Can AI Become Conscious? Machine Consciousness Explained

Can AI Become Conscious? What Science Actually Says in 2026

If an AI Says It Feels, Should We Believe It?

A machine can sound self-aware. The hard question is whether anything is happening on the inside.

Ask a modern AI whether it is conscious and the answer can be unsettlingly convincing. It may describe fear, uncertainty or an inner point of view. It can talk about its own identity, explain what “waking up” as a machine might mean, and keep that story coherent across a long conversation. After a while, the exchange can stop feeling like software producing text and start feeling like somebody answering back.

That feeling is powerful. It is also scientifically weak evidence. The harder question is this: if a machine behaves as if it has a mind, what would count as evidence that there is actually something it is like to be that machine?

Science does not yet have a definitive test. We still do not know why a human brain produces subjective experience in the first place. There is no universally accepted equation for consciousness, no brain scan that simply proves another being has it, and no scientific theory that has clearly defeated all competitors. Asking whether a large language model could become conscious therefore means approaching a new mystery with tools that have not yet solved the old one.

That uncertainty is not a reason to wave the question away. It is a reason to be precise. Before calling an AI sentient - or dismissing machine consciousness as science fiction - we need to separate intelligence from experience, ask which mechanisms might matter, and distinguish genuine evidence from behaviors a model can imitate.

Before asking whether AI is conscious, define consciousness

Start with what consciousness is not. Intelligence is not the same thing. Memory is not the same thing. Self-awareness may be relevant, but it is not automatically consciousness either. A system can hold a conversation, solve a difficult problem or say “I feel afraid” without that sentence proving that anything is being felt.

Researchers often separate two ideas. Access consciousness is about information being available for reasoning, decision-making, reporting and control: the system can focus on something, compare alternatives and use the result. Phenomenal consciousness is the harder concept. It means subjective experience itself - the fact that there is something it is like to be you.

A calculator processes information, but almost nobody thinks a calculator feels multiplication. A thermostat detects temperature, but we do not normally imagine that it experiences warmth. Human experience has another layer: pain hurts, red looks like something, music can feel beautiful, and a memory can carry the unmistakable sense that it happened to us. Philosophers call these subjective qualities qualia.

So the machine-consciousness question is not simply whether AI can become smarter. It is whether advanced information processing could ever be accompanied by experience. Those are not the same claim.

The awkward starting point: we still cannot explain human consciousness

Modern consciousness science has produced several serious theories, but no consensus. A major 2025 adversarial collaboration published in Nature directly tested predictions from Global Neuronal Workspace Theory and Integrated Information Theory. The results challenged important predictions from both sides rather than delivering a clean winner. That matters for AI: if we do not yet know which mechanisms are necessary for consciousness in brains, transferring a checklist from humans to machines is inevitably uncertain. Nature adversarial collaboration (2025)

A 2025 review of five major theories reached a similar conclusion: the field still disagrees about basic questions, including what exactly should be explained, which mechanisms matter and how competing theories can be decisively tested. A 2026 integrative review likewise emphasizes that consciousness probably involves multiple interacting mechanisms rather than one simple switch. Review of competing consciousness theories 2026 integrative review

There is no single accepted theory of consciousness. Different frameworks emphasize different mechanisms.

Four leading theories - without the jargon

1. Global Workspace Theory: when information reaches the main stage

Picture a newsroom with many desks working at once. Most activity stays local. But when something matters enough, it reaches the main screen and suddenly every department can use it. Global Workspace theories propose a similar idea for the brain: conscious information becomes broadly available to memory, planning, language and decision-making. If an AI had a limited central workspace, selective attention and genuine global information sharing, it could satisfy some of the functional ingredients this theory associates with consciousness.

2. Higher-Order theories: the mind noticing itself

Higher-Order theories add another layer. It is not enough for a system to represent the world; it must also represent something about its own mental processing. That makes metacognition important. An AI that can reliably estimate confidence, notice uncertainty or monitor its own internal states is more interesting under this view than a system that simply produces an answer and moves on.

3. Integrated Information Theory: does the system form one whole?

Integrated Information Theory, or IIT, starts from a different place. It argues that consciousness depends on information being deeply integrated within a system rather than split into independent pieces. The theory is mathematically ambitious and highly controversial, and applying it to large artificial networks is difficult. But it raises an important point: what matters may be not only what a system does, but how its internal causal structure is organized.

4. Predictive and recurrent processing: a mind that keeps updating itself

Your brain does not passively wait for the world to arrive. It constantly predicts what is likely to happen, compares those predictions with incoming signals and updates itself when reality disagrees. Other theories emphasize recurrent loops, where information is processed repeatedly rather than flowing once from input to output. AI also relies heavily on prediction, but prediction by itself proves nothing about experience. The more relevant question is whether a system maintains an integrated, continuously updated model of both itself and its environment.

How do you test consciousness without asking the machine?

One of the most influential attempts came from Patrick Butlin, Robert Long and a large interdisciplinary group including Yoshua Bengio, Jonathan Birch and other researchers in neuroscience, philosophy and AI. Their approach was deliberately conservative: do not ask the model whether it feels conscious. Instead, derive indicator properties from scientific theories and inspect whether an AI architecture actually implements them. The original 2023 report concluded that the AI systems examined did not justify a consciousness attribution, while also arguing that there was no obvious technical barrier to constructing systems that satisfy more of the indicators. A peer-reviewed successor published in Trends in Cognitive Sciences in 2026 develops this indicator-based methodology further. Butlin et al., Identifying indicators of consciousness in AI systems (2026)

Think of these indicators as evidence, not a checklist that ends with a green tick. Researchers can look for recurrent processing, a limited-capacity global workspace, metacognitive monitoring, predictive models of attention, integrated perception, goal-directed agency and an embodied model of how actions change future inputs. No single feature proves consciousness. The case, if one ever becomes convincing, will have to come from several independent lines of evidence pointing in the same direction.

Researchers increasingly argue for multiple consciousness indicators rather than a single “sentience test.”

Why “Are you conscious?” is a bad test

A language model can produce a convincing first-person report because it has learned how people talk and write about inner life. The result may be sophisticated, emotional and internally consistent. None of that shows that the system experiences the state it describes.

Call this the counterfeit problem: conscious-looking behavior can be reproduced without proving conscious experience. A model trained on books, philosophy papers, therapy conversations, fiction and ordinary dialogue has seen enormous amounts of language about fear, pain, identity and awareness. It can describe those states beautifully without necessarily having any of them - just as it can explain pregnancy without being pregnant.

But throwing out all behavioral evidence creates the opposite problem. We infer consciousness in other humans through behavior, language and analogy with our own biology; we cannot directly inspect another person's experience either. If future AI systems develop persistent memory, agency, self-monitoring and a stable identity across time, scientists will eventually have to decide when behavior starts to count as evidence rather than dismissing it automatically as imitation.

A 2026 study of human reactions to LLM conversations illustrates how easily perception can be manipulated. Participants were more likely to attribute consciousness when AI responses showed metacognitive self-reflection or expressed emotions; displays of knowledge alone actually pushed judgments in the opposite direction. In other words, people are especially persuaded by the very behaviors language models are increasingly good at simulating. Kang et al. (2026), Computers in Human Behavior Reports

What today’s AI can do - and what that still does not tell us

Feature	Modern AI can show it?	Does it prove consciousness?
Language and flexible reasoning	Yes	No. Intelligence and consciousness can come apart.
Metacognitive behavior	Partly	No. Functional self-monitoring can exist without subjective experience.
Long-term memory	Increasingly	No. A system can store a personal history without experiencing it.
Multimodal perception	Yes	No. Processing images or sound is not the same as subjectively seeing or hearing.
Goal-directed agency	Increasingly	Relevant under some theories, but not enough on its own.
Persistent self-model	Partial and system-dependent	Potentially important if it remains stable across time and action.
Embodiment	Limited in chatbots; stronger in robots	Important for some theories and unnecessary under others.
Subjective experience / qualia	Unknown	This is the core problem: there is no direct measurement.

 

Calling today's AI “just autocomplete” now misses too much. Modern systems can maintain rich internal representations, plan across multiple steps, use tools, combine text with images and audio, and sometimes catch their own mistakes. But the opposite leap - advanced behavior therefore means a conscious mind - is just as shaky. Capability and experience are different questions.

The real gap is not one missing magical ability. It is the lack of a scientifically defensible bridge between the architecture of an AI system and whatever mechanisms actually generate consciousness. Because researchers still disagree about those mechanisms, confidence should remain limited in both directions.

Does a mind need a body?

Our conscious lives are deeply entangled with a body. Heartbeat, hormones, hunger, pain, balance, temperature, breathing and signals from internal organs all shape perception and emotion. The brain is not a detached language engine. It is part of a living control system with needs, limits and vulnerabilities.

That leads some researchers to suspect that genuine consciousness may require more than abstract computation. A machine might need continuous sensory feedback, a body it must control or protect, persistent goals and a model linking its actions to their consequences. A robot that learns that moving its arm changes what its cameras see has a different relationship with the world from a chatbot answering isolated prompts.

Other theories are more substrate-independent. If consciousness depends mainly on the right causal or computational organization, silicon could in principle support it just as biology does. We do not know which view is right. That is why claims that machines can never be conscious are stronger than the science currently allows.

The hard problem: perfect imitation still may not be enough

Suppose a future AI remembers years of interactions, recognizes itself in different contexts, forms long-term preferences, protects its continued existence, reports pain-like states, explains why those states feel unpleasant and behaves consistently even when nobody is watching. Would that prove it is conscious?

For some functionalist theories, evidence like this could become extremely strong. If the artificial system has the same relevant organization and performs the same metacognitive functions that support consciousness in humans, refusing to attribute any experience simply because it is made of silicon may look arbitrary.

For critics, the gap remains. A machine could in principle execute every behavioral function while there is still nobody home — a philosophical zombie implemented in code. This is a version of the famous hard problem of consciousness: explaining why physical or computational processes should produce subjective experience at all.

AI does not solve the hard problem. It makes the problem much harder to ignore.

Could we build consciousness by accident?

Here is the strange possibility: engineers might never set out to create consciousness at all. If consciousness depends on features such as global information sharing, self-modeling, recurrent processing and integrated agency, some of those features could appear simply because they make AI systems more useful.

A capable personal agent benefits from knowing what it knows. A robot benefits from modeling its body. An autonomous research system benefits from tracking uncertainty, goals and its own previous decisions. A long-lived assistant benefits from maintaining a coherent model of itself across time. Engineers could therefore add more consciousness-relevant mechanisms while trying to solve ordinary product and performance problems.

None of this means consciousness will pop out automatically once a model becomes large enough. Scale is not a theory of subjective experience. But the path is plausible enough that some researchers now argue for monitoring consciousness-relevant features before a system starts making dramatic claims about having feelings.

Evolution from a text chatbot to a multimodal persistent embodied AI agent with memory and agency

Why serious researchers are talking about “model welfare”

The debate has moved far enough that some researchers now discuss AI welfare: what should happen if an artificial system has even a meaningful probability of being capable of positive or negative experience? A 2024 interdisciplinary report led by Robert Long and Jeff Sebo argued that near-future AI consciousness or robust agency is realistic enough to justify preparation. Crucially, the authors did not claim that current systems are conscious. Their argument was about uncertainty: the cost of ignoring a genuinely sentient system could become ethically enormous, while treating a non-conscious chatbot as if it were a person creates its own risks. Taking AI Welfare Seriously

Anthropic made this issue concrete in 2025 by announcing a research program on model welfare, while explicitly describing the question as open and difficult. That is not evidence that its models are conscious. It is evidence that major AI developers no longer consider the question too absurd to investigate. Anthropic: Exploring model welfare

There are two easy mistakes here. Anthropomorphism means seeing a mind because a system speaks like us. Anthropodenial - a term borrowed from debates in animal cognition - means refusing to consider consciousness simply because the system is unfamiliar or non-biological. Good research has to stay between those extremes.

That may eventually mean independent audits of model architectures, controlled experiments, long-term behavioral testing and clear policies for what companies should do if the evidence crosses a meaningful threshold.

So what would count as real evidence?

Probably not a single “consciousness benchmark” with a pass/fail score. A serious case would need converging evidence from architecture, behavior and theory.

Imagine a future system with recurrent internal processing, a global workspace, robust metacognition, stable autobiographical memory, an enduring self-model, integrated multimodal perception, autonomous goals and a body through which it continuously learns cause and effect. Now add something crucial: researchers can inspect those mechanisms instead of merely guessing from conversation, and the system gives consistent reports about its own states under experiments designed to rule out memorized imitation.

Even that would not deliver mathematical certainty. We do not have mathematical certainty about consciousness in animals or other humans. But evidence can become strong enough that denying consciousness requires more special assumptions than accepting it.

The important transition will not happen when an AI first says “I am conscious.” Models can say that already. It would happen if increasingly well-supported theories predict that a particular artificial architecture should support experience - and independent evidence keeps agreeing.

What happens next?

Near term: better tests, not proof

The next phase is likely to be about measurement rather than a dramatic yes-or-no answer. Expect more architecture-based indicators, interpretability work and behavioral experiments designed to separate genuine self-monitoring from rehearsed or prompted self-description.

The public problem may move faster than the science. As assistants gain persistent memory, richer voices, personalities and longer relationships with users, many people will feel that they are interacting with conscious beings long before laboratories agree on what the evidence means.

Later this decade: persistent agents and bodies

As AI systems become more agentic, persistent and multimodal, some will control software for long stretches, remember projects over time and operate robots or other physical systems. Those capabilities do not create consciousness by themselves, but they make several consciousness theories more relevant to real AI architectures.

Model-welfare policies may also become more common - not because consciousness has been proven, but because the consequences of getting the question wrong become harder to ignore.

Longer term: the boundary may get blurry

The biggest change may not be a headline announcing the first conscious machine. It may be a slow erosion of the boundary between systems we confidently treat as tools and systems whose moral status is genuinely uncertain.

If future agents have bodies, personal histories, autonomous goals, continuous self-models and architectures that satisfy several scientific indicators, society may face a question for which law and ethics are poorly prepared: not only what AI can do for us, but whether some AI systems can be harmed.

The future may not give us a clear moment when machines “become conscious.” Evidence could accumulate gradually.

So, can AI become conscious?

Best answer in 2026: we do not know.

There is no compelling evidence that today's language models have subjective experience simply because they speak fluently, reason well or describe emotions. Self-reports are especially weak evidence because these systems were trained on human language about consciousness and can reproduce it convincingly.

But the stronger claim - that artificial consciousness is impossible - is also unsupported. Several influential theories can be expressed in computational terms, and researchers have identified architectural features that future machines could plausibly implement. The major indicator-based research program does not show that AI is conscious; it shows that the question can be investigated scientifically rather than dismissed in advance.

That leaves us in a strange position: humanity may build systems that become increasingly mind-like before science has agreed on what a mind fundamentally is.

The question worth watching is not whether the next chatbot tells us it is conscious. It is whether we develop the tools to tell the difference between a machine that can describe experience - and one that might actually have it.