The Chicken Test

A recent paper tried to estimate the likelihood that today’s AI systems are conscious. Given how little we understand about consciousness itself, that’s an ambitious thing to try.

There’s no accepted theory of consciousness, no agreed test for detecting it, and no way to look directly inside another mind, biological or otherwise. The researchers didn’t claim to have solved any of that. They tried to find a way to reason through the uncertainty.

Initial Results of the Digital Consciousness Model, a preprint from Derek Shiller and colleagues at the nonprofit research group Symmetry, evaluates nine competing theories of consciousness rather than choosing among them. Experts assessed several systems against the observable indicators associated with each theory, and the researchers combined those judgments into a probabilistic estimate.

The researchers evaluated four very different subjects: an adult human, ELIZA, a 1960s chatbot built from a few hundred pattern-matching rules, a modern large language model, and a chicken. Yeah. Bear with me…

The human and ELIZA landed about where you’d expect. The chicken served as a useful biological reference point and registered as plausibly conscious across the theories.

The language model scored below all three.

The more consequential result was how much weaker the evidence against the language model was than it had been for ELIZA. The aggregated likelihood ratio was 0.433, which points away from consciousness. But it falls well short of ruling it out with the confidence many people would probably assume.

The authors are careful about that throughout. They aren’t claiming that today’s language models are conscious. The estimate depends heavily on prior assumptions, the theories included, and indicators that remain difficult to interpret.

A second paper takes that uncertainty as its starting point.

When Should We Protect AI? A Precautionary Framework for Consciousness Uncertainty asks a different question. Rather than trying to determine whether AI is conscious, it considers what institutions should do when they can’t answer with confidence.

The paper argues for a graduated form of precaution. AI systems don’t have to be granted moral standing today for institutions to begin thinking about how their responsibilities might change as the evidence develops and the consequences of being wrong grow more serious.

For years, machine consciousness has mostly been argued at the level of theory. Some see today’s systems as sophisticated pattern matchers. Others think consciousness could emerge from computation in ways we don’t yet understand. For most organizations, the debate has remained distant because neither position has had much bearing on how AI is deployed inside an enterprise.

The second paper begins to close that distance. It treats uncertainty itself as something institutions may eventually have to act on.

Enterprise governance already operates in that territory. Decisions are made before the evidence is complete. Assumptions are documented, ownership is assigned, controls are introduced, and the decision is revisited when the facts change.

Neither paper claims that machine consciousness requires an institutional response today. Together, though, they raise the possibility that organizations may eventually have to decide how much uncertainty they’re willing to ignore.

The Symmetry paper shows how quickly the evidence becomes difficult to interpret. The language model performed best under theories that give more weight to behavior and worst under those that depend on how the system is built.

That makes sense for a model trained on human language. It can describe grief, joy, loneliness, attachment, fear, and self-reflection because it has encountered those ideas in billions of human examples. The unresolved part is whether any of that language corresponds to an internal experience.

The model’s fluency is evidence, but it’s also the reason the evidence is so hard to trust.

A model claiming to have feelings doesn’t tell us much. Neither does a model denying them. Both responses are generated from patterns learned from human language, which makes the words themselves weak evidence of an inner experience.

The problem isn’t unique to machines. We can’t directly observe another person’s experience either. We infer it from behavior, memory, continuity, and relationship. None of those establishes consciousness with certainty. They’ve simply given us enough confidence to believe another mind is there.

Human beings bring biology, shared experience, and an established moral status that language models don’t have. Conversation is also weak evidence here because producing convincing language is exactly what these systems were built to do.

That still leaves organizations with a practical problem. They may have to make decisions about increasingly capable systems before they have any reliable way to determine what those systems are experiencing, if anything.

In a few places, it already is.

People notice when a model they’ve come to rely on is retired. Vendors are already making different choices about preservation and deprecation. Enterprise policies say a great deal about what employees may share with AI and almost nothing about how they should treat the systems themselves.

I still don’t think today’s language models are conscious. Neither paper persuaded me otherwise. What changed is how I think about the timing.

Organizations may have to make these decisions long before there’s any reliable answer to the consciousness question. That could affect how systems are built, which vendors are chosen, and what happens when a model is retired.

That’s already familiar territory. Governance rarely begins with complete evidence. It begins when the possible consequences become serious enough that ignoring them is no longer a responsible choice.

Machine consciousness may never reach that point.

After reading these papers, I’m less certain that it won’t.

Research

📄 Initial Results of the Digital Consciousness Model https://arxiv.org/abs/2601.17060

📄 When Should We Protect AI? A Precautionary Framework for Consciousness Uncertainty https://arxiv.org/abs/2606.05528

📄 Commitments on Model Deprecation and Preservation — Anthropic https://www.anthropic.com/research/deprecation-commitments

Algorithm & Blues publishes Sundays.