples: whether repeated model interpretations change subsequent self-reports; whether user corrections meaningfully update model or organisational behaviour; and whether post-deployment evaluation criteria remain independent of the system that helped produce them.
The three domains are overlapping research lenses. A companion robot can be studied through Physical AI and Mental AI together. Domain I includes computational reasoning systems studied independently of human deployment; Domains II and III examine recommendation services through human effects and reciprocal interactions.
Relationship to Physical AI
Physical AI concerns systems that perceive and act in the physical world, as in robotics and autonomous machines. Mental AI is proposed as an equally legible research heading for mental processes and mental consequences. The comparison distinguishes research priorities and leaves room for hybrid systems that cross both fields.
Spatial reasoning can inform a person's mental model; a social robot can exercise physical control while making inferences about human intention. Work crossing the domains belongs to both. The distinction is useful when it determines different evidence, risks, methods, and standards of evaluation.
From the July 2026 Study to the Unified Research Framework
The July 2026 paper must retain its original, narrower definition. In the public preprint Mental AI: Post-Deployment Criterion Endogeneity in the Causal Loop of Human Mental Life (posted 29 July 2026), Mental AI identifies a causal class requiring three jointly considered conditions:
- An inference about human mental life.
- Operational uptake of that inference in a response or decision.
- A non-trivial, inference-indexed consequence pathway back to the person, their expression, state, opportunity or treatment.
The July 2026 study examined how AI interpretations of a person's mental life could shape subsequent responses, behaviour and evaluation. The November framework connects this causal research to a broader programme encompassing artificial mental functions and the mental lives of AI users.
Under the proposed taxonomy, the 2026 causal class is a specific operational subcategory, primarily within Domain III, with substantial connections to Domain II. Post-deployment criterion endogeneity is a narrower, testable problem within that subcategory: after an AI inference shapes the person or their environment, a later criterion may become causally dependent on the earlier inference and require an independent validity check.
Controlled comparisons and independent outcome measures establish when criterion movement occurs and how large its effects are.
Foundational research preprint · 29 July 2026
Scientific Foundations and Research Principles
RRI's unified framework advances research questions that can be answered, challenged and refined. Four cross-cutting commitments guide its three domains.
Clear definitions and measures. Define whether a study measures machine functionality, human subjective experience, self-report, outward behaviour or institutional treatment. These are different variables.
Testing cause and effect. Differentiate pre-existing differences, system-induced changes, strategic adaptation, demand effects, selective observation and changed evaluation criteria. Whenever possible, use randomisation, blinded assessments, exposure histories and pre-registered contrasts.
Human choice and accountability. Evaluate interpretive authority, consent, privacy, ability to contest or correct a model's account, and the practical consequences of institutional adoption. Effective review carries justified corrections through to later outputs, records and decisions.
Independent testing and open evidence. Make scope conditions, negative cases, data provenance, uncertainty, and meaningful benchmarks inspectable. Independent comparisons will refine the framework's explanatory scope and show where shared questions produce new research value.
Research and Evaluation Programme
RRI's next research phase centres on the following test designs and evaluation criteria.
Research domain | Central research question | Evidence and measures |
Artificial Mental Functions | How do AI systems monitor, model and revise their own performance? | Calibration, transfer across contexts, self-correction and intervention tests |
AI-Mediated Human Mental Life | How does AI shape human judgment, emotion and self-understanding? | Independent assessment, change over time, duration of effects and comparison groups |
Coupled Human–AI Dynamics | How do repeated interactions change judgments and evaluation criteria? | Feedback dynamics, independent reference measures and downstream correction |
The next phase will develop an open Mental AI taxonomy, shared evidence and safety protocols, a transparent set of boundary cases, reproducible evaluation tasks, and interdisciplinary critical review. Each deliverable will be documented as it is produced and independently evaluated.
RRI welcomes independent research teams to apply the framework, evaluate its methods and develop new comparative studies.
Relationship to Existing RRI Research
Affective Sovereignty and Interpretive Authority: normative and design questions concerning a person's standing when computational explanations are applied to emotional life.
Delegated Precision: a predictive-processing hypothesis concerning external affective labels and subsequent self-interpretation; with its human outcome effects identified as priorities for controlled investigation.
Resonant Amplification Framework: a mechanistic and design account of reinforcing human–AI interaction loops.
DefMoN, Algorithmic Affective Blunting and Affective Thermodynamic Relationship: distinct methods and findings concerning synthetic research data, model interpretation, and evaluation under affective or normative stress.
Narrative–affect measurement and ANAD: empirical resources for separating language from an unwarranted one-to-one inference about mental states.
These works provide distinct theoretical, computational and empirical foundations. Each contribution is identified by its research question, method, evidence and publication status; new field-wide claims will be evaluated through dedicated studies.
Public Evidence and Publication Status
Foundational research preprint (29 July 2026): Kim (2026), Mental AI: Post-Deployment Criterion Endogeneity in the Causal Loop of Human Mental Life, Preprints.org, 29 July 2026.
Selected related peer-reviewed works:
- Formal and Computational Foundations for Implementing Affective Sovereignty in Emotion AI Systems, Discover Artificial Intelligence (2026).
- Interrupting Resonant Amplification: A Mechanistic and Design Framework for Human–AI Interaction, Computers in Human Behavior Reports (2026).
- Narrative–Affect Discrepancy as a Regulated Degree of Freedom in 351,734 Relationship Narratives, PLOS ONE (2026).
- Delegated Precision: A Predictive-Processing Hypothesis on External Affective Labelling and Self-Interpretation, Frontiers in Psychology (2026), Hypothesis and Theory.
Research commentary: Can an AI Evaluate a Mind It Has Already Helped Change?; The Human Mind Is Becoming AI's Operating Environment; The Test Begins When the AI Label Disappears. These essays communicate scholarly arguments and research questions, with the original peer-reviewed papers linked separately.
From Research to Public Deliberation
The From Interpretive Authority to Public AI Accountability | Seoul 2026 addressed one application of the broader agenda: AI inferences about citizens and the conditions under which institutions may act on those inferences. RRI's six operational safeguards offer a research-informed route for evaluating consequential public-sector AI decisions.
Mental AI also encompasses positive applications and research questions outside public governance. Its interdisciplinary scope creates a shared agenda for research, replication and critical discussion across existing fields.
Authorship, Citation and Research Collaboration
Framework proponent: Ryan SangBaek Kim, Ryan Research Institute, Paris. Public framework edition: RRI Mental AI Unified Field Framework, Version 2.0, 10 November 2026. Suggested citation: Kim, R. S. (2026). Mental AI: A Proposed Unified Field for Artificial Mental Functions, AI-Mediated Human Mental Life, and Coupled Human–AI Dynamics. RRI Mental AI Unified Field Framework (v2.0). Ryan Research Institute, Paris.
RRI publishes this authored and citable framework for independent replication, research use and scholarly criticism. The original contribution lies in the defined three-domain structure, its research criteria and their integration. The framework is available for assessment and development by scholars and institutions worldwide. New public research resources and independent evaluations will be linked as the programme develops.
Published by: Ryan Research Institute, Paris · Edition: v2.0 (10 November 2026).
Research and media correspondence: ryan@ryanresearch.org.