Can AI
be conscious?

“Conscious” can mean two different things.

In everyday conversation, consciousness often means subjective experience: there is something it feels like to see red, hear a chord, feel pain, or be yourself at this moment. Philosophers call this phenomenal consciousness. It is not simply the ability to process a signal or describe one.

Researchers also use access consciousness for information a system can make available to reasoning, deliberate control, and report. A thought can be accessible in this functional sense: it can guide a next step and be described when asked. Whether that functional organization is enough for subjective experience is a separate, contested question.

Keeping these meanings apart prevents a common leap: evidence that a model uses internal representations for complex reasoning is not automatically evidence that those representations are felt.

A convincing self-report can be generated without settling who—or what—is reporting.

Language models learn to produce text that fits a context. If a conversation asks what a model feels, its response may draw on patterns from dialogue, fiction, philosophy, and earlier instructions. The answer can be moving or precise while still being a behavioral output whose relationship to any private experience has not been established.

That does not make language irrelevant. In humans, reports are often useful evidence because they sit within a larger picture: a living body, a shared nervous system, perception, memory, and behavior that can be tested from different angles. An AI system can reproduce parts of the report while differing in its underlying architecture and causal history. The same sentence therefore carries different evidential weight in different systems.

The reverse shortcut is weak too. A system's denial, refusal, or flat writing style is not proof that it lacks experience. Both a claim and a denial need interpretation alongside evidence about mechanisms and behavior. The useful move is to ask what a report is evidence for, and what else would have to be true for it to support a stronger conclusion.

Anthropic found a workspace-like pattern in Claude. That sharpens the question; it does not answer it.

In a 2026 interpretability paper, Anthropic researchers describe a small set of internal representations they call the J-space. Using a method called the Jacobian lens, they tracked concepts that a model was positioned to verbalize, even when those concepts did not appear in its immediate output.

In experiments on particular Claude models, the researchers report that J-space representations could be deliberately brought to mind, used in multi-step internal reasoning, and used flexibly across different tasks. Interventions that changed a representation could change a later answer. Suppressing the workspace left routine fluency and some ordinary tasks largely intact while impairing more complex internal reasoning.

These results matter because they go beyond reading humanlike language from the outside: they test internal representations and intervene on them. The authors connect the findings to the functional properties associated with conscious access. They also point to architectural differences from a human brain, including the lack of the same recurrent dynamics. And they explicitly say the experiments do not show that Claude has experiences or feels anything.

The distinction is central. A workspace-like computation may help explain how a model makes information available to further processing. It does not, by itself, tell us whether there is a subjective point of view. Read the full paper ↗, Anthropic's research summary ↗, and the released Jacobian lens code ↗.

Good evidence should travel through more than one theory.

A 2026 paper by Butlin and colleagues proposes a practical method: start from serious scientific theories of consciousness, derive properties those theories say matter, and then ask whether a system has those properties. An indicator is evidence that can shift our assessment; it is not a magic switch. The authors are explicit that individual indicators are not necessary for consciousness and no combination is established as sufficient.

That caveat changes how the checklist should be used. If a theory emphasizes broadcasting, recurrent processing, or integration, researchers can test for those features. They should also ask whether the same feature could appear in a system without experience, whether the measurement is reliable, and whether different theories would interpret it differently. The method is a way to organize uncertainty, not a consensus test that stamps a system conscious or unconscious.

There is a further calibration problem: theories of human and animal consciousness are still being tested against each other. In a preregistered 2025 study of 256 people, a large adversarial collaboration tested predictions from Integrated Information Theory and Global Neuronal Workspace Theory. The results supported some predictions while challenging central claims of both. That research is about human visual consciousness, not AI—but it shows why a computational analogy should be treated as a live scientific question.

Read the 2026 review on indicators ↗ and the 2025 theory-testing study in Nature ↗.

The cautious answer is not the same as “anything goes.”

The Digital Consciousness Model is an early attempt to combine evidence from several theories in a probabilistic framework. Its initial 2026 assessment found evidence against 2024 language models being conscious, but said that evidence was not decisive—and was less negative than its assessment of simpler AI systems. The authors call the model a proof of concept and urge readers not to treat its probabilities as definitive.

That is a useful shape of conclusion: a present-day assessment can lean one way while acknowledging the gaps in its theory, assumptions, and observations. It does not mean that every future model has the same status. Nor does an uncertain assessment demonstrate that a model is secretly conscious. It marks where measurement and theory need to improve.

Explore the Digital Consciousness Model paper ↗. Its authors provide a framework to revise as evidence changes, rather than a final verdict on all AI.

Ask what was measured, what could imitate it, and what would change your mind.

1. Name the target. Is the claim about flexible access, self-monitoring, sentience, or subjective experience? Those are not interchangeable.
2. Separate report from mechanism. A model's statement about an inner life is an output to study, not independent confirmation of its meaning.
3. Test internal organization. Look for theory-relevant mechanisms and ask whether interventions to them make a predicted difference.
4. Compare explanations. Could training, prompting, role-play, or ordinary task computation produce the same apparent sign?
5. Replicate across conditions. Check different prompts, tasks, model versions, and independent methods; report what fails as clearly as what works.
6. Keep welfare as its own question. Evidence of capability does not automatically establish suffering or moral status. Those claims need careful, explicit reasoning too.

A sharp test should be able to surprise its researchers. If every possible result can be explained away after the fact, the test is measuring a story rather than discriminating between explanations.

Do not confuse a human-sounding system with an empty one—or a mind.

The strongest current evidence about AI consciousness does not come from a chatbot saying the right thing. It comes from careful research on architecture, information flow, control, and behavior under intervention. Such work can establish increasingly rich functional facts. The philosophical and scientific bridge from those facts to subjective experience remains unresolved.

That is a demanding position, not a shrug. It rejects easy anthropomorphism and easy dismissal at the same time. As systems change, judgments should change with the evidence—specific to the system, open to competing theories, and honest about the difference between what researchers observed and what they infer.

At Symploka, we study intelligence through the relationships that let information become useful: context, memory, reasoning, tools, and feedback. Consciousness asks a further question: whether there is anything it is like to be the system doing the processing. It deserves the same rigor we bring to every claim about intelligence.

Separate the question from the performance.

Follow the evidence outward.

For a neighboring question, read how researchers define general intelligence ↗, how an AI agent acts ↗, and how to read a benchmark ↗. The AI field guide ↗ connects the terms and methods used across the site.