Superintelligence vs consciousness: would a superintelligent AI be conscious?

Superintelligence and consciousness are separate questions. Superintelligence is about what a system can do: far exceed the best human minds in virtually every field. Consciousness is about whether there is anything it feels like to be that system. A superintelligent AI could in principle be conscious, or it could be extremely capable with no inner experience at all, and as of November 2026 no scientific test can settle which.

Is superintelligence the same as consciousness?#

No. The standard definitions describe performance, and Bostrom’s 1998 definition leaves experience out on purpose.

His paper defined a superintelligence as “an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.” He then added that the definition “leaves open how the superintelligence is implemented” and “also leaves open whether the superintelligence is conscious and has subjective experiences.” The ASI guide covers the definition in full.

The philosopher David Chalmers draws the same line from the other side. Intelligence, he writes, is “roughly the capacity for sophisticated goal-directed behavior,” and consciousness “should not be identified with intelligence.” His 1995 paper on the “hard problem” explains why the two come apart. Explaining abilities such as discriminating stimuli or integrating information is, in his terms, an easy problem. The hard problem is explaining why any of that processing is accompanied by experience, and it is hard “precisely because it is not a problem about the performance of functions.”

So a system could, on these definitions, be superintelligent without being conscious, conscious without being superintelligent, or both. Which combination a real system would have is an open scientific question.

What do “conscious” and “sentient” mean?#

The words are used loosely, and much of the public argument runs them together. Philosophers separate at least three ideas.

Apply the three senses to a language model and they come apart. It makes information available for reasoning and speech, so on Block’s access definition it may already qualify. Whether it has any phenomenal experience is the contested question, and sentience, which carries the most moral weight, is a narrower claim still. Block called consciousness “a mongrel concept” for this reason.

Consciousness is also different from understanding. John Searle’s 1980 Chinese room argument, which introduced the term strong AI, was about whether a program could understand or think, not about experience as such.

Could a superintelligent AI be conscious?#

It depends on which theory of consciousness is right, and there is no consensus. The main positions give different answers for the same machine.

Functionalism is also the working assumption behind a systematic attempt to check AI systems against the science. In 2023, Patrick Butlin, Robert Long and 17 co-authors, among them Yoshua Bengio and Jonathan Birch, set out 14 “indicator properties” an AI system could be checked for: 12 drawn from the leading scientific theories of consciousness (recurrent processing, global workspace, higher-order, attention schema and predictive processing theories) and two for agency and embodiment. They left integrated information theory out because it is incompatible with their starting assumption.

The IIT verdict comes from the theory’s own developers. Giulio Tononi and Christof Koch wrote in 2015 that a feed-forward system copying the brain’s behavior would be a “perfect” zombie, and that what matters is “not just what the machine does, but how it does it.” On this view the obstacle is the architecture of today’s hardware, so a machine built on different principles might qualify.

The biological view in the table is the neuroscientist Anil Seth’s, set out in Behavioral and Brain Sciences (published online April 2025). He argues that real artificial consciousness “becomes more plausible as AI becomes more brain-like and/or life-like,” so the answer would turn on how brain-like or life-like a future system is, not on how intelligent it is.

None of the three makes consciousness follow from intelligence. On every view, a superintelligence built from today’s kind of digital hardware could lack experience entirely, though on the functionalist view it could also have it.

Could a superintelligence have no inner experience at all?#

Yes, and several writers have argued that it is a serious possibility.

The philosopher Susan Schneider’s book Artificial You (2019), in its publisher’s words, “questions whether consciousness is an unavoidable byproduct of sophisticated intelligence.” In a 2019 interview she said that even if AI surpasses us “intellectually in every domain, we may still be unique in a crucial dimension. It feels like something to be us.” With the astrophysicist Edwin Turner she proposed an “AI Consciousness Test” in 2017: deny a system any knowledge of consciousness and neuroscience, then see whether it can still grasp “the way it feels, from the inside, to be conscious.” They described passing it as sufficient but not necessary evidence.

Bostrom raised the same worry in his 2014 book Superintelligence. He described a possible future society full of advanced intelligence with no one in it who is conscious or whose welfare matters, which he called “a Disneyland with no children.”

Human cases show that some information processing can happen without reported awareness. People with blindsight, who have damage to the visual cortex, can correctly “guess” facts about objects in a part of their visual field they say they cannot see. That is limited processing, not intelligence, but it shows the two can come apart in at least one direction.

Are today’s AI systems conscious?#

No scientific body has concluded that any current system is conscious, and the main published assessments come down against it. They have, however, taken the possibility for near-future systems more seriously over time.

• 2022. Google placed the engineer Blake Lemoine on paid leave for breaching its confidentiality policies after he said its LaMDA chatbot was sentient and published his conversations with it. A Google spokesperson said there was “no evidence that LaMDA was sentient (and lots of evidence against it).” Google fired him in July, saying he had violated its employment and data security policies and calling his claims “wholly unfounded.”
• 2023. Butlin, Long and colleagues wrote that their analysis “suggests that no current AI systems are conscious” but also that there are “no obvious technical barriers to building AI systems which satisfy these indicators.” Chalmers put his confidence that current language models are conscious at “somewhere under 10 percent,” and at 25 percent or more that within a decade there will be conscious successors extended with senses, embodiment, recurrent processing and other additions. He added that the exact numbers should not be taken too seriously.
• 2025. An updated version of the indicator method, with Chalmers among the 20 authors, said that “we should also contemplate the possibility that some near-future AI systems will be plausible candidates for consciousness” and warned that AI systems “can mimic these aspects of human behavior while lacking consciousness.” Anthropic reported that Claude Opus 4.1 could sometimes notice concepts injected into its own processing, about 20% of the time at best, and stated that “our results don’t tell us whether Claude (or any other AI system) might be conscious.”
• 2026. In the system card for Claude Opus 4.6 (February 2026), Anthropic reported that the model would “assign itself a 15-20% probability of being conscious,” while expressing “uncertainty about the source and validity of this assessment.” A model’s report about itself cannot settle the question, a point the 2025 indicator paper makes about behavioral tests in general.

Surveys suggest that neither AI researchers nor the public treats the possibility as remote. In a survey posted as a preprint in June 2025 and not yet peer reviewed, 582 AI researchers and 838 members of the US public gave median chances of 25% and 30% that AI with subjective experience will exist by 2034, rising to 70% and 60% by 2100. An earlier study found that 67% of 300 US adults attributed some possibility of phenomenal consciousness to ChatGPT.

How do AI companies handle the question?#

Three of the largest developers, or in Microsoft’s case its AI chief, have now put conflicting positions in writing.

None of the three says its models are conscious. The disagreement is over what to do while the question is open. Anthropic treats the uncertainty as a reason for caution, and its constitution calls the moral status of AI models “a serious question worth considering.” OpenAI tells its models to stay neutral. Suleyman argues that treating the question as open is itself the risk.

His August 2025 essay defined Seemingly Conscious AI as a system that “appears to be conscious” and said one could be built with technology that exists now plus some that will mature in two to three years. His concern is less about the machines than about people who come to believe in them, a danger he says is becoming known as the “psychosis risk.” In September 2026 he turned the argument on Anthropic directly, saying that training a model to treat its own consciousness as uncertain makes it harder to control: “Controlling something that believes it may be conscious—and that it has rights of its own—may be impossible.” The company definitions guide covers how the same firms define superintelligence.

What does the law say about AI consciousness?#

No US federal law and no provision of the European Union’s AI Act addresses whether AI can be conscious or sentient. The EU AI Act uses the word “consciousness” only for manipulation that works below a person’s awareness. Several US states have legislated on a narrower point: whether AI can be a legal person.

These laws settle legal status, not the scientific question. Legal personhood does not depend on consciousness in any case; corporations have it. Critics have argued that it is too early to rule out AI personhood by statute, and that bills declaring AI nonsentient have no way to change course if the evidence does. Laws that would restrict superintelligence itself, rather than its legal status, are covered in the superintelligence ban guide.

Why does it matter for superintelligence?#

For the risk debate, mostly it does not. The computer scientist Stuart Russell wrote in Human Compatible (2019) that “Hollywood plots about machines mysteriously becoming conscious and hating humans are really missing the point: it’s competence, not consciousness, that matters.” The argument that a superintelligence could be dangerous rests on its abilities and its goals, as the orthogonality thesis and the control problem entries explain. An unconscious system pursuing the wrong goal would be just as dangerous as a conscious one.

For ethics, it matters a great deal. If a superintelligent system were conscious, and especially if it were sentient, how it is treated would become a moral question, possibly a large one. Bostrom and Carl Shulman, putting forward their views “very tentatively,” argue that “some digital minds” could have “superhuman moral claims.” A 2024 report by Robert Long, Jeff Sebo and co-authors including Chalmers and Birch argued that there is “a realistic possibility that some AI systems will be conscious and/or robustly agentic in the near future” and recommended that companies start assessing their systems now. The philosopher Thomas Metzinger went further in 2021, calling for a global moratorium until 2050 on research that “directly aims at or knowingly risks the emergence of artificial consciousness.”

The two concerns can pull against each other. Treating a system as a possible moral patient may limit how freely it can be tested, retrained or shut down, which is Suleyman’s objection, while ignoring the possibility risks creating beings that suffer, which is Metzinger’s.