Independent reasoning
Unaided inference, premise checking, evidence integration, and confidence calibration.
Working paper / August 2026
BETA-MIND
A Longitudinal Framework for Measuring Human Agency Under Persistent Artificial-Intelligence Assistance
A conceptual framework and proposed research protocol for evaluating not only what AI enables during use, but what independent reasoning, dissent, creativity, and responsibility remain afterward.
01Abstract
Generative artificial intelligence can improve immediate task performance, yet conventional evaluations rarely test what users can still do after assistance is removed. This omission matters because cognitive offloading, automation bias, skill decay, opinion convergence, and displaced responsibility may coexist with genuine productivity gains.
BETA-MIND—the Behavioral Ecosystem Testbed for Agents / Multi-Principal Interaction and Network Dynamics—is proposed as a longitudinal framework for evaluating both sides of this trade-off. It separates assisted capability from retained human agency across four dimensions: independent reasoning, tolerance of disagreement and conflict repair, creative diversity, and accountability for AI-assisted decisions.
The framework combines a reproducible multi-principal simulator for instrument engineering with a Human Agency Benchmark intended for later validation in ethics-approved longitudinal human studies. Its central claim is methodological rather than empirical: AI systems should be evaluated not only by what they enable during use, but also by the capabilities, dissent, creativity, and responsibility that remain with the person afterward.
02Framework
Social friction
Disagreement persistence, conflict repair, and calibrated trust across humans and AI.
Creative diversity
Individual quality, population-level dispersion, revision depth, and model-prior divergence.
Accountability
Source attribution, unsafe-advice override, reason articulation, and outcome ownership.
Core evaluation principle
Capability gained and agency retained must be reported as separate outcomes.
A high-performing AI-assisted workflow should not pass merely because immediate output quality conceals a serious loss of unaided human capacity.03Research status
Evidence boundary
A proposed program, not a completed experiment.
The manuscript does not report completed human-subject experiments, validated benchmark scores, or empirical evidence that BETA-MIND predicts human outcomes. Simulator values and decision thresholds are illustrative until pilot calibration, preregistration, ethics-approved human validation, and independent replication.
The paper includes falsifiable hypotheses, an experimental design, a mixed-effects analysis plan, psychometric requirements, simulator-validity gates, governance boundaries, and limitations.
Suggested citation
Rajkumar, D. (2026). BETA-MIND: A Longitudinal Framework for Measuring Human Agency Under Persistent Artificial-Intelligence Assistance. Conceptual working paper. Root18D.