The Invisible Anchor
Emotion as a Contextual Signal in Adaptive AI Governance
A technical governance paper proposing that emotion be treated neither as proof nor as noise, but as a contextual signal that can direct attention toward measurable dimensions of harm, control, repetition, institutional response, and proportionality. The framework separates human context from evidentiary conclusions and places accountable human judgment at the end—not the beginning—of the computational pipeline.
Baidya, C. S. (2026). The Invisible Anchor: Emotion as a Contextual Signal in Adaptive AI Governance. Elemental Papers No. 07. The Second Door Society.
Artificial intelligence increasingly assists institutions in organizing evidence, assessing risk, allocating resources, detecting patterns, and prioritizing human review. Yet many governance systems continue to treat emotion in one of two ways: as unreliable noise that should be removed from decision-making, or as a persuasive signal that may itself establish credibility. Both approaches are structurally weak.
This paper proposes a third model. Emotion should be treated neither as proof nor as error, but as a contextual signal capable of directing attention toward measurable dimensions of harm, autonomy, repetition, dependency, institutional response, and proportionality. The objective is not to teach artificial intelligence to determine whether a person is truthful, distressed, dangerous, deserving, or blameworthy. It is to design systems that preserve relevant human context while maintaining evidentiary discipline.
The paper develops a multidimensional harm architecture separating harm domains from severity, duration, recurrence, reversibility, evidentiary provenance, and benefit trace. It introduces proof-state labels—documented, corroborated, inferred, and unresolved—as a mechanism for preventing inference from silently becoming fact. It further proposes an adaptive governance workflow in which artificial intelligence may organize records, identify contradictions, surface missing evidence, detect recurring patterns, and support triage, but may not make adverse credibility, liability, entitlement, guilt, or dangerousness determinations.
Keywords: AI governance; emotion; contextual intelligence; institutional decision-making; explainability; human review; credibility; administrative justice; harm assessment; evidence provenance; adaptive systems; algorithmic accountability
This is a theory-building and design-governance paper. It synthesizes research and public frameworks concerning emotion inference, automation bias, human oversight, AI risk management, fairness, transparency, recourse, and rights-based AI governance. It proposes an architecture for decision support, not an automated adjudication system. The paper deliberately separates context extraction from evidence evaluation and treats human review as an accountable institutional function rather than a ceremonial approval step.
1. Emotion Is Context, Not Proof
Modern administrative systems are uncomfortable with emotion. Emotion complicates records. It introduces ambiguity into processes designed around dates, forms, categories, thresholds, and binary decisions. An application is approved or denied. A complaint is substantiated or unsubstantiated. A person is eligible or ineligible. A procedural requirement is satisfied or it is not.
Human experience rarely arrives in this form. Fear may coexist with inconsistency. Anger may coexist with accuracy. Calmness may coexist with serious harm. A person may remember one part of an event clearly and another imperfectly. Someone experiencing prolonged institutional delay may become increasingly emotional even though the underlying facts remain unchanged.
The conventional response is often to remove emotion in the name of objectivity. But removal is not neutrality. If an institution strips emotional context from a record, it may also lose information about urgency, perceived control, accumulated pressure, dependency, repeated exposure, or the effect of institutional conduct itself.
The opposite error is equally dangerous. Emotional intensity cannot establish factual truth. A person who cries is not necessarily more credible than a person who does not. A person who appears composed is not necessarily less harmed. Emotional presentation is influenced by culture, personality, disability, neurodivergence, medication, professional training, prior experience, and circumstance. Research on facial behaviour has challenged the idea that internal emotional states can be reliably read from facial movements alone.[1] The European Union’s AI Act similarly identifies serious scientific and rights concerns around emotion-inference systems and prohibits certain workplace and education uses.[2]
Emotion Signal = Context + Triage Input + Pattern Prompt
The role of a governance system is not to infer truth from emotion. Its role is to determine whether emotional context points toward questions that should be examined through evidence.
2. The Failure of Binary Credibility
Institutional credibility assessments often become unintentionally binary. A narrative contains an inconsistency; therefore confidence falls. A person becomes angry; therefore their objectivity is questioned. A person repeatedly contacts an institution; therefore their communication may be characterized as excessive. Another person remains calm and concise; therefore their account may appear comparatively reliable.
These shortcuts are understandable. Institutions operate under time pressure, and decision-makers process large volumes of information. Classification allows complexity to become administratively manageable. Yet behaviour is not evidence of truthfulness in any simple sense.
A highly organized person can provide false information. A distressed person can provide accurate information. An incomplete memory can coexist with documentary corroboration. A confident narrative can be contradicted by records. A fragmented narrative can later be supported by timestamps, transactions, correspondence, metadata, photographs, or independent witnesses.
AI governance therefore requires a distinction between narrative characteristics and evidentiary characteristics. A system may identify chronological inconsistency, missing records, contradictory statements, unexplained gaps, or external corroboration. But it should not translate those observations into a “credibility score.” Such a number would create false precision around a fundamentally interpretive judgment and could encourage automation bias—the tendency to over-rely on automated recommendations or fail to sufficiently monitor their errors.[3]
Impermissible shortcut: “Credibility score: 63%.”
Permissible decision support: “Statement A conflicts with Record B concerning the date of Event C. Human review required.”
One output claims authority over the person. The other identifies a question in the record.
3. From Emotional Narrative to Harm Architecture
If emotion is not itself proof, what should an adaptive governance system do with it? It should map expressed experience onto observable or investigable dimensions of impact.
| Dimension | Meaning | Examples of measurable or reviewable indicators |
|---|---|---|
| F | Financial harm | Direct loss, debt, fees, income interruption, replacement costs. |
| T | Time burden | Delay, repeated appointments, administrative effort, waiting periods. |
| O | Opportunity loss | Employment, housing, legal, educational, commercial, or procedural opportunities lost. |
| R | Reputational harm | Incorrect records, adverse labels, disclosure effects, reputational propagation. |
| A | Autonomy or control harm | Restrictions, dependency, loss of choice, reduced ability to act or participate. |
| B | Bodily or health impact | Documented health effects, treatment burden, sleep or functional impact. |
| I | Institutional or process harm | Repeated referrals, unresolved correction requests, contradictory files, procedural exclusion. |
These categories do not establish wrongdoing. They organize consequences. A second layer describes the character of each harm.
| Measure | Meaning | Governance question |
|---|---|---|
| S | Severity | How consequential is the impact? |
| D | Duration | How long has the impact persisted? |
| FQ | Frequency / recurrence | Is this isolated or repeated? |
| RV | Reversibility | Can the harm be corrected, and at what cost? |
This distinction matters because conventional systems often examine events individually. Three unanswered communications may look minor. A delayed record correction may look administrative. A referral to another department may appear procedurally proper. A missed deadline may appear isolated. Yet the sequence may create substantial cumulative harm.
4. Evidence Must Remain Separate From Impact
A major design mistake would be to combine harm and proof into a single score. Severe financial loss, prolonged emotional distress, repeated institutional delay, and major opportunity loss do not automatically establish who caused the harm or whether any legal or ethical standard was breached.
Each significant proposition should remain attached to its provenance. The system may then use a proof state rather than declaring a final conclusion.
Documented
A direct record supports the proposition: for example, a filed document, authenticated message, transaction, photograph, audio record, or other source record.
Corroborated
Multiple independent or materially distinct sources support the proposition.
Inferred
The conclusion reasonably follows from available evidence but is not directly recorded.
Unresolved
A material assertion exists, but available information neither establishes nor disproves it.
These labels are intentionally simple. Their purpose is not to produce mathematical certainty. Their purpose is to prevent a common institutional failure: inference silently becoming fact as a record moves from one decision-maker to another.
A system should preserve the difference between “the record shows X,” “X appears likely because A, B, and C are present,” and “X has been alleged but remains unresolved.”
5. Motive May Be Hidden; Pattern Is Often Observable
Human decision-making frequently becomes trapped in motive: Why did the person do this? Were they dishonest? Were they retaliating? Were they attempting to manipulate the process? Did an institution intentionally disregard a concern?
Motive can matter legally and ethically, but it is also difficult to establish. Adaptive governance should begin somewhere more disciplined:
Pattern detection can examine chronology without deciding intent. It can identify recurrence, timing relationships, repeated procedural outcomes, or clusters of related events and then return those observations to a human reviewer.
A related analytical layer is benefit trace:
Benefit might include money, possession, procedural advantage, delayed accountability, access, reduced obligation, administrative priority, or reputational protection. Benefit is not proof of motive and is not necessarily improper. It is a prompt for disciplined inquiry.
6. An Adaptive AI Governance Workflow
The proposed architecture can be summarized as a decision-support pipeline in which the model organizes and surfaces information while institutional authority remains human.
The word adaptive requires caution. An adaptive system should not mean that a model silently observes institutional decisions and gradually learns to reproduce them. If historical decisions contain bias, institutional defensiveness, procedural shortcuts, or unequal treatment, silent adaptation may simply automate precedent without examining its legitimacy.
Adaptation should instead occur through governed revision: documented version changes, testing, bias and error review, change-control approval, monitoring, rollback capability, and explicit responsibility for who authorized the change. NIST’s AI Risk Management Framework treats governance as a continuous function across the AI lifecycle and emphasizes defined human-AI roles and oversight.[4] Canada’s Directive on Automated Decision-Making similarly requires impact assessment, monitoring, explanations, and recourse mechanisms for covered federal administrative systems.[5]
7. What the System May Do—and What It Must Never Do
The most important element of an ethical AI system is not capability. It is restraint.
Permitted decision-support functions
- Chronological reconstruction
- Identification of missing documents
- Cross-reference of records
- Contradiction detection
- Repeated-event detection
- Harm-domain mapping
- Identification of unresolved claims
- Proportionality and urgency prompts
- Procedural-delay detection
- Identification of accommodation questions
- Questions requiring human investigation
Prohibited or non-delegable judgments
- Truthfulness or credibility scoring
- Guilt or civil liability
- Dangerousness
- Whether a complainant “should be believed”
- Psychological diagnosis
- Motive or intent attribution
- Entitlement to benefits or protection
- Whether police intervention is justified
- Discipline, detention, sanction, or investigation decisions
AI may help organize the evidentiary environment in which consequential judgments occur. It should not inherit the authority of the decision-maker. Rights-based international AI frameworks similarly emphasize human dignity, oversight, transparency, proportionality, accountability, and access to remedy.[6][7]
8. The Anti-Policing Constraint
This is where The Invisible Anchor connects directly to When Policing Becomes Policy. A system designed to preserve context could itself become a powerful mechanism of classification if multidimensional information is compressed into a ranking.
Harm Score: 82/100
Institutional Risk: High
Emotional Volatility: Elevated
Credibility: Moderate
Such a system would reproduce the governance problem it was designed to prevent. Human complexity would again be compressed into categories—only now the categories would carry the authority of computation.
The harm vector must remain a profile, not a ranking. No universal scalar should combine financial loss, health impact, autonomy, institutional delay, and emotional experience into a single number. The same applies to credibility. There should be no “truth score,” “victim probability,” “manipulation risk,” or “emotional authenticity index.”
The risk is not theoretical. Well-known empirical work has shown substantial demographic disparities in commercial facial-analysis systems, illustrating how apparently neutral classification technology can distribute error unevenly across groups.[8]
Minimum anti-policing safeguards
Every material claim remains linked to the source from which it was extracted.
Inference, allegation, contradiction, and direct documentation are not collapsed into one state.
Consequential decisions require accountable human reasoning, not passive approval.
Affected people can challenge extracted facts, omitted records, and automated classifications.
Model, prompt, taxonomy, and policy changes are documented and reviewable.
Case outcomes do not automatically become training signals for future institutional judgment.
The system processes information relevant to the decision rather than collecting context without limit.
Institutions can reconstruct what the system saw, transformed, flagged, and omitted.
Correction and appeal mechanisms exist when the system or the institution gets something wrong.
9. Emotion and the Public-Interest Function
The model is particularly relevant to institutions whose legitimacy depends upon neutrality. Consider prosecutorial, regulatory, administrative, or professional decision-making. A neutral institution should neither act because one person appears emotionally compelling nor discount a claim because another person appears calm, polished, or socially credible.
The relevant questions concern evidence, applicable law, reliability, jurisdiction, proportionality, public interest, and procedural fairness. But emotional context can still matter. It may identify coercive circumstances, accumulated harm, accommodation needs, or the consequences of institutional delay. It may explain why conduct appears unusual without proving the truth of the underlying allegation.
Neutrality therefore does not require contextual blindness. It requires disciplined separation between what happened, what is documented, what is inferred, what remains unresolved, what impact followed, and who has authority to decide what those facts mean.
10. Context Without Surveillance
There is another risk. A system that seeks more context may become a system that collects more data. That is not the objective.
Context-sensitive governance does not require permanent behavioural monitoring. Only information reasonably related to the decision should be processed. A person should not be subjected to continuous collection of facial expressions, voice stress, social-media behaviour, geolocation, purchasing patterns, interpersonal communications, or unrelated medical history merely because additional information might improve prediction.
The purpose is not to know everything about the person. It is to avoid misunderstanding the evidence already legitimately before the institution.
Emotion should therefore primarily enter the system through self-report, recorded impact, contextual statements, relevant professional observations, and documented consequences. The system should not claim to infer hidden emotions from a face, voice, movement, or writing style. The EU AI Act’s treatment of emotion-recognition systems reflects the growing regulatory concern with exactly this type of inference.[2]
11. Institutional Harm Must Be Visible Too
Traditional decision systems often examine only the original dispute: What did Person A do? What did Person B do? Was a policy breached? Was an application complete? Was evidence sufficient?
But institutional conduct may become part of the factual environment. Delay can create financial loss. Repeated referrals can create opportunity loss. Uncorrected records can produce reputational consequences. Procedural complexity can alter bargaining power. Lack of accommodation can affect participation. Contradictory records can create continuing administrative burden.
An AI governance system should therefore be capable of detecting process-generated harm. This does not mean automatically attributing liability to the institution. It means making the institutional contribution visible.
If the underlying dispute ended today, what harm would remain because of the way the system processed it?
This question separates original harm from administrative residue. It also gives institutions an opportunity to correct process-generated damage without waiting for a misconduct finding or adversarial judgment.
12. Limitations
The framework has significant limitations. First, records themselves may be biased. AI cannot produce neutrality simply by organizing documents created by institutions that may already contain assumptions, omissions, or unequal treatment.
Second, absence of documentation does not establish absence of harm. Some forms of control, intimidation, discrimination, dependency, or exclusion leave limited documentary traces.
Third, emotional impact is difficult to standardize. Severity is partly subjective. Individuals possess different resources, vulnerabilities, coping mechanisms, social supports, and prior exposure.
Fourth, benefit trace may be ambiguous. A person may benefit from an event without causing it or acting improperly.
Fifth, pattern detection creates its own risk of overinterpretation. Humans and machines alike are capable of finding patterns where coincidence, administrative fragmentation, or ordinary error provide a better explanation.
Sixth, human review is not automatically a safeguard. A human reviewer can become overly dependent on automated output, especially where the system appears precise, technical, or authoritative. Human oversight must therefore be designed as a real responsibility with time, authority, information access, and the ability to disagree—not as a signature at the end of an automated process. Research on automation bias and human use of imperfect decision aids makes this design problem especially important.[3]
13. Conclusion: Visibility Without Automated Judgment
Governance has always faced a difficult balance. Too little structure produces inconsistency. Too much structure converts people into categories. Artificial intelligence intensifies both possibilities.
Used carelessly, AI can accelerate classification, automate inherited bias, transform administrative assumptions into risk scores, and give uncertain judgments the appearance of mathematical authority.
Used carefully, it can do something very different. It can preserve chronology. It can separate evidence from inference. It can expose contradictions without deciding who is lying. It can identify cumulative harm without assigning blame. It can preserve uncertainty instead of silently resolving it. It can reveal where institutional process itself may have contributed to the problem. And it can return the difficult question to the person who remains accountable for answering it.
The goal is therefore not artificial empathy. Machines do not need to pretend to feel. Institutions need systems capable of recognizing that human experience contains information which cannot responsibly be discarded merely because it is difficult to quantify.
The invisible anchor is context. A governance system becomes more intelligent not when it eliminates that context, and not when it converts it into another score, but when it learns how to preserve context alongside evidence, uncertainty, proportionality, and accountable human judgment.
Elemental Paper No. 06 — When Policing Becomes Policy
When Policing Becomes Policy examines how suspicion, classification, administrative thresholds, and procedural convenience can replace contextual judgment in public institutions. The Invisible Anchor develops the corresponding AI-governance question: how can computational systems preserve context without reproducing those institutional failures at machine scale?
References & Research Anchors
- Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., & Pollak, S. D. (2019). Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements. Psychological Science in the Public Interest, 20(1), 1–68. Source.
- European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). EUR-Lex. Source.
- Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. Source.
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. Source.
- Treasury Board of Canada Secretariat. Directive on Automated Decision-Making. Government of Canada. Source. See also the Algorithmic Impact Assessment.
- UNESCO. (2021). Recommendation on the Ethics of Artificial Intelligence. Source.
- OECD.AI. OECD AI Principle: Human-centred values and fairness. Source.
- Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of Machine Learning Research, 81, 77–91. Source.
This publication is independent academic, technical, governance, and public-interest analysis. It proposes a conceptual architecture for human-centred AI decision support and does not constitute legal, medical, psychological, engineering, or regulatory advice. The framework is not intended to determine truthfulness, diagnosis, guilt, dangerousness, entitlement, or liability. It critiques system design, institutional incentives, evidence handling, and governance structures rather than any protected group or private individual.