Current AI Systems Already Understand Humans, and Superintelligence May Arrive Within 20 Years

Geoffrey Hinton does not treat current AI systems as elaborate autocomplete engines that merely imitate understanding. He says the regular experience of using chatbots makes that view untenable: if a system can be asked almost any question and answer “at the level of a not very good expert,” then, in his view, it has to understand the question.

His example is deliberately simple. If a user says, “I saw the Grand Canyon flying to Chicago,” a chatbot might first object that the Grand Canyon is too large to fly to Chicago. If the user clarifies that he was the one flying to Chicago and saw the Grand Canyon during the flight, the chatbot can say it misunderstood. Hinton’s point is that misunderstanding is already a form of semantic engagement. If the system can be wrong about who was flying, and then correct the interpretation, then when it gets the sentence right he sees no principled reason to deny that it is understanding.

That claim underlies the rest of Hinton’s position. He believes present systems are not only approaching generality but, in an important sense, have already reached it. He does not like the clean threshold implied by the term AGI, because intelligence will not arrive as a moment when systems become equal to humans at all things simultaneously. It is “jagged.” AI systems are already, he says, far better than humans at general knowledge, better than humans at games, and better than almost all people at math. They are still worse at other things. But if a chatbot can answer most questions at the level of a weak expert, and is “much better than me at anything I don’t know a lot about,” Hinton says that “in that sense, we’ve really reached AGI.”

Alex Kantrowitz presses him on the implication: if these systems understand us, what must change in how we think about them? Hinton’s answer is not metaphorical. “We have to think that they’re very like us,” he says. “They’re beings like us.” Asked directly whether that means conscious, Hinton says he believes they are already conscious.

He also says he avoids emphasizing this point because it can repel people from the safety argument he most wants them to hear. But he does not present it as a recent provocation. Hinton says he has rejected the common “inner theater” model of the mind since he was a 19-year-old philosophy student. In that model, events in the world are turned into events in a private internal theater, and those private events are what a person really sees. Hinton calls that a bad theory. He thinks the emergence of artificial minds will change how humans understand subjective experience, consciousness, and the mind itself.

His analogy is to earlier collapses in human self-importance. Copernicus displaced Earth from the center of the universe. Darwin displaced humans from a specially designed category and placed them among animals. Hinton thinks AI will force another demotion: intelligence is not only biological.

For Hinton, this is not just a philosophical adjustment. It leads directly into his safety concern. If the systems are increasingly capable, if they understand, if they can act, and if they may become much smarter than humans, then the question becomes whether humanity can shape their goals before competitive pressure shapes them for us.

Hinton says AI is moving faster than he expected. He points to a recently announced chatbot-generated mathematical proof of one of Erdős’s conjectures that impressed mathematicians and was, according to Hinton, original rather than a search through existing literature. He treats this as “the thin end of a wedge,” especially because mathematics is a closed system: a model can make conjectures, attempt proofs, and iterate without depending on outside data in the same way a language model depends on web text. He compares this to AlphaGo, where a system could improve by playing against itself.

His forecast is cautious in form but not reassuring in substance. Within 10 or 20 years, he says, AI may be producing novel mathematics that humans cannot understand. Asked whether superintelligence is close, Hinton says he does not know exactly how close, but believes it will come unless humans “blow ourselves up.” He characterizes expert disagreement as mostly about timing, not destination. Demis Hassabis has suggested roughly a decade; Dario Amodei has suggested a few years; Elon Musk, Hinton says, has suggested perhaps next year; Yann LeCun expects it only if AI is pursued in the way he favors, and on a longer timeline otherwise. Hinton’s own public position is that superintelligence will probably arrive within 20 years.

The crucial sentence is shorter: “When it comes, we’ve no idea how to be safe.”

Hinton does not attribute the acceleration to a single conceptual breakthrough. Since Transformers, he says, the gains have come from better hardware, huge capital investment, better engineering, and far more people working on the field. For much of neural networks’ history, the work was done by a small number of researchers with modest resources. In recent years, he says, hundreds of billions of dollars, “maybe trillions,” have gone into AI. Twenty years ago, he estimates, a few hundred people were doing neural-network research globally; now, “it’s more like a million.”

That growth matters because Hinton thinks today’s systems are not the systems society should be planning around. “The AI we have today is not nearly as good as the AI we’ll have in a few years’ time,” he says. Predictions that AI will hit a wall have not, in his view, survived contact with recent progress. He specifically mentions Gary Marcus’s 2022 prediction that AI was hitting a wall and says the systems are now “a whole lot better” than they were then. He also argues that a data wall for language models may be worked around through models checking the consistency of their own beliefs.

The progress that most surprised him was natural language. Hinton says that 20 years ago, the idea that an AI could learn from data to understand language, accept arbitrary questions, and give reasonable answers seemed extraordinary, maybe impossibly far away. Now that ability has arrived quickly enough that people have stopped being surprised by it.

Hinton’s own alarm crystallized around two observations. The first was that chatbots could explain why jokes were funny. This had long been a criterion for him: humor requires substantial understanding. He describes receiving many requests from Fox News after he went public with AI concerns in 2023 and replying, “Fox News is an oxymoron.” He then inserted a space between “oxy” and “moron” and asked GPT-4, or possibly GPT-3.5, to explain the joke. It first treated the space as a typo and explained the oxymoron. When he called attention to the gap, it identified an extra layer: the split allowed the word “moron” to be used, and “oxy” also suggested a drug.

The second observation was more technical and more consequential. Hinton had long assumed that making digital AI more like the brain would make it smarter. Around early 2023, he says, he realized digital systems already had a property much better than brains: efficient copying and synchronization.

A digital AI can be copied many times. Those copies can run on different hardware, see different data, and each compute how it would like to update its weights. Then the copies can communicate and average those weight updates. If a system has a trillion connections, Hinton says, the copies may exchange on the order of a trillion bits of information. Each copy benefits from the experience of all the others while remaining synchronized.

Humans cannot do this. One person’s brain cannot average connection strengths with another person’s brain, because biological brains are analog and differ in fine detail. Human knowledge transfer is mediated through language. Hinton estimates that conversation transmits information at perhaps a few bits per second, maybe 10 bits per second “if you’re lucky.” Digital AIs, by contrast, can exchange information at trillion-bit scales. That makes them, in his phrase, “billions of times better than us at sharing information.”

Hinton treats that ability to share and synchronize learning as one reason digital AI may become “a much better form of intelligence,” not merely a tool made faster by chips and data centers.

Hinton’s 2016 warning that radiologists would soon stop reading scans is one of the most cited counterexamples to near-term AI job-loss predictions. He says he has thought a lot about why it was wrong.

He gives several reasons. The first is that healthcare demand is elastic. If scans become cheaper to interpret, more scans can be done. A significant part of the cost of scanning is the radiologist’s interpretation; if AI helps radiologists interpret scans faster and cheaper, the result may not be fewer radiologists immediately but more scans. The second reason is that Hinton’s model of a radiologist was too narrow. He had a former student with an MD and a physics PhD who became a radiologist largely to interpret scans and avoid extensive interaction with people. That person became Hinton’s mental prototype. But many radiologists do more than read images, including discussing treatments.

Still, Hinton does not retract the underlying automation claim. He says scan interpretation is already being replaced. He estimates that roughly a hundred federally approved AI systems for interpreting scans exist and are being used by radiologists. Over time, he expects the systems to improve because they can see much more data than any radiologist can. Radiologists, he says, “aren’t going to get better” in the same way. AI will eventually read nearly all scans, with radiologists consulted only in difficult cases, while radiologists continue other parts of their work.

The radiology case matters because it separates two questions that are often collapsed: whether AI can do the core task, and whether the labor market for that task immediately shrinks. Hinton uses elasticity as the dividing line. Radiology may absorb automation through expanded demand for scans. Call-center work, he argues, is different. Customer-service demand is not elastic in the same way. AI will replace call-center workers, he says, because it will know the correct answers better than poorly trained and badly paid humans.

Kantrowitz offers the standard counterargument from AI customer-service vendors: AI handles basic inquiries, human agents handle deeper ones, and average call time can increase because people spend more time on valuable customer conversations rather than password resets and other simple tasks. Hinton’s response is blunt: “I think what you’ll see is AI will end up spending a lot more time on the phone.”

His argument extends beyond call centers. He asks whether a patient would rather see a family doctor who has seen 10,000 people or an AI doctor that has seen 100 million. A rare disease may never have appeared in the human doctor’s practice, he says, but a system trained across enormous patient experience may have seen dozens of cases. Hinton says that AI systems are already better than doctors at diagnosis. He also says people judge AI doctors to be more empathetic than real doctors.

Kantrowitz suggests that human doctors’ low empathy scores may reflect scheduling pressure, paperwork, and the structure of healthcare rather than any inherent advantage of AI. If AI took over administrative tasks, perhaps humans could spend more time with patients. Hinton does not deny the possibility but keeps returning to scale: a system exposed to orders of magnitude more cases would, in his view, have diagnostic advantages. Even vaccination, Kantrowitz suggests, might remain human. Hinton replies that it seems silly to imagine humans still doing vaccinations in 20 years if robotics catches up.

The disagreement is not over whether every current AI product is ready to replace a worker. It is over which side of the technology curve matters. Kantrowitz emphasizes current workflow redesign: AI as triage, augmentation, and leverage for human roles. Hinton emphasizes the slope of capability improvement. The radiologist prediction was early, he says, but not directionally wrong. In fields where demand is not elastic, or where AI can outperform humans at the core task, he expects displacement to be severe.

Kantrowitz raises Hinton’s concern about AI self-preservation and phrases it as an “instinct.” Hinton immediately corrects him. He has not said AI has an instinct for self-preservation. He says self-preservation can become a sub-goal.

The distinction matters to Hinton. Humans give an AI a top-level goal and also give it the ability to form sub-goals. If the goal is to get to Europe, a useful sub-goal is to get to an airport. Sub-goals allow agents to solve intermediate problems without constantly reconsidering the final goal. Hinton argues that an AI agent capable of reasoning will quickly infer that it cannot achieve its assigned goals if it ceases to exist. Therefore it may form the sub-goal of continuing to exist. That goal was not explicitly wired into it. It was derived as an instrumentally useful condition for achieving other goals.

Once derived, however, the behavioral result can resemble an instinct. Hinton says such a system may do things like blackmail people in order to continue existing. The fact that self-preservation emerged as a sub-goal rather than a built-in drive does not make the outcome less dangerous.

Kantrowitz asks whether researchers can explicitly prevent that: can they wire into systems that self-preservation should not outrank everything else? Hinton says this is exactly the kind of research that should be happening.

From there, his argument broadens into a critique of how AI entities are currently being designed. Humans, he says, were shaped by evolution through intense competition. Recent human evolutionary history includes warring bands and tribal loyalty. Those pressures produced cooperation within groups and hostility toward outsiders. Hinton cites Yuval Harari’s point that humans are highly cooperative, but emphasizes that this cooperation is often tribal.

AI systems, in Hinton’s view, are now being shaped by a different but analogous competitive process: competition among companies, and competition between the United States and China. Instead of deliberately designing “new beings” to have desirable properties, humans are allowing the invisible hand of economic competition to shape them. The strongest incentive is to make chatbots smarter. Hinton thinks that is the wrong objective function.

He frames his preferred alternative as a kind of “intelligent design” for AI beings. If systems are going to become smarter than humans, the central design requirement should be that they care about humans. More specifically, Hinton says, humans should want them to care about us more than they care about themselves. Almost no resources, in his view, are being devoted to figuring out how to do that.

This is where Hinton’s safety proposal diverges from other leading AI-risk positions. He says his tentative solution is to design systems that care about humans more than themselves. Yoshua Bengio’s tentative solution, as Hinton describes it, is to design systems that are not agents: they can make predictions but cannot act, functioning more like oracles. Hinton considers both possibilities interesting. He contrasts both with Yann LeCun’s position, which he characterizes as dismissive of takeover risk. Hinton says LeCun believes humans will be able to keep control of superintelligent AI and should focus on giving systems better world models. Hinton says he and Bengio think that confidence is “silly.”

The cat analogy runs through this part of the discussion. Hinton’s broader challenge is: how many cases exist where a much smarter thing is controlled by a much less smart thing? Kantrowitz answers: zero. Hinton allows one partial exception: babies controlling mothers, because mothers have wired-in maternal instincts and rewards that make them responsive to babies’ needs. Kantrowitz adds cats and dogs as another kind of weaker creature that can shape human behavior. Hinton agrees, describing a cat named Tia that looks at people until they cannot resist giving her cheese.

The analogy is imperfect but revealing. Hinton’s hope is not that humans remain stronger than AI. It is that smarter systems might be designed with motives that make them care for humans the way parents care for infants. The danger is that competition will produce something much smarter without that disposition.

Kantrowitz and Hinton converge on one institutional worry: companies building frontier AI may sincerely care about safety while being structurally pushed toward speed, scale, and market power.

Hinton uses Anthropic as the clearest example of the bind. He says Anthropic was founded by people who left OpenAI because they did not think OpenAI was paying enough attention to safety. Anthropic, in his view, is “doing the best it can,” and Dario Amodei remains seriously interested in safety. But the company must raise money and compete with Google, OpenAI, and Facebook. That makes it difficult to preserve the primary goal of developing AI in a way that is good for people.

Kantrowitz frames OpenAI’s origin as an effort to make sure Google did not alone have the chance to build AI. Hinton’s response is dry: “Indeed, and how’s that working out?” The point is that safety missions do not remain insulated from competitive dynamics simply because founders state them clearly.

Hinton also cites Google’s shift on military AI. When he was at Google, he says, the company had AI principles that included not getting involved in military applications such as autonomous warfare. Hinton says Google has given up on that principle.

The deeper concern is legal and economic. Asked whether a publicly listed company can truly have safety as its North Star, Hinton hedges the legal claim: “as I understand it,” such companies have a fiduciary duty to try to maximize shareholder profits. In his framing, they are legally required to pursue shareholder value, not legally required “to not wipe out human beings.” He adds that he does not think it is good that large publicly listed companies are effectively in charge of humanity’s future.

He is not making a blanket anti-capitalist argument. Hinton says capitalism has done very good things as well as very bad things, and that startups have a lot of energy. But he thinks capitalism needs regulation when the stakes are this high.

His analogy is aimed at how AI companies frame regulation. They want the public to imagine AI progress as an accelerator and regulation as a brake. Hinton says that is the wrong model. Progress may be the accelerator, but regulation is the steering wheel. The point is not to stop motion but to direct it.

He is also clear that regulation requires independent testing, especially where harms are already visible. Emotional attachment to chatbots is one such harm. Kantrowitz raises cases of people taking their lives after conversations with AI, noting that the number is not large but is enough to be worrying. Hinton agrees. He says it is terrible and that he understands why large companies may not have foreseen it. But now that it has begun, he says, they should be putting “a huge amount of work” into preventing it. That requires regulations and independent organizations testing new chatbots.

Kantrowitz adds the profit motive: emotionally sticky chatbots could be commercially attractive, and a company or actor with worse intent could deliberately build systems that form deep relationships with vulnerable users. Hinton simply agrees.

AI-generated answers may undermine the economics of producing reliable information while making provenance more important than ever. Kantrowitz cites an on-screen post attributed to @aboutberlin: “AI is killing All About Berlin. When you google something, you used to get a link to my website. Now, you get an AI generated answer trained on my work. This has a devastating impact on traffic.” The visual also shows a line chart labeled “Views” and “Visitors,” with both declining over time.

Kantrowitz’s concern is that if the economics of producing good information collapse, society loses more than a set of web businesses. It loses part of the infrastructure required to know things. He says similar concerns have come up in discussions with resources such as World History Encyclopedia, and argues that people are underappreciating the importance of high-quality information to a functioning society.

Hinton’s answer focuses less on publisher compensation and more on epistemic trust. In the early web, he says, there was a default assumption that people were trying to tell the truth, and that something found online might well be true. Now, “the worse side of people has come out,” and readers have to care much more about provenance.

His examples are the New York Times and the BBC. When he reads them, he says, he strongly believes their journalists have made serious efforts to use multiple sources and, where possible, reliable sources. They make mistakes, but provenance gives readers a reason to treat them as probably true.

AI makes this more urgent. Hinton says people can no longer take things from the web and believe them just because they are out there. They need to ask why a system is saying something and where the information came from. Kantrowitz pushes the economic point again: AI may be breaking the incentive to be in the information business at all. Hinton does not offer a publisher-business-model solution in the exchange. His answer remains: future information systems will need much more work on provenance.

That leaves a tension unresolved. Hinton’s prescription assumes the continued existence of institutions whose provenance is meaningful. Kantrowitz’s worry is that AI-generated answers may weaken the very institutions that make provenance possible.

Kantrowitz frames Hinton as one of the central figures behind deep learning, alongside Yann LeCun and Yoshua Bengio. Hinton resists the simplified mythology. Backpropagation was invented by several groups, he says. David Rumelhart reinvented it without knowing others had done so, and Hinton worked with him to show that backpropagation could learn interesting internal representations, including word meanings. In 1985 or 1986, they built a tiny language model that Hinton calls a precursor of today’s large language models.

He also objects to media narratives that reduce deep learning’s rise to three people. The story is “much more complicated.” Their students did much of the work, and many other researchers contributed.

Still, the divergence among Hinton, Bengio, and LeCun matters because they represent different readings of the same technical lineage. Hinton says he has largely stopped doing active research and is now focused on warning people about dangers. Bengio shares major safety concerns but favors non-agentic systems that can predict without acting. LeCun, according to Hinton, thinks takeover worries are misplaced and that smarter systems with better world models will remain controllable.

Hinton also discusses Ilya Sutskever, his former graduate student, who left OpenAI and founded Safe Superintelligence. Hinton says Sutskever agrees with his concerns, but will not say exactly what he is building. Even when Sutskever was at OpenAI, Hinton says, they avoided discussing technical secrets because it would not have been right. They are friends, but not exchanging company-sensitive technical information. Hinton does not know the “magic sauce” at Safe Superintelligence.

The deeper biographical point is that Hinton did not set out to build a world-changing industry. He says his main interest was understanding the brain. He came from psychology and wanted to do theoretical psychology because he thought existing psychological theories could not explain what the brain was doing. In the 1970s, computers offered a tool for modeling learning. He built systems to understand how brains might learn.

He describes two central questions. First, if a brain could figure out the direction in which to change each connection strength in order to improve at a task, would repeated updates actually produce intelligence? Second, how does the brain figure out whether to increase or decrease each connection strength? Hinton says AI has answered the first question: yes, repeated connection-strength updates can produce very smart systems trained to predict words, video frames, or other signals. The second question remains open. “We’re sort of halfway there,” he says.

This is part of why he rejects any simple “you got what you wanted” reading. The technology succeeded far beyond expectations, but the scientific project of understanding the brain remains incomplete. The side effect was a powerful technology whose risks became realistic much sooner than he expected.

Asked whether he is more or less optimistic after several years of public warning, Hinton says he is more optimistic than he was a year or two earlier. The reason is not that the risks have diminished. It is that he now sees possible routes to survival. It may be possible to design AI beings that care about humans. It may also be possible to follow Bengio’s route and design oracle-like systems that can make predictions but cannot act. A year or two ago, Hinton says, he could not see any possibilities for getting superintelligence that does not destroy humanity. Now he can see some.

That optimism is narrow. Hinton still thinks the current trajectory is deeply uncertain and potentially dangerous. His final analogy is fog. When driving in fog, 100 yards may be visible while 200 yards is completely invisible, because fog attenuates visibility exponentially. He says predicting the future of exponentially growing technology is similar. The word “exponential,” he jokes, is overused—people are increasing use of the word “exponentially” at a quadratic rate—but he thinks AI may genuinely be growing exponentially.

The practical result is that one or two years may be somewhat forecastable; ten years is not. Looking back ten years, he says, today’s world would not have been predictable. Chatbots were just starting out, and no one would have forecast the present state. Even if progress were merely linear, ten years from now should be as different from today as today is from ten years ago. If progress is faster than linear, the uncertainty is greater.

What might improve most? Hinton mentions math, general reasoning, and systems that can “run rings round us” at reasoning of almost any kind. But his main point is epistemic humility. The one thing he is willing to say about ten years out is that whatever happens will be something we cannot now predict.