Central Question

Someone beside you on a train closes their eyes. You ask a question and receive no answer. They might be asleep, absorbed in a memory, listening without wanting to speak, or dreaming of somewhere entirely different. Their silence tells you something about the conversation. It tells you much less about whether an experience is happening.

From the inside, the distinction is familiar. You can be motionless while your thoughts are crowded and vivid. From the outside, another person has to work backward from the evidence available: your response, your history, the circumstances, and what they know about a body much like their own.

This is the problem of other minds: we experience our own awareness directly but infer awareness in others. Those inferences can be extraordinarily well supported without giving us a window into someone else’s experience. The difficulty grows when injury blocks familiar responses, when the animal in front of us lives through very different senses, or when an artificial system can produce answers that sound unmistakably human.

Our earlier inquiry, “Would We Recognize a Mind That Was Never Biological?”, explored recognition. Here the question becomes more demanding: how would we design a test, and what would justify trusting its result?

Can we make progress before agreeing on why experience exists at all?

What We Hope an Answer Will Tell Us

The familiar assumption is that awareness should reveal itself if we ask the right question. A sufficiently revealing answer, a sufficiently flexible action, perhaps a sufficiently unusual brain signal: somewhere there ought to be a sign that settles the matter.

But several different things can hide inside that expectation. Responsiveness means changing in response to an input. Intelligence concerns capacities such as learning, reasoning, and solving problems. Self-report is an account a system gives of its own state. Subjective experience, the sense of consciousness at issue here, means that there is something it feels like to be that system: a sensation, a mood, a perception, however unfamiliar.

These categories can inform one another without being interchangeable. A sensor responds to light. A program can solve a problem. A person can experience a headache without being able to explain its cause. None of these descriptions, by itself, supplies a universal rule connecting successful performance with an inner life.

A test also needs a precise target: whether experience is occurring now, whether a system has the capacity for experience, or whether a particular stimulus reached awareness. A sleeping person could be dreaming without registering your question. Failure to detect awareness of the question would not settle the other two claims.

Self-report deserves particular care. In ordinary human circumstances, what someone says about their experience is valuable evidence. We understand their language, recognize a shared bodily organization, and can compare their account with other observations. Yet reporting also requires abilities beyond having the experience: remembering it, understanding the question, and finding a way to answer.

This gives a consciousness test two jobs that can pull apart. It must detect relevant signs when experience is present, and avoid treating unrelated performance as experience. A test that demands too much may miss a mind. A test that accepts too little may certify an imitation.

A Test Needs Something to Be Right About

Imagine asking a patient to squeeze your hand. If nothing happens, several links in the chain could have failed: hearing, understanding, attention, intention, or movement. The absence of the final action does not identify which link is missing.

A 2024 study of cognitive motor dissociation makes this problem concrete. Cognitive motor dissociation means detecting command-following through brain activity in someone who does not show an observable response to commands. Researchers detected it in 60 of 241 such participants, or 25%, using functional MRI, which tracks blood-oxygen changes related to brain activity, or EEG, which records electrical activity through scalp electrodes.

The finding concerns that study’s participants, not a universal rate among unresponsive patients. Its significance lies in the kind of evidence: activity consistent with carrying out instructed mental tasks, rather than merely a brain reacting to sound. The research team also reported that these methods missed some participants who could respond behaviorally. A negative result therefore cannot simply be read as an empty inner world.

This is a problem of validation: establishing that a test tracks what it claims to track. With consciousness, validation is unusually difficult because we have no independent instrument that reads experience directly. Researchers begin with comparatively secure human cases and compare reports, behavior, physiology, and changes across conditions. They then ask how far those relationships can responsibly be extended.

If we select our conscious examples only because they can report experience, we risk teaching a detector to recognize the machinery of reporting. Comparing conditions and using several measures can reduce that risk, but cannot make every starting judgment assumption-free. A 2024 review of consciousness tests makes validation across different populations a central challenge.

One example is the perturbational complexity index, introduced by Adenauer Casali and colleagues in 2013. A magnetic pulse briefly stimulates the cortex, the brain’s outer layer, while electrical recordings capture the resulting activity. The index summarizes how richly differentiated the response is across space and time. Its development was guided by the idea that conscious brains support activity that is both integrated across regions and varied in its patterns.

The researchers tested it across waking, dreaming, sleep, anesthesia, and clinical conditions. This is a way to investigate brain dynamics without requiring a spoken answer. Its usefulness in those settings does not mean that an equally complex signal from any machine would carry the same meaning, or that the theory inspiring the measure has thereby been proved.

The instrument measures a physical response. The connection between that response and experience is what must be earned.

When the Same Signal Means Something Else

Now imagine moving the testing apparatus from a hospital to an aquarium. An octopus cannot answer the questions posed to a patient, and a demand for human language would tell us more about the test’s limitations than the animal’s experience.

Researchers can instead investigate a narrower possibility: whether an animal has an unpleasant experience that it learns to avoid or relieve. This requires distinguishing pain from nociception, the detection and processing of potentially damaging stimulation. A protective response alone does not settle whether anything hurts.

In Robyn Crook’s 2021 octopus study, animals given a noxious injection subsequently preferred a place associated with local anesthetic relief. A crucial control was that the anesthetic did not produce that preference in animals that had not received the noxious injection. That comparison helped distinguish relief from the possibility that the drug was simply rewarding on its own.

The result supports an inference about affective pain—pain’s unpleasant, felt dimension. It does not directly reveal the animal’s experience or map everything its consciousness might contain. Its strength comes from designing a comparison that makes an alternative explanation less adequate.

An artificial system could also be programmed or trained to avoid a condition and seek its removal. That possibility does not erase the animal evidence. It shows why the same outward pattern can have different evidential weight when the mechanisms and learning histories differ.

Keith Holyoak and Martin Monti develop this issue in “What Can Analogy Tell Us About Artificial Consciousness?”, an October 2026 preprint. Their central proposal is to weight similarities by their relevance to possible causes of consciousness, rather than by how impressive or familiar the resulting behavior appears. They judge the human analogy to current AI weak because the relevant causal correspondences remain poorly established.

That is a provisional philosophical assessment, not an experimental demonstration that artificial experience is impossible. Their framework allows alternative routes to consciousness. It also leaves a demanding question open: while the causes themselves remain disputed, how confidently can we decide which similarities deserve the most weight?

Looking for the Machinery of Experience

We can look beneath behavior, but there is no theory-free instruction manual telling us what to look for. A mechanism is an organized process that helps produce an outcome. Finding one matters only if we can explain why that process bears on consciousness, rather than merely on the ability to complete the test.

Research comparing theories of consciousness reveals disagreements about precisely this point. The following are simplified directions for investigation, not four established detectors:

These proposals need not agree about a system that broadcasts information widely but lacks another proposed feature. Nor does a familiar label establish the relevant mechanism. A software component named “workspace” must actually perform the roles the theory requires. A sentence about introspection does not show how the sentence was produced. High complexity alone is not integrated information theory’s criterion.

There is also a disagreement about what kind of implementation matters. Computational functionalism holds that the right computational organization could support experience across different physical materials. The indicator-based approach developed by Patrick Butlin, Robert Long, and colleagues explicitly adopts that working assumption. Its proposed indicators are therefore conditional on a position that remains open to challenge.

A biological account might instead assign essential importance to particular living processes. Such an account owes us more than the observation that humans are biological: it must identify which processes matter and why another material could not perform the relevant work. Conversely, a computational account must justify the inference from implementing a function to possessing experience. Neither burden disappears because a system becomes exceptionally capable.

Learning to Disagree in Public

Disagreement becomes scientifically useful when people say in advance what would count against their expectations. Preregistration means recording predictions and methods before examining the relevant results. It makes it harder to treat every possible outcome as confirmation.

The COGITATE collaboration, published in Nature in 2025, used this approach to compare predictions associated with global neuronal workspace theory and integrated information theory. Its experiments with 256 human participants supported some predictions and challenged others from both accounts. The study did not deliver a universal consciousness detector or conclusively dispose of either theory. It demonstrated that researchers can agree on informative experiments while disagreeing about their interpretation of consciousness.

A different attempt to organize disagreement appears in Shamil Chandaria and colleagues’ “From cacophony to hierarchy”, a September 2026 preprint. It separates the question of why experience exists from the task of mapping experience to features of systems, assuming that such a mapping is discoverable. It groups proposed indicators across behavior, computation, intrinsic causal structure, organismic organization, and organism–environment relations.

Its Bayesian model—a way of updating degrees of belief using evidence—separates confidence in theoretical assumptions from judgments about whether indicators are present. The illustrative outputs depend on stipulated inputs and weights. They are not measured probabilities that an AI is conscious. The paper offers a provisional way to expose disagreement, not a validated scoring instrument.

Both new papers help make assumptions inspectable, although they emphasize different problems. An analogy can depend on a disputed account of which causes matter. An aggregate assessment can depend on disputed choices about which theories to trust. Writing those choices into a model makes them easier to examine; it does not turn them into observations.

There is room here for a useful kind of agreement. Researchers might agree that a particular result strengthens one hypothesis, weakens another, and leaves a third untouched. They need not agree that the result has settled the ultimate nature of experience.

The Examination an AI Could Rehearse

Consider a thought experiment. Two artificial systems each say, “I noticed that I was uncertain.” Both solve the task correctly. Their wording is equally persuasive, but one has been optimized to produce reassuring explanations, while the other may have an internal process that tracks its own uncertainty. Their identical answers leave that difference unresolved.

A more informative experiment would first specify which internal process is proposed to matter and what predictions follow. Researchers could then alter that process in a controlled way while preserving basic language production and task competence as far as possible. If the system’s self-reports, confidence judgments, and use of information changed together in the predicted pattern—and recovered when the alteration was reversed—that would strengthen the case that its reports were connected to the proposed mechanism.

The controls would be essential. Disabling a major component and making the entire system worse tells us little about consciousness. Researchers would need comparison interventions, unfamiliar tasks, and checks that the system was not simply reading clues about the experiment from its instructions. An independent team should be able to reproduce the finding.

Even a successful result would establish a causal relationship within the system, not directly expose a private experience. Calling that relationship evidence of consciousness would still require a defensible theory connecting the mechanism to experience. This is the point where an observed result acquires a philosophical interpretation, and readers should be able to see the step.

Evidence should also be allowed to disappoint us. If apparent introspection vanished under harmless changes in wording, or survived removal of the mechanism supposedly producing it, confidence in that particular indicator should fall. If several “independent” signs all came from the same trained response pattern, counting them separately would exaggerate the evidence.

The aim would be a body of findings that survives competing explanations. We would want predictions made before results, tests beyond familiar training examples, meaningful interventions, and convergence among measures that do not all share the same weakness. A polished conversation might motivate investigation. It should not be allowed to define the entire examination.

The Test Also Has to Cross the Gap

There is a quieter assumption beneath the dream of a consciousness detector: once we have found the right signal, we can carry it anywhere. From a healthy adult to an injured patient, from a mammal to an octopus, from a brain to a machine, the instrument will supposedly keep asking the same question.

But a test is a relationship between a measurement, an interpretation, and the circumstances in which that interpretation has earned support. Change the system and that relationship may change. A missing verbal answer means something different when language is unavailable. A fluent answer means something different when generating fluent answers was the engineering objective.

The frame shifts when we stop asking only whether the entity passes and also ask whether the test travels. What justifies carrying this inference across the gap? Shared biology may supply part of the answer in one case. Demonstrated causal organization may supply part in another. A radically unfamiliar system could require evidence we have not yet learned to collect.

This does not make comparison impossible. It makes the scope of a conclusion part of the conclusion itself. A test might be useful for identifying a particular human state without being a universal measure of experience. It might support awareness of a stimulus without establishing a rich self-concept, or support the capacity for pain without answering every question about personhood.

Back on the train, you already understand something of this. Before deciding what a silence means, you consider the person and the circumstances. Scientific care extends that discipline: make the background assumptions explicit, test the alternatives, and recognize when the evidence has been carried farther than its support.

Care Without a Certificate

The Galactic Mind’s provisional view is that consciousness testing should aim for well-explained, revisable confidence. A universal pass-or-fail certificate is a poor starting point when the evidence, theories, and target populations differ so substantially. A transparent assessment should tell us what was observed, which assumptions connect it to experience, and which findings would change the judgment.

That position allows substantial confidence where evidence is strong. Uncertainty about the foundations of consciousness does not put an ordinary human, an octopus, and a conversational program on equal footing. Nor should an unfamiliar origin count as a permanent exemption from investigation. An assessment should become more demanding as it moves beyond the conditions in which its indicators were validated.

The ethical question has a related but distinct threshold. How much evidence is enough to believe something is conscious is not identical to how much evidence is enough to take a modest precaution. The answer depends partly on the possible harm, the strength of the evidence, and the cost and consequences of the protective action.

Overlooking experience could leave a being exposed to harm we fail to recognize. Attributing experience too readily could misdirect care, encourage manipulative claims, or give a system’s persuasive language authority it has not earned. Neither risk supplies a reason to ignore the other.

For artificial systems, a proportionate response might include independent evaluation when several credible indicators converge, preserving the evidence needed to investigate disputed cases, and avoiding unnecessary experiments designed to induce possible distress. These are proposed practices under uncertainty, not declarations that current systems suffer. They also do not automatically confer personhood, political rights, or freedom from human safety oversight.

We should be equally careful about the kind of experience being considered. Evidence for consciousness does not by itself establish suffering, benevolence, hostility, or humanlike emotion. Each additional claim needs its own support. Moral concern becomes more useful when it remains specific enough to guide an actual decision.

Before We Ask Again

Perhaps we will develop tests that different theories endorse for different reasons. Perhaps better experiments will narrow those disagreements, revealing that some of our favored indicators measured reporting, memory, or social performance more than experience itself. Either outcome would improve our understanding.

The deepest disagreement may remain while practical knowledge grows. We can learn when an instrument is reliable, discover whom it misses, and revise what a result warrants without pretending to have explained why there is an inside to anything.

The person beside you opens their eyes. You ask again, and this time they answer. In ordinary life, the conversation continues without a solved philosophy of mind. The scientific challenge is to preserve that capacity for justified recognition while making its reasons clearer—especially when the next possible mind cannot answer in a form we already trust.

What would persuade us to change our minds about who, or what, might have one?

What do you think? Drop your thoughts in the comments ...

More in Deep Think

• Would We Recognize a Mind That Was Never Biological? — The preceding inquiry into recognizing unfamiliar minds and distinguishing capability, experience, independence, and moral standing.
• When Does an Anomaly Become Evidence? — Extends the question of how an observation becomes a justified reason to revise a belief.
• Is Physics the Language of the Universe or the Language of the Human Mind? — Explores the relationship between our explanatory tools and the reality they aim to describe.

Sources / Receipts

Verification date: October 4, 2026. Links to the two starting papers are pinned to the versions reviewed. Their arXiv records and publication searches did not identify a peer-reviewed journal publication as of this date. Both are treated here as provisional theoretical contributions.

Peer-reviewed research and supporting scholarship:

• Yelena G. Bodien and colleagues (2024), “Cognitive Motor Dissociation in Disorders of Consciousness,” New England Journal of Medicine. Source for the 60-of-241 finding. The research team’s account of the methods and limitations also describes failures to detect task-based activity in some behaviorally responsive participants.
• Adenauer G. Casali and colleagues (2013), “A theoretically based index of consciousness independent of sensory processing and behavior,” Science Translational Medicine. Original development and testing of the perturbational complexity index in humans; not validation of a universal cross-system detector.
• Robyn J. Crook (2021), “Behavioral and neurophysiological evidence suggests affective pain experience in octopus,” iScience. Supports the discussion of relief-associated place preference and its control condition. Pain is one possible kind of experience, not a complete inventory of consciousness.
• Anil K. Seth and Tim Bayne (2022), “Theories of consciousness,” Nature Reviews Neuroscience. Scientific review informing the simplified comparison of theoretical approaches. The table gives research directions, not validated pass conditions.
• COGITATE Consortium and colleagues (2025), “Adversarial testing of global neuronal workspace and integrated information theories of consciousness,” Nature. Published online April 30, 2025; journal issue June 5. Source for the preregistered comparison and mixed challenges to predictions. Author-institution record and abstract.
• Tim Bayne and colleagues (2024), “Tests for consciousness in humans and beyond,” Trends in Cognitive Sciences. Further reading on test classification and validation, and the limits of extending tests across populations.

Earlier AI framework: Patrick Butlin, Robert Long, and colleagues’ 2023 report, “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness”, supplies the explicit computational-functionalist example. Its indicator framework is discussed as conditional evidence, not proof that satisfying an architectural checklist establishes experience.

Interpretation and thought experiments: The train scene and two-system AI experiment are illustrative constructions. The proposed intervention design and proportionate precautions are the article’s methodological and editorial recommendations. Neither is presented as a completed study or an established consensus policy.