Anders begins by separating several meanings of consciousness, including awareness of internal and external states, subjective sensation, volition and a continuing sense of self. This matters because evidence for one functional property does not automatically establish the full human experience implied by the word conscious.
Anthropic's work tests whether information can enter a shared internal workspace and become available to multiple model processes. The results provide evidence for conscious access as a computational function, but the researchers explicitly distinguish that from phenomenal consciousness, the felt experience that remains difficult to explain even in humans.
Language models also differ from people in important ways. Their activity is driven by token processing, they do not maintain an uninterrupted stream of experience between prompts, and ordinary sessions lack the persistent memory needed for a stable personal history. Model statements about their own consciousness are therefore interesting evidence to analyze, not reliable self-verification.
Anders concludes that the research raises better questions rather than settling the debate. More persistent memory, multiple interacting processes and richer system architecture may create additional consciousness-like functions, but the current evidence does not justify claiming that language models have human-like subjective awareness.
Watch on YouTube
