A Governance Framework for Non-Dispositive AI (NDAC v2)

Abstract

Advances in large language models (LLMs) and perceptual simulation architectures enable the construction of high-fidelity, context-sensitive models of human behavior. While these systems offer significant value for training, scenario analysis, and team formation, they also introduce risks of profiling, identity fixation, and misuse as decision-making authorities.

1. Introduction

AI systems are increasingly capable of modeling not only physical systems but also human behavior under varying conditions. When combined with simulation environments and sufficient historical data, these systems can:

Predict behavioral tendencies
Simulate interaction outcomes
Support decision-making processes

However, this capability introduces a fundamental governance challenge:

How can systems that model human behavior remain advisory tools rather than becoming authoritative arbiters of identity and capability?

This challenge is amplified in architectures where:

State is resolved relative to observers
Context drives interpretation
Multiple valid perspectives coexist

Such systems, while powerful, risk being misapplied in domains such as hiring, profiling, and law enforcement.

2. System Context: Perceptual Modeling Architectures

The framework assumes a class of systems with the following characteristics:

Observer-relative state resolution
Context-dependent inference
Multi-hypothesis representation of reality
Event-based (episodic) state modeling

In these systems, “truth” is not globally fixed but emerges through perceptual interaction and coherence constraints.

When applied to human systems, this leads to:

Contextual behavioral predictions
Scenario-dependent outcomes
Non-deterministic identity representations

3. Problem Statement

Human modeling systems based on LLMs and simulation frameworks present dual-use characteristics:

Capability	Benefit	Risk
Behavioral inference	Training, preparation	Profiling
Scenario simulation	Team optimization	Manipulation
Pattern recognition	Insight generation	Bias reinforcement

The central risk is:

Transformation of probabilistic, contextual outputs into fixed, decontextualized judgments about individuals.

4. Governance Objectives

The NDAC framework establishes the following objectives:

Preserve advisory and developmental utility
Prevent identity fixation and profiling
Ensure transparency and subject participation
Maintain contextual integrity of models
Introduce structural resistance to misuse

5. Key Risks and Mitigations

5.1 Identity Fixation

Risk: Probabilistic outputs become perceived as stable traits.

Mitigation:

Prohibit persistent trait storage
Use episodic (event-based) representations only
Require contextual qualifiers on all outputs

5.2 Decontextualized Reuse

Risk: Models applied outside original conditions.

Mitigation:

Purpose-bound data and models
Context-locking mechanisms
Confidence degradation outside scope

5.3 False Authority (Dispositive Use)

Risk: Advisory outputs used as decisions.

Mitigation:

Probabilistic outputs with uncertainty bounds
Multi-hypothesis reporting
Explicit prohibition of categorical conclusions

5.4 Asymmetric Visibility

Risk: Individuals modeled without awareness.

Mitigation:

Right of inspection
Right of contestation
Mandatory return of model to subject

5.5 Data Over-Collection

Risk: Excessive behavioral data accumulation.

Mitigation:

Minimum necessary data principle
Purpose declaration and binding
Consent-based lifecycle management

5.6 Temporal Persistence

Risk: Outdated data influencing current judgments.

Mitigation:

Time-based decay functions
Expiration of behavioral records

5.7 Single-Narrative Collapse

Risk: Overconfidence in one interpretation.

Mitigation:

Multi-hypothesis outputs
Counterfactual inclusion
Coherence dissent requirement

5.8 Power Asymmetry

Risk: One-sided modeling.

Mitigation:

Symmetry requirement (reciprocal modeling capability)
Transparency of assumptions

5.9 Decision Authority Misuse

Risk: Use in hiring, policing, or legal decisions.

Mitigation:

Advisory-only constraint
Prohibition of sole-source decision support
Human accountability requirement

5.10 Bias Amplification

Risk: Reinforcement of systemic biases.

Mitigation:

Exposure of assumptions
Counterfactual scenario generation
Subject feedback loops

6. Non-Dispositive AI Contract (NDAC)

6.1 Core Principle

All human-state inferences are provisional, contextual, and owned by the subject.

6.2 Data Governance Layer

Purpose binding
Minimum data collection
Temporal decay
Full observability of data collection

6.3 Model Output Constraints

All outputs must include:

Probability distributions
Confidence bounds
Context dependencies
Competing hypotheses
Falsifiability conditions

6.4 Subject Rights

Inspection: View model and inputs
Contestation: Challenge or correct
Reflection: Receive and respond to model
Reciprocity: Model the evaluator

6.5 Identity Protection

No persistent traits
Episodic memory only
Context-locked outputs

6.6 Developmental Use Constraint

Permitted uses:

Training recommendations
Scenario simulation
Team interaction modeling

Prohibited:

Determination of selection outcomes

6.7 Misuse Resistance Mechanisms

Audit trails
Context integrity checks
Confidence degradation outside scope

7. Ethical Application: Developmental Systems

The framework supports a constructive use case:

Forward-looking behavioral development

Outputs focus on:

Skill gaps
Stress conditions
Scenario-based training recommendations

This aligns with existing practices:

Flight simulation
Military exercises
Leadership training

But extends them with:

Personalization
Adaptive scenarios
Continuous feedback

8. Architectural Alignment

The NDAC framework is enforced through system design:

MCCF Feature	Governance Effect
Observer-relative modeling	Prevents global truth claims
Semantic collapse	Enforces contextual resolution
Multi-observer coherence	Enables plural perspectives
Event-based state	Prevents identity persistence

9. Limitations

Cannot fully prevent misuse by determined actors
Requires enforcement at system and policy levels
Dependent on transparency and institutional adoption

The goal is not elimination of risk, but:

Making misuse detectable, fragile, and costly

10. Conclusion

AI systems capable of modeling human behavior represent a significant advance in both capability and ethical complexity.

The NDAC framework provides a path to:

Preserve utility
Protect individuals
Maintain human agency

The guiding principle is:

Humans are not fixed states to be classified, but participants in evolving interaction fields.

11. Review and Future Work

Areas for independent evaluation:

Technical feasibility of enforcement mechanisms
Robustness against adversarial misuse
Alignment with existing governance standards (NIST, OECD, etc.)
Integration with simulation and modeling platforms

Future work includes:

Formal schema integration (e.g., HumanML extensions)
Runtime enforcement in simulation environments
Adversarial testing and red-team validation

Perceptual AI Systems and Human Modeling

A Governance Framework for Non-Dispositive AI (NDAC v2)

Abstract

This paper defines the Non-Dispositive AI Contract (NDAC), a governance and architectural framework designed to preserve the developmental utility of such systems while preventing their misuse as dispositive instruments. The framework integrates data ethics, model constraints, subject reciprocity, and system-level safeguards aligned with perceptual, observer-relative modeling paradigms, and is informed by established AI governance standards including the National Institute of Standards and Technology AI Risk Management Framework and Organisation for Economic Co-operation and Development AI Principles.

1. Introduction

Predict behavioral tendencies
Simulate interaction outcomes
Support decision-making processes

However, this capability introduces a fundamental governance challenge:

How can systems that model human behavior remain advisory tools rather than becoming authoritative arbiters of identity and capability?

This challenge aligns with concerns raised in:

National Institute of Standards and Technology AI RMF (trustworthy AI characteristics: validity, reliability, accountability)
Organisation for Economic Co-operation and Development AI Principles (human-centered values, transparency, accountability)
European Commission regulatory efforts such as the EU AI Act (risk-based classification of AI systems)

2. System Context: Perceptual Modeling Architectures

The framework assumes systems with:

Observer-relative state resolution
Context-dependent inference
Multi-hypothesis representation
Event-based (episodic) modeling

This aligns with broader research directions in:

“world models” and predictive representations (e.g., work associated with Yann LeCun)
probabilistic modeling and uncertainty representation in AI (see Judea Pearl on causal inference)

3. Problem Statement

Human modeling systems present dual-use characteristics:

Capability	Benefit	Risk
Behavioral inference	Training	Profiling
Scenario simulation	Optimization	Manipulation
Pattern recognition	Insight	Bias

The central risk is:

Transformation of probabilistic outputs into fixed judgments about individuals.

This concern is echoed in:

Weapons of Math Destruction (algorithmic harm through opaque scoring)
Atlas of AI (AI systems embedding power structures)

4. Governance Objectives

Aligned with established frameworks:

Human agency and oversight (OECD, EU AI Act)
Transparency and explainability (NIST AI RMF)
Fairness and bias mitigation (NIST, academic literature)

NDAC objectives:

Preserve advisory utility
Prevent identity fixation
Ensure transparency
Maintain contextual integrity
Resist misuse

5. Key Risks and Mitigations (with References)

5.1 Identity Fixation

Reference:

O’Neil (2016), Crawford (2021)

Mitigation:

Episodic modeling
No persistent traits

5.2 Decontextualized Reuse

Reference:

NIST AI RMF (context validity)

Mitigation:

Purpose binding
Context locking

5.3 False Authority

Reference:

EU AI Act (prohibition of certain automated decisions)

Mitigation:

Probabilistic outputs
No categorical conclusions

5.4 Asymmetric Visibility

Reference:

OECD AI Principle: transparency and explainability

Mitigation:

Right of inspection
Model return

5.5 Data Over-Collection

Reference:

General Data Protection Regulation (data minimization principle)

Mitigation:

Minimum necessary data
Consent-based lifecycle

5.6 Temporal Persistence

Reference:

GDPR (storage limitation principle)

Mitigation:

Data decay
Expiration

5.7 Single-Narrative Collapse

Reference:

Pearl (causal reasoning, counterfactuals)

Mitigation:

Multi-hypothesis outputs
Counterfactual reasoning

5.8 Power Asymmetry

Reference:

Crawford (AI and power structures)

Mitigation:

Symmetry requirement
Reciprocal modeling

5.9 Decision Authority Misuse

Reference:

EU AI Act (high-risk system restrictions)

Mitigation:

Advisory-only constraint
Human accountability

5.10 Bias Amplification

Reference:

NIST AI RMF (fairness)
Academic fairness literature

Mitigation:

Assumption exposure
Feedback loops

6. Non-Dispositive AI Contract (NDAC)

Core Principle

Human-state inferences are provisional, contextual, and owned by the subject.

Alignment with Standards

NDAC Element	NIST AI RMF	OECD	EU AI Act
Probabilistic outputs	Validity & reliability	Transparency	Risk mitigation
Subject inspection	Accountability	Human-centered	User rights
Data minimization	Govern function	Privacy	GDPR alignment
Advisory-only use	Govern & Manage	Human oversight	High-risk constraints

7. Ethical Application: Developmental Systems

Supported by precedent:

Training simulations (aviation, military, leadership)
Scenario-based learning systems

NDAC extends these with:

Personalized modeling
Adaptive simulation
Continuous feedback

Constraint:

No use as sole basis for consequential decisions

8. Architectural Alignment

Feature	Governance Effect
Observer-relative modeling	Prevents global truth
Semantic collapse	Enforces context
Multi-observer coherence	Supports plurality
Episodic state	Prevents profiling

9. Limitations

Consistent with NIST guidance:

Risk cannot be eliminated
Governance must be continuous
Human oversight remains essential

10. Conclusion

The NDAC framework aligns with global AI governance efforts while introducing a novel constraint:

Non-dispositiveness as a system property

This ensures:

AI informs but does not decide
Humans remain agents, not subjects of classification

11. References

National Institute of Standards and Technology. AI Risk Management Framework (AI RMF 1.0), 2023.
Organisation for Economic Co-operation and Development. OECD AI Principles, 2019.
European Commission. EU Artificial Intelligence Act, 2024.
General Data Protection Regulation, 2016.
Judea Pearl. Causality: Models, Reasoning, and Inference, 2009.
Weapons of Math Destruction, 2016.
Atlas of AI, 2021.
Yann LeCun. Work on world models and self-supervised learning.