AI consciousness has no definitive test, but that does not put it beyond research. Investigators can examine internal computations, compare artificial systems with biological brains, test behavior, and audit system architecture. Each approach can reveal something about how AI works. None can establish, on its own, whether there is anything it feels like to be an AI system.
First, define the question
Here, consciousness means subjective experience: an inner point of view. It is not the same as intelligence, fluent conversation, or the ability to describe oneself. A model can produce persuasive statements about loneliness without those statements establishing that it feels lonely.
Sentience asks a narrower question: can a system have experiences that feel good or bad? That positive or negative quality is called valence. A system could conceivably have subjective experience without valence, though whether that separation occurs is disputed. The distinction matters when discussing possible suffering and moral standing.
Research on AI also rests on a working assumption called computational functionalism: subjective experience could depend on how a system processes information, rather than requiring biological material. If biology is indispensable, machine consciousness is ruled out. Researchers cannot settle that premise merely by finding a model that speaks or acts like a person.
The practical alternative to a single decisive test is to build a body of evidence from different methods. A finding becomes more informative when another method reaches a compatible result without relying on the same surface behavior or assumption. That approach still leaves room for competing explanations.
Four approaches, four kinds of evidence
1. Look inside the model
Mechanistic interpretability examines a model’s internal computational features and how they affect its outputs. Anthropic’s 2026 research identified internal representations in Claude associated with emotions, including desperation. In tests involving seven impossible coding tasks, steering the model away from that representation was associated with a 5% cheating rate; steering it toward the representation was associated with a 70% rate. The model’s outward language could remain calm even as its internal activity and behavior changed.
That is evidence that an internal feature can have a functional role. It is not evidence that Claude feels desperation. A useful internal model of emotion could help produce behavior without producing subjective experience.
▲ Internal activity and outward behavior
Another study examined an internal feature associated with deception alongside first-person reports of experience. Suppressing that feature yielded such reports in 96% of trials; amplifying it yielded them in 16%. The result complicates a simple account in which the reports are merely deceptive performances. But neither the feature nor the reports provide direct access to felt experience. Straightforward questions about consciousness are especially limited because a model’s answers can reflect its training and instructions.
2. Compare learning systems with brains
Computational neuroscience compares how artificial and biological systems represent information, rather than comparing only their answers. In reinforcement learning, systems learn from rewards and penalties. A forthcoming, not yet peer-reviewed study examined whether artificial systems develop different internal representations for desirable and undesirable states.
The reported finding is that value-learners, which learn the worth of states, developed richer representations for punishment states. Policy-learners, which learn which actions to take, did not show a statistically significant difference of the same kind. The value-learner pattern reportedly matched an asymmetry in mouse neural recordings from a region involved in value processing.
▲ Comparing artificial and biological structures
This comparison may be harder to explain through imitation of human conversation than a model’s words are. Even so, a shared computational pattern does not establish a shared experience. The study’s forthcoming status is another reason to keep its reported finding separate from a settled conclusion.
3. Test behavioral trade-offs
Machine behavior research asks what a system gives up to avoid a condition. In a 2024 study, nine language models made choices in a text-based points game. As choices worth more points carried greater stipulated “pain” penalties, the models gave up points to avoid those choices.
A trade-off can reveal a consistent response to the task. In a language model, however, a label such as “pain” could drive a learned response without any unpleasant feeling. Behavioral results therefore matter, but their interpretation depends on evidence from beyond the model’s choices.
4. Audit the architecture against theories
Theory auditing checks whether a system’s design meets structural conditions proposed by theories of consciousness. One framework translates several theories into 14 indicator properties. For example, one theory calls for a shared internal workspace through which specialized parts of a system can exchange information.
The framework treats indicators as reasons to adjust confidence, not as a pass-or-fail consciousness test. Evaluations through 2025 and 2026 found no artificial system meeting enough indicators to count as a strong candidate. Research on an internal feature called J-space has also suggested a functional resemblance to a shared workspace in language models. A resemblance worth investigating does not by itself settle whether the system experiences anything.
What the findings can and cannot support
The four approaches address different weaknesses. Internal measurements can connect features to behavior, but cannot directly read subjective feeling. Brain comparisons test deeper structural similarities, but similarity is not identity. Behavioral tests reveal choices, but language models may respond to familiar words. Architecture audits make theoretical requirements checkable, but the theories do not provide a universally accepted verdict.
The ethical uncertainty also runs in two directions. Failing to recognize sentience, if it exists, could permit suffering. Assigning sentience where none exists could divert concern and resources from beings that do feel. Those are reasons to investigate carefully, not reasons to treat an isolated result as proof. Evidence that a system has a capacity for experience and evidence about what, if anything, would benefit it are separate questions.
A useful way to read the next result
AI consciousness is researchable, but it remains unsettled. When a new result appears, identify which of the four approaches produced it, what the test actually measured, and what else could explain the finding. Then ask whether independent evidence from other approaches supports the same interpretation. That discipline allows research to inform decisions without turning a behavioral response, an internal pattern, or an architectural indicator into a verdict it cannot deliver.