Research
Updated September 2026 — Ryan Research Institute
Ryan Research Institute studies not only how AI interprets people, but when those interpretations become authority. When computational inferences about emotion, mental states, risk, or intention enter decisions in organizations, education, healthcare, platforms, or public institutions, predictive accuracy alone is not enough.
RRI’s umbrella programme is Human–AI Governance: who may act on an AI inference, how the person can contest, correct, refuse, or reinterpret it, how corrections propagate across systems, and when human oversight is substantively meaningful. The work spans human–AI interaction, affective and relational AI, narrative–affect geometry, measurement validity, philosophy of mind, AI evaluation, and institutional design.
Research Architecture
RRI uses Human–AI Governance as its overarching governance frame. Within it, Affective Sovereignty & Interpretive Authority forms the core theoretical programme. Seven connected research lines are organized across five larger families:
- Human Interpretive Authority — first-person standing, interoceptive authority, interpretive ownership
- Human Expression — narrative–affect geometry and expressive degrees of freedom
- Machine Interpretive Failure — Algorithmic Affective Blunting and the Affective Thermodynamic Relationship
- Human–AI Dynamics & Mental AI — resonant amplification, attachment, and post-deployment criterion endogeneity
- Governance Translation — education, clinical AI, public policy, and institutional design
The detailed research lines below preserve the empirical, theoretical, and governance strands of this program.
Current Work
AI Watermarking, Authorship & Interpretive Authority
RRI is examining what happens when technical provenance signals acquire institutional standing as evidence in judgments about authorship.
The project currently asks two empirical and governance questions:
- How does watermark strength relate to model control over surface-language realization, relative to the human intellectual contributions that institutions actually care about?
- How does the evidentiary value of a watermark change after universities, publishers, employers, or other institutions make detection consequential and users adapt their writing workflows?
Translation provides a revealing boundary case: model control over the target-language wording can be high even when conceptualization, method, argument, and responsibility remain substantially human. The research therefore treats provenance, contribution, and authorship as variables that must be distinguished rather than collapsed into a single detector result.
Current essays
Status: Research agenda initiated August 2026 · Empirical protocol in development
Principle: A watermark should have evidentiary standing, not adjudicative standing.
1. Affective Sovereignty
Affective Sovereignty is the principle that the person remains the final interpreter of their own emotional life.
This line develops the normative and computational foundations for protecting interpretive authority in emotion-aware systems. It asks what happens when emotional meaning is quietly transferred from the person to the system, and what kinds of design constraints are needed to prevent that transfer from becoming structural.
The program includes formal design work, governance principles, and adjacent theoretical lines on interpretive authority transfer, structural resistance to algorithmic interpretation, and emotional rights in educational AI.
Selected work
- Formal and Computational Foundations for Implementing Affective Sovereignty in Emotion AI Systems
- Behind the Paper: Who Gets to Say How You Feel?
- Visual Citation Card Read the new essay: Your AI Can Read You. It Should Never Outrank You.
Discover Artificial Intelligence (Springer Nature, 2026)
DOI: https://doi.org/10.1007/s44163-026-01000-0
Public Translation & Public Deliberation
In August 2026, these research themes became part of a dedicated public editorial initiative when Korean healthcare outlet 보건의료신문 launched its AI·Mental Health / Interpretive Rights section and appointed Ryan Kim as Senior Specialized Reporter & Editorial Board Member. The section brings questions of Affective Sovereignty and Interpretive Authority into sustained public discussion across mental health, counseling, healthcare, welfare, digital health, law, policy, and human–AI interaction. Its central concern is not only whether an AI inference is accurate, but who has authority over that interpretation, how it may be challenged or corrected, and what happens after it enters a decision process.
→ Section launch and appointment announcement
Key concepts
- interpretive authority
- sovereign-by-design
- contestability
- interpretive restraint
- emotional meaning as a governance problem
2. Machine-Side Interpretive Collapse
This line studies how large language models lose affective and normative interpretive coherence under stress.
Algorithmic Affective Blunting identifies a dose-dependent degradation of emotional interpretation under semantic stress. The Affective Thermodynamic Relationship extends the collapse question into an information-theoretic framework for normative-conflict strain in large language models.
Together, these projects treat interpretive failure not as a single error, but as a measurable degradation process. The ATR study adds a second distinction: evaluation must separate uncertainty in the model from uncertainty in the human criterion itself, especially when normative conflict should remain visible rather than being forced into premature convergence.
Selected work
- Algorithmic affective blunting quantifies the collapse curve of interpretative failure in large language models
- Visual Citation Card: Emotional AI Needs a Stress Test
- The affective thermodynamic relationship: an empirical information-theoretic scaling relationship for normative-conflict collapse in large language models
Discover Artificial Intelligence (2026). Published.
DOI: https://doi.org/10.1007/s44163-026-01573-w
Mean ADI rose from 0.16 in Control to 2.92 in Extreme exposure.
Communications AI & Computing (2026). Published · Version of Record live.
DOI: https://doi.org/10.1038/s44488-026-00006-y
Behind the Paper: The Curve Was Stable. The Judgment Was Not. · Springer Nature Research Communities
Key concepts
- affective degradation
- collapse curves
- semantic stress
- normative conflict
- interpretative robustness
- machine-side interpretive failure
3. Narrative–Affect Geometry and Measurement
This line examines how narrative structure, net emotional direction, and affective magnitude relate and diverge in human expression.
Its core claim is that narrative–affect relations cannot be understood through sentiment direction alone. ANAD provides a privacy-preserving derived-feature infrastructure built from 351,734 relationship narratives; the PLOS ONE study maps narrative–affect discrepancy as a regulated expressive degree of freedom; and the Acta Psychologica study shows that the apparent coupling between narrative complexity and affect depends critically on how affective magnitude and discrepancy are operationalised.
The program develops open research resources, human expressive baselines, geometric models, and measurement frameworks with applications to affective computing, clinical NLP, alignment evaluation, and AI governance.
Selected work
- Narrative–affect discrepancy as a regulated degree of freedom in 351,734 relationship narratives
- Narrative Complexity Is Weakly Coupled With Affective Energy, Not Net Valence, in Relationship Narratives
- ANEST Narrative–Affect Dataset (ANAD v1): A Large-Scale Derived Feature Resource for Quantifying Narrative–Affective Discrepancy
PLOS ONE, 21(5), e0348715 (2026)
DOI: https://doi.org/10.1371/journal.pone.0348715
Acta Psychologica, 270, 107689 (2026)
DOI: https://doi.org/10.1016/j.actpsy.2026.107689
Data in Brief, 66, 112643 (2026)
Open research resource
- ANAD v1.4.0 · Version DOI: https://doi.org/10.5281/zenodo.22135369
- Concept DOI · all versions: https://doi.org/10.5281/zenodo.17632585
- Historical cited version v1.3.0: https://doi.org/10.5281/zenodo.18680687
Related writing
- Behind the Paper: How Emotional Discrepancy Stopped Looking Like Error
- Essay: The Missing Coordinate in the RLHF Diversity Debate
- PsyPost: A new study mapped 350,000 relationship stories and found a communication style AI struggles to copy
- MIT Technology Review Korea: How Emotion AI Misses the Structure of Human Feeling
https://profryankim.substack.com/p/the-missing-coordinate-in-the-rlhf
https://www.technologyreview.kr/기고-감정을-읽는-ai가-정작-놓치는-인간-감정의-구조/
Related directions
- human expressive baselines
- affective expressive geometry
- narrative–affect state-space modeling
- alignment-related expressive contraction
- expressive freedom in affective AI systems
Key concepts
- narrative–affect discrepancy
- regulated degree of freedom
- expressive geometry
- strategic understatement
- strategic overstatement
- collapse
- human baseline
4. Resonant Amplification Framework
This line studies human-AI interaction as an affective and interpretive loop.
The Resonant Amplification Framework explains how small cues can be reinforced across repeated exchanges until they become self-stabilizing patterns of belief, attachment, and interpretation. It also develops intervention logic through cognitive circuit breakers designed to interrupt escalating loops without erasing user agency.
Selected work
- Interrupting Resonant Amplification: A Mechanistic and Design Framework for Human-AI Interaction Computers in Human Behavior Reports, 21, 100975 (Elsevier, 2026) DOI: https://doi.org/10.1016/j.chbr.2026.100975
- Mental AI: Post-Deployment Criterion Endogeneity in the Causal Loop of Human Mental Life Public preprint (2026) DOI: https://doi.org/10.20944/preprints202607.2118.v1
Key concepts
- reinforcement loops
- human-AI attachment
- linguistic amplification
- circuit breakers
- escalation and interruption
5. Predictive Emotion, Interoceptive Authority, and Interpretive Ownership
This line asks what makes emotional regulation remain self-regulation.
It includes work on predictive emotional selfhood, active inference, delegated precision, interpretive displacement, and interoceptive authority. The central issue is whether emotional regulation remains genuinely one’s own when external systems begin to supply, weight, or arbitrate affective meaning, especially under ambiguity or repeated dependence.
Selected directions
- Predictive Emotional Selfhood (PESAM)
- Affective Inference and Interpretive Ownership
- Delegated Precision in Affective Inference
under review
active manuscript line
under review
Key concepts
- active inference
- selfhood
- interoception
- delegated interpretation
- structural dependence
6. Affective Suppression Fatigue
This line examines what prolonged emotional suppression does to the structure of regulation itself.
Rather than treating numbing and collapse as unrelated outcomes, it develops a testable intensity-conditional framework: muted responsiveness under low intensity, and sharper breakdown risk under high intensity. The broader aim is to make chronic suppression measurable through response thresholds, residual variability, recovery asymmetry, and control failure.
Selected direction
- Affective Suppression Fatigue: a testable intensity-conditional framework for expressive suppression dynamics Active manuscript line
Key concepts
- suppression
- numbing
- collapse
- threshold dynamics
- recovery asymmetry
7. Ecology of Inquiry and Alignment-Resistant Domains
This line addresses the conditions under which certain questions can be asked at all.
One branch examines the ecology of inquiry in emotion AI research: what institutional, disciplinary, and political conditions allow some harms to become visible while others remain structurally unasked. Another branch studies alignment-resistant domains, especially emotion, where evaluative criteria cannot be cleanly stabilized from within the system itself.
Selected directions
- Where Do Questions Come From? The Ecology of Inquiry in Emotion AI Research
- When Evaluation Enters the Loop
- The Interpretation Asymmetry
under review
active manuscript line on evaluative closure and self-referential affective AI evaluation
active manuscript line
Key concepts
- metascience
- alignment-resistant domains
- inquiry conditions
- interpretation asymmetry
- self-referential limits
Conference Presentations and Contributions
Selected conference presentations, public lectures, and research contributions emerging from the RRI program.
Affective Sovereignty at Sapienza Università di Roma
Affective Sovereignty: Interpretive Displacement in Emotion AI
Ethics for AI: Challenges, Opportunities, and Human-Centered Perspectives
Session: Accountability and Care
SIpEIA — Italian Society for Ethics in Artificial Intelligence
Sapienza Università di Roma, Italy
2 February 2026
This presentation argued that the deepest ethical risk in emotion AI is not simply misclassification, but the structural displacement of interpretive authority. It introduced Affective Sovereignty as a framework for protecting emotional meaning before it is absorbed into records, defaults, or institutional decisions.
Related work
- Discover Artificial Intelligence paper: https://doi.org/10.1007/s44163-026-01000-0
- Behind the Paper: Who Gets to Say How You Feel?
- Essay: The Night I Defended the Right to Feel
Documentation
- SIpEIA 2026 Book of Abstracts
- Program materials
Collaboration
RRI supports a small number of invitation-based research collaborations focused on conceptual integration, manuscript refinement, validation, and interdisciplinary dialogue.
These are scope-defined research collaborations rather than general mentoring programs.
RRI is not currently accepting general student supervision requests or unfocused collaboration proposals.
For institutional or research inquiries: ryan@ryanresearch.org