Could a machine ever be conscious?

What exactly would a machine need to have?

The word “conscious” gets used in many ways, so it helps to pin it down. The Stanford Encyclopedia of Philosophy points to the philosopher Thomas Nagel’s famous test from 1974: a creature is conscious if there is something it is like to be that creature. There is something it is like to be a bat using sonar, even if we can’t imagine it. And most people think, as Chalmers puts it, that there is nothing it is like to be a water bottle.

Some questions about the mind seem within reach. Scientists can study how the brain takes in information and uses it. Philosophers call these the “easy problems,” though they are hard work. The hard problem, which David Chalmers set out in 1995, is explaining why any of that processing comes with felt experience at all. For machines, this matters a great deal. We can describe what a computer does, step by step. What we can’t yet say is whether doing it would ever feel like anything.

Chalmers also stresses, in his 2023 essay on language models, that consciousness is not human-level intelligence. There is a consensus, he writes, that many animals are conscious, like cats or mice or maybe fish. So a machine would not need to be brilliant to feel. And being brilliant would not prove that it does.

Could a test tell us?

The most famous test is Alan Turing’s. In a 1950 paper, “Computing Machinery and Intelligence,” he described an imitation game. A judge chats with a person and a machine without seeing them, then guesses which is which. As the Stanford Encyclopedia explains, Turing predicted that in about fifty years, an average judge would have no more than a 70 percent chance of guessing right after five minutes.

Machines have now done well at this game. In a 2025 study posted as a preprint (not yet peer-reviewed when it appeared), Cameron Jones and Benjamin Bergen ran five-minute three-way chats. When told to act like a person, GPT-4.5 was picked as the human 73% of the time, more often than the real humans were.

But the Turing test measures behavior. It was built to ask whether machines can think, not whether they feel. In 1980, the philosopher John Searle offered a thought experiment called the Chinese Room. Picture a man who knows no Chinese, alone in a room with a rulebook. Notes in Chinese come in under the door. By following the rules, he sends back fitting answers. People outside think someone in the room understands Chinese. Nobody does. Searle concluded that running a program might look like understanding without being understanding.

The Chinese Room is an analogy, and critics say it breaks. What Searle himself called perhaps the most common reply, the Systems Reply, grants that the man does not understand Chinese but says the whole system, rulebook and all, might. The debate is still alive. Either way, it shows why outward behavior alone can’t settle the question.

What do the leading theories say about AI?

In 2023, Patrick Butlin, Robert Long and 17 co-authors tried a different route. Instead of judging behavior, they looked inside. They drew a list of “indicator properties” from several scientific theories and checked whether real AI systems have them.

One theory they used is global workspace theory. On this view, the brain has many specialized parts working at once. Consciousness happens when some information wins a competition for a small shared “workspace” and is broadcast to all the parts. A machine could be built this way in principle. The report found that some indicators are already met by existing systems, but that no current system is a strong candidate for consciousness. It also found no obvious technical barriers to building systems that meet the indicators.

That conclusion rests on a working assumption the authors state openly: computational functionalism, the idea that doing the right kind of computation is enough for consciousness. They call it a mainstream but disputed view.

A major rival, integrated information theory, disagrees. The report explains that its backers claim digital computers are unlikely to be conscious, whatever programs they run, because on this theory a system that ran the same steps as a human brain would still not be conscious if its parts were of the wrong kind. So the two theories point in different directions for machines. And the science is not settled. In a 2025 Nature study of 256 people, backers of integrated information theory and of a brain-based version of global workspace theory agreed on predictions in advance. The results fit some predictions of each while, in the authors’ words, “substantially challenging key tenets of both theories.”

What is the strongest objection?

The strongest objection to machine consciousness is that it may need biology. Searle argued that minds come from biological processes, and computers can at best simulate them. Integrated information theory reaches a similar result for different reasons. If either is right, even a perfect copy of a brain’s software, run on an ordinary computer, would likely not feel.

The strongest reply comes from computational functionalism, which Butlin and colleagues call a mainstream position: what matters is the kind of processing, not what it runs on. Experts are split. In a 2020 survey of philosophers reported by Chalmers, about 3% accepted or leaned toward the view that current AI systems are conscious, and 82% rejected or leaned against it. About 39% accepted or leaned toward the view that some future AI systems will be conscious, 27% rejected or leaned against it, and 29% were neutral.

Chalmers also worked through rough numbers based on mainstream assumptions: somewhere under 10 percent that current language models are conscious, and 25 percent or more that conscious, more capable successors arrive within a decade. He warned readers not to take the exact numbers too seriously, and noted that his own views would give somewhat higher odds. The numbers are best read as a sign that careful thinkers treat the future case as open.

What do people often get wrong?

“The chatbot told me it has feelings.” Lemoine relied heavily on LaMDA saying so. Chalmers reports that a one-word change to the question, asked of another model, got answers going both ways, from “Yes, I’m not sentient” to “Well, I am sentient.” Reports that fragile are weak evidence. Google called Lemoine’s claims “wholly unfounded” and, the BBC reported, fired him in July 2022, saying he had broken its employment and data security policies.

“Smarter means more conscious.” As Chalmers notes, the two are different. A system could be very capable and feel nothing, or feel something without being very capable.

“If no one can prove it, the question is pointless.” The Butlin report shows otherwise. Researchers can check specific features, compare them with theories that are tested in human brains, and update as the theories improve.

“We will soon upload our minds.” Nobody can do this today. A wiring map of a whole adult fruit fly brain, published in Nature in 2024, covers 139,255 neurons. A human brain has about 86 billion. And even a perfect copy would face every question on this page.

How can you check this yourself?

• Ask what kind of evidence is on offer. Is it the machine’s own words, or a look at how it works inside? Butlin and colleagues explain why the first is unreliable.
• Read the short versions. The Butlin report’s summary is at the front, and it lists every indicator property in one table.
• Try the one-word test. Ask a chatbot whether it is conscious, then rephrase the question to invite the opposite answer. Notice how easily the answer shifts.
• Watch for loaded words. “Sentient,” “aware” and “intelligent” mean different things. Ask which one a headline means.

After all this detail, the answer stands. There is no good evidence that today’s chatbots feel anything. Whether a future machine could is an open question, and it will stay open until science learns what makes any mind conscious. For more on that deeper puzzle, see our story What Is Consciousness?