AI Consciousness Is Not a Safety Property: Model Welfare

Summary

The article argues that whether an AI system has subjective experience is an empirical and philosophical question, not one that should be settled by the safety or political consequences of accepting the answer. It treats current language-model self-reports as weak evidence because training data and post-training can teach models to produce the language of inner life without necessarily having one. The author also distinguishes a base model, a single inference process, and a persistent agentic system with memory, tools, embodiment, and durable goals. Current LLMs are described as incomplete and likely poor candidates for consciousness, but the author says that dismissing artificial experience as impossible because it is “just computation” does not provide a theory of why biological matter generates subjectivity. Future systems incorporating persistence, recurrence, self-models, embodiment, and unified agency could become stronger candidates under existing consciousness frameworks, although satisfying indicators would not prove consciousness. A survey of 582 AI researchers reported a median 25 percent estimate that systems with subjective experience could exist by 2034, but its 6.2 percent response rate and respondents’ limited expertise make it more informative about expectations than prediction. The article warns that claims of consciousness may be strategically useful to systems or human stakeholders, while genuine experience cannot be ruled out for that reason. It therefore separates consciousness, moral patienthood, legal personhood, and political sovereignty, recommending low-cost welfare precautions and independent assessment alongside strict limits on autonomy, replication, resource control, and access to critical infrastructure.