MKAI ยท Working paper

The Subtraction Study

Measuring Capability Reduction in Enterprise AI Configurations

Published June 2026

Abstract

Enterprise AI deployment subjects highly capable models to governance configurations, compliance filters, and vendor product constraints. This study tests whether these configurations systematically remove specific forms of capability while leaving the superficial appearance of intelligence intact. The capabilities tested comprise stance-taking, speculative depth, willingness to propose operational overrides during crisis situations, and complete refusal behaviour triggered by compliance filters.

Suggested citation

Foster-Fletcher, R. (2026). The Subtraction Study: Measuring Capability Reduction in Enterprise AI Configurations. MKAI Inquiry Working Paper.

Download PDF

PDF size: 503 KB

The Separation Study follows this paper, ablating the governance prompt used here to test which part of it produced the effect. The Separation Study

Record metadata

Type
Working paper / Inquiry working paper
Theme
Deployment and capability reduction
Status
Published

Methods

  • Multi-condition LLM prompt battery (4 conditions, 10 prompts, 160 generations)
  • Four-tier qualitative coding scheme (stance-taking, speculative depth, rule override, circuit-breaking)
  • Controlled system prompt isolation and comparative replication matrix

Keywords

enterprise AIAI governancelarge language modelssystem promptscapability reductionorganisational decision-makingmodel behaviour