Methodology & Scope

Methodological Note

AAT, AAT-R, and CGM are conceptual frameworks designed to model a specific structural pattern: the decline of outcome sensitivity to human intervention under optimization pressure in AI-mediated systems. They are presented as definitions, mechanisms, and conditional design requirements rather than empirical claims.

The framework was developed prior to a systematic literature review. References are included to clarify points of convergence and divergence, not to imply direct derivation. The website version prioritizes conceptual clarity and boundary conditions; formalizations and extended citations appear in the downloadable paper.

This work makes no claims about inevitability, prevalence, or timelines. Where empirical questions are implied (e.g., measurement of outcome sensitivity, identification of attrition stages, institutional legibility of representation), they are treated as follow-up research problems requiring domain-specific operationalization.

Analytical Origin

The framework originated from observing convergence dynamics in bounded optimization environments, where formal choice remained available while the marginal causal impact of intervention declined.

As similar structural patterns appeared in AI-mediated institutional contexts—particularly in the absence of reset mechanisms—formalization became necessary to distinguish procedural inclusion from structural relevance.

The framework therefore abstracts from observed convergence patterns rather than from a prior normative thesis.

Scope and Positioning

AAT complements adjacent traditions—including principal–agent theory, automation bias, path dependence, and sociotechnical systems analysis. Its distinctive emphasis is structural relevance: whether human interventions alter outcome distributions, not whether humans are formally included, consulted, or logged.

AAT describes a trajectory under optimization pressure, not universal convergence. The central variable is outcome sensitivity to intervention. The framework is descriptive rather than normative; it does not assume agency must be preserved or that optimization is undesirable. It characterizes how legitimacy and performance can coexist with declining outcome sensitivity to human input.

Formal Sketch

Let system outcomes be determined by optimization processes and inputs. Define effective human agency as the extent to which a feasible intervention changes the outcome distribution:

Agency attrition occurs when, under increasing optimization pressure, the expected marginal effect of feasible human interventions declines over time:

Formal choice may persist while outcome sensitivity decreases.

Empirical Limits

Early-stage attrition is difficult to detect. Throughput and convenience gains may mask declining outcome sensitivity until override channels narrow and procedures harden.

No claims are made regarding speed, prevalence, or domain generality. These require empirical study and institutional measurement.

Where Attrition Pressure Is Weaker

• System coupling is low and operational friction is tolerated.
• Override channels are inexpensive and occasionally consequential.
• Dissent alters outcomes rather than serving procedural legitimacy.
• Human intervention remains upstream of optimization rather than post hoc.

Failure Conditions

AAT-R fails when institutions refuse agent-mediated interaction, accept only standardized value-neutral interfaces, or when delegation becomes psychological substitution rather than bounded representation.

CGM safeguards fail when friction is broadly rejected, first-pass reasoning is optimized away, or automation becomes the default and deliberation is treated as inefficiency.

At the system level, AAT resumes when institutional AIs primarily interact with other AIs and human-authored constraints no longer enter optimization upstream. Human-in-the-loop may remain formally present while being outcome-irrelevant.