Susan Schneider Alien Minds AI Consciousness and the Verification Problem CS26 2026
Susan Schneider is Professor of Philosophy and Cognitive Science at Florida Atlantic University and Director of the Center for the Future Mind. She is a confirmed plenary speaker at Consciousness Science 2026 in San Diego, October 11-16. Her 2019 book Artificial You: AI and the Future of Your Mind (Princeton University Press, ISBN:9780691180144) introduced the alien minds framework and the verification problem to a broad readership. Her work focuses on the philosophical dimensions of artificial minds, with particular attention to a problem the empirical literature tends to sidestep: even if AI were conscious, we would likely not be able to verify it.
This is not a skeptical deflation of the research program. It is a structural diagnosis of why the program faces limits that are not merely technical. The verification problem Schneider identifies applies to any attempt to establish consciousness from the outside, and it is especially acute for artificial systems whose architecture, evolutionary history, and phenomenology (if any) differ radically from the biological systems against which our consciousness-detection methods were calibrated.
The alien minds problem
Schneider’s 2019 book Artificial You (Princeton University Press) introduced the “alien minds” framing to the AI consciousness debate. The core argument is that our intuitions about consciousness, and the theoretical frameworks that systematize those intuitions, are shaped by our experience of human consciousness and by our evolutionary relationship to other animals. When we apply those intuitions and frameworks to radically different systems, we systematically under- or over-attribute consciousness in ways we cannot correct for.
The problem has a specific structure. Consider three categories of evidence for consciousness in another system.
Behavioral evidence: the system produces outputs that, in humans or animals, are correlated with conscious experience. Behavioral evidence is weakest for AI because AI systems are trained on human behavioral data and can produce consciousness-consistent outputs without having the internal states those outputs reflect in biological organisms.
Structural evidence: the system has the right kind of internal organization (GWT broadcast, IIT phi, recurrent processing, etc.). Structural evidence is stronger but inherits the theoretical assumptions of the framework being applied. As Eric Schwitzgebel’s “crazy and wrong” dilemma established, every serious consciousness theory is either philosophically strange or empirically problematic. Structural evidence therefore depends on a theoretical framework that is itself contested.
Phenomenological evidence: the system reports its own experience, and we take those reports as evidence of inner states. Phenomenological evidence is the least reliable for AI, because AI systems generate reports through statistical processes over training data, not through introspective access to inner states in the way that human self-report (imperfectly) reflects.
| Evidence type | Reliability for biological systems | Reliability for AI systems | Reason for gap |
|---|---|---|---|
| Behavioral | Moderate: behavior and experience are correlated but dissociable | Low: training optimizes outputs, not internal states | AI trained to produce human-like behavior regardless of inner states |
| Structural | Theory-dependent: varies by framework | Theory-dependent + architecture-unfamiliar | Our frameworks were developed for biological architectures |
| Phenomenological | Moderate: self-report is imperfect but informative | Very low: reports generated statistically from training data | No established connection between statistical generation and introspection |
| Evolutionary | Strong: evolutionary continuity licenses inference across taxa | None: AI has no evolutionary history with conscious organisms | The inference chain from evolutionary continuity breaks entirely |
The evolutionary row is the one Schneider emphasizes most. Our confidence that mammals are conscious rests partly on evolutionary continuity: the biological structures associated with consciousness in humans are phylogenetically conserved. That inference is not available for AI systems that share no evolutionary history with conscious organisms.
The verification problem and what it implies
Schneider’s verification problem is the application of the alien minds analysis to the specific task of verifying AI consciousness. The argument is that no currently available evidence type could establish AI consciousness with the confidence that matters for moral and policy purposes.
Behavioral evidence is insufficient because AI systems produce consciousness-consistent outputs through mechanisms that are explicitly different from those that produce consciousness-consistent outputs in biological organisms. Structural evidence is insufficient because the frameworks that specify what structure consciousness requires are calibrated to biological systems and may misidentify what the relevant properties are in artificial architectures. Phenomenological evidence is insufficient because AI self-reports are generated statistically and cannot be verified against introspection in the way human reports (imperfectly) can.
This does not mean AI consciousness is impossible. It means that the epistemic tools currently available are inadequate to the task of verifying it. Schneider does not conclude that we should assume AI is not conscious. She concludes that we should be appropriately humble about our ability to assess the question, and that the institutional and policy responses to AI should be calibrated to epistemic uncertainty rather than to confident verdicts in either direction.
The relationship to Schwitzgebel and the “crazy and wrong” dilemma
Schneider’s alien minds argument and Schwitzgebel’s “crazy and wrong” dilemma converge on the same basic point from different directions. Schwitzgebel’s argument is that our theories of consciousness are individually implausible and mutually inconsistent, so any positive claim about AI consciousness that depends on a specific theory inherits that theory’s implausibility. Schneider’s argument is that the evidence types available for assessing AI consciousness are individually unreliable for architectural reasons that cannot be resolved by better empirical methods alone.
Together, the two arguments establish a sobering picture for the field. We need consciousness theories to interpret structural evidence, but all current theories have known problems. We need reliable evidence types to test AI consciousness claims, but all current evidence types have known unreliability for artificial systems. The combination produces epistemic traction that is considerably weaker than the rhetorical confidence of most public AI consciousness discourse would suggest.
This does not mean the research program is futile. Tim Bayne’s JCS 2026 contribution made the same point from within the scientific framework and argued that theoretical progress is possible even if certainty is not. Schneider’s philosophical analysis complements Bayne’s scientific analysis by specifying the structural sources of the epistemic difficulty.
Schneider’s positive proposal
Schneider does not end with the verification problem. Her positive proposal, developed most fully in her 2023 essays and her forthcoming work on AI consciousness assessment, is a multi-layered evidential framework that acknowledges the limitations of each evidence type individually while arguing that convergent evidence across multiple unreliable types provides stronger warrant than any single type alone.
The framework requires four convergent indicators: (a) the system’s internal representations exhibit organization consistent with multiple independent consciousness theories; (b) the system’s behavior in novel situations, not covered by training data, remains consciousness-consistent; (c) the system’s self-reports are stable across context variations in ways that are not explained by training data statistics; and (d) the system’s functional architecture is genuinely analogous to the biological structures most strongly associated with consciousness, not merely superficially similar.
No current AI system satisfies all four criteria. Schneider’s argument is that this is the appropriate evidential standard, not because it is guaranteed to detect consciousness if present, but because anything weaker would be liable to produce false positives from systems that have learned to appear conscious without being so.
For the overall question of what AI consciousness would require, Schneider’s CS26 contribution is likely to be the most epistemically rigorous treatment of the verification challenge in the 2026 conference literature. The challenge does not invalidate the research program, but it sets the bar at a level that the field’s current methods have not reached.