Triangulating Evidence for Machine Consciousness Claims: A Validity-Centered Stack of Behavioral Batteries, Mechanistic Indicators, Perturbation Tests, and Credence Reporting

Summary

Frontier AI systems produce responses that lead people to wonder if they might be conscious. A new framework, the Triangulated Consciousness Assessment Stack (TCAS), combines four evidence streams to distinguish genuine indicators from optimized artifacts or surface-level cues: behavioral batteries with robustness controls, mechanistic indicators with explicit assumptions, perturbation tests probing causal sensitivity and proxy failures, and observer-confound controls separating anthropomorphic attribution from evidence. When all streams are available, TCAS produces theory-indexed credence bands and standardized disclosure cards rather than binary verdicts. An empirical evaluation of GPT-5.2 Pro covered only behavioral and perturbation streams; missing streams meant credence bands were withheld.

Study at a glance

Abstract

Frontier AI systems are now producing responses that make users, developers, and policymakers genuinely pause and ask: could these models have conscious experiences? Yet the field still lacks rigorous, hard-to-game tools that can distinguish genuine indicators from optimized artifacts or surface-level cues. We introduce the Triangulated Consciousness Assessment Stack (TCAS), a validity-centered framework that combines four evidence streams: behavioral batteries with robustness controls (B), mechanistic indicators with explicit assumptions (M), perturbation tests that probe causal sensitivity and proxy failures (P), and observer-confound controls that separate anthropomorphic attribution from evidence (O). When all streams are available, TCAS produces theoryindexed credence bands and standardized disclosure cards (TCAS Cards) rather than binary detection verdicts. We report an empirical evaluation of GPT-5.2 Pro via OpenRouter (2026-02-19 UTC) covering B and P streams only, including a pre-specified role-play negative control. M and O were not run in this black-box walkthrough, so theory-indexed credence bands are explicitly withheld under the missing-stream rule. Prompts, rubric, judge prompt, raw outputs, and provenance manifest are released at the repository commit cited in the camera-ready build.