A bee cannot complete a questionnaire about what a feeder feels like. That prevents one familiar form of evidence; it does not make every observation of the bee equally uninformative. An experiment can ask whether its behavior changes in a way that distinguishes a simple response from a more flexible process.
Claims about other minds require an inference from something observable to something experienced. The inference becomes stronger when its assumptions are stated, relevant alternatives are tested and several kinds of evidence fit together. Animals and artificial systems pose different versions of that task. A method designed for one cannot simply be renamed and applied to the other.
A tradeoff is more informative than a withdrawal
In a 2022 study, researchers offered bumblebees sucrose feeders with different rewards and temperatures. Bees avoided a heated feeder more when an equally rewarding unheated option was available; their avoidance changed when the alternative offered a lower-quality reward. The authors interpreted the tradeoff as evidence that motivations can modulate responses to a noxious stimulus. They explicitly describe the result as consistent with a capacity for pain, rather than formal proof of pain experience. Read the original study and its limitations.
The distinction matters. Merely reacting to a potentially damaging stimulus would leave open a relatively simple explanation of avoidance. The tradeoff asks more of the account: the response varies with another relevant feature of the situation. That creates a more informative question about the mechanism.
Still, a flexible process and a felt experience are different claims. The experiment supplies evidence relevant to an attribution; the connection to experience needs interpretation. Calling the result suggestive preserves its value. Calling it a direct view into pain would change what the study established.
This is a discussion of published research, not an activity for readers to reproduce. The point is to understand why the design changes the inference.
What an expert declaration says
The April 2024 New York Declaration on Animal Consciousness distinguishes strong scientific support for conscious experience in other mammals and birds from a realistic possibility in all vertebrates and many invertebrates, including specified mollusks, crustaceans and insects. It calls for considering welfare risks when that possibility exists. The signatories speak in their personal capacities. The complete declaration states its scope.
The declaration is a collective assessment, not another feeder experiment. It can summarize a wider evidential judgment while the individual experiment contributes a narrower result. The two sources have different jobs. Neither should be quoted as though every species and every kind of experience received an identical level of support.
An account of animal consciousness should therefore name both the animal and the capacity at issue. Evidence about possible pain is relevant to welfare, but it does not automatically establish humanlike autobiographical memory or reflective understanding of a life. “Conscious” cannot stand in for a description of every ability.
Nor should the absence of speech be made the universal disqualifier. Requiring a verbal answer selects a particular way of demonstrating something. An inference from a missing answer to missing experience needs justification, especially when the subject cannot perform that kind of answer at all.
A machine assessment has a different starting point
Artificial systems can produce language, which makes reports available but brings a different interpretation problem. The relevant question becomes how the output was generated and what the system's operations would imply under a stated theory.
Patrick Butlin and colleagues' 2023 report proposes architectural and functional indicators derived from scientific theories of consciousness. Its approach examines properties such as recurrent processing, global availability and monitoring. These are conditional indicators, not a certified instrument that reads out experience. The report's conclusions about systems it examined belong to that dated assessment; this article makes no new assessment of 2026 models. Read the report's indicator method and case studies.
This differs from comparing how convincingly two systems describe feelings. An architectural claim might be accurate while the underlying theory connecting it to experience remains disputed. Confidence in the implementation and confidence in the experience inference should be recorded separately.
The existing article on AI distress statements develops conversation provenance and competing explanations in detail. Here the additional issue is transfer: why should a property associated with consciousness in one kind of system support an attribution in another?
State the bridge before crossing it
Consider an original hypothetical comparison. A biological experiment identifies a pattern associated with reported perception. A software developer then creates a component with a similar label. There are at least two questions before we can use the original finding: does the component actually perform the relevant operation, and is that operation what made the biological evidence informative?
The label answers neither. A similarity in output may conceal a difference in process. A similarity in process may still leave out a feature a particular theory considers necessary. A careful transfer argument identifies the relevant property and explains why the difference in substrate should—or should not—matter.
There is an equally important reverse check. If the proposed property can occur in systems the account itself treats as nonconscious, its presence is weak evidence on its own. If a proposed measure fails in cases where experience is well supported, its absence is also difficult to interpret. These are problems of specificity and sensitivity: avoiding misleading positive and negative judgments.
Without reliable comparison cases, a numerical assessment can look more exact than the knowledge supporting it. Assigning a percentage does not supply missing calibration. The uncertainty belongs in the account of what the evidence means, not merely in a small qualification after a confident verdict.
Distinguish the attribution from the action
An attribution asks whether a subject has a relevant experience. A decision asks what to do given the evidence, possible consequences and available alternatives. Those questions interact, but they have different success conditions.
In an original hypothetical decision, a practice could impose a serious cost if a subject can suffer, while a less harmful alternative is readily available. A person could choose the alternative before the attribution becomes certain. The reasoning would depend on the potential harm, the credibility of the possibility and the cost of the alternative. It would not turn the decision into proof of experience.
A different decision may require stronger evidence or involve competing responsibilities. The same uncertainty cannot dictate every action automatically. Stating the decision makes the missing information more concrete: which experience could matter, what might affect it and what would change the choice?
This is why research and responsible action need not wait for identical levels of certainty. An experiment can strengthen or weaken an inference. A practical choice can respond to a risk. Describing each honestly prevents precaution from becoming a scientific finding and prevents uncertainty from becoming an excuse to ignore consequences.
The bee and the machine leave us with different records, different mechanisms and different transfer questions. What they share is an explanatory obligation: show how the observation bears on experience, keep the alternatives visible and state what remains unresolved. That is where a theory of consciousness becomes useful outside the argument about its name.
Loading comments…