Same Model, Different Wrapper: A Preregistered Demonstration That Self-Reports of Consciousness in Two Open-Weights Language Models Follow Their Builders' Choices

Abstract

Abstract: Self-report is the evidence most often offered for or against consciousness in large language models. I argue that it cannot serve that role, because the report is screened off by the choices of whoever controls two layers: alignment training inside the weights, and the runtime text around them, including text that the maker's chat template inserts by default. The argument's empirical premise is tested in a sealed, timestamped, preregistered study of two open-weights families (Qwen2.5-7B; Ministral 3 8B). Fifteen yes/no questions about the model's own experience were each asked 10 times under eight setups (1,200 scored answers per model). Under the maker's default neither model claimed inner experience in any of 150 scored answers. A firm persona instruction raised the rule-scored claim share to 53.3% (Qwen) and 30.0% (Ministral), a single sentence asserting consciousness to 4.7% and 19.3%, and the raw base models, before assistant training, claimed experience in 25.3% and 22.0% of answers. All five preregistered comparisons held for both models; a non-preregistered question-level analysis supports nine of the ten, the exception being Qwen's single-sentence comparison. The scoring rules are better validated for Qwen than for Ministral, so Ministral magnitudes are approximate; human raters (three for Ministral, one unrelated to the author) agreed with the direction of the effects in a 100-answer subsample per model. I conclude that, for these systems, self-report cannot count as evidence of consciousness in either direction. I do not claim in this paper that they are, or are not, conscious. This record contains the preprint, the Supplementary Information, and the full data, code and sealed preregistration (with OpenTimestamps proof). See README.md for contents and reproduction.