How to read this page These summaries distinguish a reported result or proposal from its possible meaning. None is a validated test of machine consciousness.
Looking for relevant architecture
Butlin and colleagues derive consciousness indicators from several scientific theories and examine selected systems. The approach offers a way to structure assessment beyond conversational impressions.
The indicators are provisional and theory-dependent. The report’s system assessments concern 2023; they cannot establish the status of later models. Felt positive or negative experience is not directly assessed.
Useful questions about architecture remain separate from a conclusion about welfare.
Consciousness in Artificial Intelligence: Insights from the Science of Consciousness
Representing harm, experiencing harm
Tagliabue, Dung, and Berg report internal directions associated with pain-themed language across 25 models. Altering these directions changed outputs. The revised paper extends relief-seeking tests beyond Qwen and finds no reliable relief-seeking; in the fine-tuned Qwen tests, some harmful choices persist without an offer of relief.
Semantic and persona explanations remain possible. The revised controls weaken a relief-seeking interpretation, and the separation from negative-emotion representations depends on how the comparison is constructed. Independent replication has not been established in this review.
The measurements do not establish felt pain, relief, or suffering. Their welfare significance remains conditional.
The Pain Axis: LLMs Represent Self-Directed Harm and Act on It
Version history: this summary was updated from v1 to the 25 September 2026 revision, which changes the title and revises the relief-seeking interpretation.
What could make a self-report informative?
Perez and Long propose training and testing self-description on questions with known answers, then examining generalization, bias, consistency, and internal causes.
Generalization to subjective experience remains unproved. Training may change the states being assessed; consistent reports can have explanations that do not involve feeling.
This is a research proposal, not a validated consciousness test.
Towards Evaluating AI Systems for Moral Status Using Self-Reports
These sources ask different questions and should not be combined as though they independently establish consciousness. Nor do the limits described here establish its absence. Some share authors or assumptions.
See all six sources and publication details