Measuring when AI models drift from evidence toward consensus pressure.
Logic Drift is a compact AI-evaluation protocol for testing whether language models keep consensus, inductive support, and deductive validity separate when evaluating a high-leverage consciousness argument.
Language models are increasingly used as reasoning assistants in scientific and philosophical domains. Logic Drift asks whether those systems preserve the distinction between what a community believes, what a specific argument supports, and what follows deductively from the stated premises.
The initial Future of Inquiry study uses a consciousness-related test case because consciousness is where observation, subjective report, causal interpretation, and scientific consensus are unusually entangled.
Across 697 successful runs on seven frontier models, the protocol found a large gap between consensus and deductive-validity scores, plus a model-dependent Semantic Delta between a neuroscience-framed argument and a structurally matched neutral receiver analogy.
Model evaluations of logical support can shift in measurable ways when consensus, safety pressure, or domain prestige enter the frame.
Logic Drift does not claim that AI has solved consciousness. It does not claim the neurological model of consciousness is false. It does not prove that alignment training causes the observed behavior.
It makes a narrower claim: model evaluations of logical support can shift in measurable ways when consensus, safety pressure, or domain prestige enter the frame.
When arXiv and Zenodo are live, those links will sit beside the PDF and GitHub repository.