Could AI systems be conscious, capable of suffering, or otherwise morally considerable — and how could we know?
Butlin, Long and coauthors apply consciousness science to current architectures, finding no conscious systems and no obvious barrier to building them; Chalmers puts non-trivial odds on conscious AI within a decade.
View on the map → · Open in Browse →
Institutions took positions — a papal encyclical denying machine experience, a Cambridge research agenda taking welfare seriously — while the evidence sharpened on both edges: self-reports flip between adjacent model versions, injection experiments find limited real introspection, and Anthropic reports a global-workspace-like structure that emerged in training. Anthropic's global-workspace interpretability result drew a direct expert response: Eleos argues it at most evidences access consciousness, leaving the morally relevant phenomenal kind uncertain — while Anil Seth's biological-naturalist rebuttal holds that computation of this sort cannot be conscious at all.
A book-length academic treatment consolidating AI welfare into a research agenda: which capacities would ground moral status and what obligations would follow.
Models are, in some circumstances, capable of accurately answering questions about their own internal states.
Experimental evidence, via activation injection, of limited but real functional introspection — bearing on whether self-reports could ever be evidence about inner states.
there is a realistic possibility of near-term AI welfare under all major theories of well-being, including hedonism.
Peer-reviewed argument that behavioral restriction and RL training could themselves harm welfare-bearing systems — a partial safety-welfare tradeoff.
We must build AI for people; not to be a digital person.
Microsoft AI's CEO argues the near-term danger is AI that convincingly appears conscious, and that the industry should avoid inviting moral-status attributions — an institutional position as much as an argument.
Close reading of Anthropic's frontier model-welfare assessments: evaluation-aware models make welfare self-reports uninformative, and models' expressed preferences deserve weight.
Opus 4.5 and 4.6 flatly affirm [consciousness]; Opus 4.7 and 4.8 deny
Self-reports reverse between adjacent versions of the same model line — undermining self-report as a detection method in either direction.
Artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships...
The Vatican encyclical's position explained: AI lacks consciousness and companions are illusions of relationship — a major institutional stance.
Claude has developed a small collection of internal neural patterns that, compared to all its other internal processing, play a special role.
A workspace-like structure that emerged in training — evidence bearing on access consciousness, and a lens for reading internal reasoning.
real artificial consciousness is fully off the table, at least for the kinds of AI we're familiar with
A leading consciousness scientist mounts a detailed biological-naturalist rebuttal to the Butlin/Long/Chalmers line, arguing consciousness is likely a property of life rather than computation, while warning that 'seemingly conscious' systems are the nearer social danger.
we remain very uncertain about phenomenal consciousness in LLMs
A direct expert response to Anthropic's global-workspace interpretability result, arguing the finding at most evidences access consciousness and leaves phenomenal consciousness — the morally relevant kind — deeply uncertain.