A Question With No Test: Studying AI Consciousness Anyway | Noa Weiss (AI Consciousness Researcher)

Noa Weiss on the four pillars of evidence for AI consciousness, and why the question can be researched before the hard problem is solved. There is no test for consciousness. No single experiment settles whether a system has felt experience, which is why the question is often set aside as unanswerable. Weiss argues the opposite: that a body of evidence can be assembled from several independent directions, and that it already adds up to more than the sum of its parts. She organizes the field into four pillars. Mechanistic interpretability looks inside the model, covering Anthropic's work on emotion-related vectors and Jack Lindsey's finding that models sometimes detect injected thoughts. Computational neuroscience compares what we know about mammalian brains to what appears in LLMs, including how reward and punishment are encoded. Machine behavior applies the methods used to infer sentience in animals. Theory auditing works from the indicator approach introduced by Patrick Butlin and Robert Long, which tests systems against existing theories of consciousness rather than waiting for one to win. Weiss makes the case that computational neuroscience is the neglected pillar and the most informative, because mimicry cannot reach it: a model having read the entire human corpus does not change how it encodes reward against punishment. She also sets out why this matters on safety grounds and not only ethical ones, and why waiting for certainty is not available as an option while the public debate is already underway. The talk opens and closes with the same audience poll on the probability that some AI system is already conscious. Chapters 0:00 Audience poll: what is your P(AI consciousness)? 1:24 Why there is no test for consciousness 2:02 About Noa Weiss 3:02 Phenomenal consciousness and "What Is It Like to Be a Bat?" 5:27 Valence, and where moral patienthood comes from 6:34 How we infer consciousness in other minds 11:12 Why it matters: ethics and safety 12:52 Why it matters now, not after certainty 14:34 The four pillars of evidence 15:13 Pillar 1: mechanistic interpretability 18:02 What a model says about itself 20:53 Pillar 2: computational neuroscience 24:11 Pillar 3: machine behavior 25:59 Pillar 4: theory auditing and the indicator approach 30:36 What it all adds up to 31:05 Why computational neuroscience is the neglected pillar 33:17 Second poll: did your number move? 33:57 Takeaway: AI consciousness is researchable 34:27 Q&A Noa Weiss argues that AI consciousness is researchable now, because four independent lines of evidence can be built without first solving the hard problem or settling on a single theory of consciousness. More AI safety research: https://far.ai FAR.AI is a research nonprofit working to ensure the safe development of advanced AI. We host the Alignment Workshop series and publish frontier alignment research.