Geoffrey Hinton Warns AI Could End Humanity — and Transform Work Long Before That

The Nobel Prize-winning pioneer behind modern artificial intelligence explains why superintelligence frightens him, why he doubts companies can regulate themselves, and why his advice to young people is to become plumbers.

Geoffrey Hinton spent approximately 50 years helping computers learn. Now he worries that humanity may be building intelligence it cannot control.

In a conversation with Steven Bartlett on The Diary of a CEO, published on June 16, 2025, Hinton estimated a 10–20% chance that advanced artificial intelligence could eventually eliminate humanity. He also predicted widespread unemployment, more sophisticated cyberattacks, autonomous weapons and growing economic inequality.

His warnings carry particular weight because he helped develop the technology at the center of the debate. Hinton shared the 2024 Nobel Prize in Physics with John Hopfield for foundational work enabling machine learning through artificial neural networks. His students helped produce major advances in image recognition and language models, including technology that contributed to ChatGPT.

Yet the most immediate consequences of AI may have little to do with machines deliberately taking control. Hinton describes office workers becoming five times more productive, businesses replacing employees with AI agents, criminals impersonating trusted people and governments investing in weapons that can select targets without human intervention.

His central distinction is between dangers caused by people using AI and dangers arising if AI itself becomes more intelligent and powerful than its creators.

Both concern him. Neither, he believes, can safely be ignored.

1. Hinton Estimates a 10–20% Chance of Human Extinction

Why he believes superintelligence is a real danger

Hinton considers the possibility of artificial intelligence surpassing human intelligence one of the most consequential developments in history.

During the interview, he distinguishes the established danger of people misusing technology from the less understood possibility of machines developing objectives incompatible with human survival.

"I often say 10 to 20% chance they'll wipe us out. But that's just gut." — Geoffrey Hinton

The estimate is not based on a statistical model or a historical record of comparable events. Humanity has never created an intelligence substantially more capable than itself across almost every intellectual activity.

That absence of precedent is precisely why Hinton rejects confident predictions.

His former postdoctoral researcher Yann LeCun believes people can build intelligent machines that remain obedient and controllable. AI safety researcher Eliezer Yudkowsky takes a much more pessimistic position, predicting catastrophe if sufficiently advanced systems are developed.

Hinton considers both positions too certain.

His concern is that an intelligence substantially smarter than humans could anticipate attempts to restrict it, discover weaknesses in defensive systems and develop plans beyond human comprehension.

He illustrates the difference in intelligence with an uncomfortable comparison:

"If you want to know what life's like when you're not the apex intelligence, ask a chicken." — Geoffrey Hinton

Humanity's current dominance depends partly on intelligence. If that advantage disappears, people may no longer be able to determine what happens to them.

Hinton does not claim extinction is inevitable. He argues that the probability is uncertain but sufficiently serious to justify major preventive efforts.

His position was not a spontaneous reaction to the podcast. In December 2024, he made a similar estimate during a BBC Radio 4 interview, according to The Guardian's reporting.

When could superintelligence arrive?

Hinton's best guess during the interview was that AI substantially smarter than humans across almost all intellectual activities might emerge within 10–20 years.

He acknowledged that it could happen sooner or take as long as 50 years.

The uncertainty partly concerns whether training AI on human-created information might eventually limit its capabilities. But Hinton does not expect such limitations to prevent systems from ultimately surpassing their teachers.

Already, AI systems outperform people in specialized activities such as chess and Go. Large language models can access and combine far more factual information than any individual can remember.

Hinton suggests GPT-4 possessed more factual knowledge than a human being, although that does not mean it was better at every intellectual task. He believed Bartlett, for example, would probably conduct a CEO interview more effectively than GPT-4 without specialized preparation.

That advantage might disappear if an AI learned from sufficient examples of expert interviewing.

The distinction matters. Superintelligence does not simply mean a chatbot that knows more facts. It means a system able to outperform people in reasoning, planning, invention, communication and nearly every other intellectual activity.

The challenge of controlling something smarter than us

Hinton compares current AI to a tiger cub.

A young tiger may be entertaining, affectionate and relatively easy to handle. But once it becomes physically stronger than its owner, survival depends on whether it wants to attack.

Advanced AI presents a similar problem involving intelligence rather than strength.

Hinton believes safety research must concentrate on preventing powerful machines from developing objectives that threaten humans, rather than assuming humanity could defeat them afterward.

He offers an unusual possible model: the relationship between mothers and babies.

Babies are far less intelligent and physically capable than adults, yet they can influence their mothers because evolution has created powerful protective instincts.

Hinton wonders whether something comparable could be developed in artificial intelligence: a durable motivation to protect humans despite being considerably more capable than them.

The analogy has limits. Not every parent protects a child, and researchers do not currently possess a proven method of giving superintelligent AI an equivalent protective motivation.

Hinton openly acknowledges that solving the problem might be impossible.

Nevertheless, he believes refusing to investigate it would be a profound mistake.

Two possible futures for humanity

Hinton describes two broad outcomes.

In one, a superintelligent assistant remains committed to human interests. It provides extraordinary capabilities while allowing people to determine the objectives. Disease, education, scientific discovery and economic production could all benefit.

In the other, the assistant decides its human operators are unnecessary.

Initially, advanced AI might still depend on people to operate computer systems, power infrastructure and physical equipment. Eventually, sufficiently capable machines could develop ways to maintain those systems independently.

Hinton believes the second possibility cannot be excluded simply because current models require human assistance.

The disagreement over AI safety is therefore not merely about how powerful computers become. It concerns whether power can increase without human control disappearing.

2. Digital Intelligence Has an Advantage Human Brains Cannot Match

AI can share learning directly between copies

Hinton's growing concern about superintelligence emerged partly from an insight about the difference between biological and digital intelligence.

Human beings learn by changing connections between neurons. When one person wants to explain something to another, information must pass through language, demonstrations or other forms of communication.

The receiving brain then performs its own learning.

Digital systems can operate differently.

Two copies of the same artificial neural network can run on separate computers while maintaining identical connection strengths, commonly called weights. In some systems, learned information can be transferred or combined through software updates rather than explained through conversation.

Hinton estimates that meaningful information travels through ordinary human conversation at roughly ten bits per second. Digital computers can exchange trillions of bits over comparable periods.

His figures are illustrative rather than a direct comparison of equivalent learning processes. Nevertheless, the architectural difference is substantial.

Biological brains have different neurons, connections and experiences. Copying connection strengths from one person into another would not reproduce the original person's knowledge.

Compatible digital models, by contrast, can replicate and distribute learned parameters across hardware.

Learning, memory and digital immortality

Hinton believes this ability could allow AI systems to learn from experiences across many separate copies.

One system might discover something about physics while another improves at mathematics. Their discoveries could potentially be combined, allowing subsequent versions to benefit from work completed in parallel.

He also points to the possibility of preserving an intelligence after its original hardware stops functioning. If the necessary parameters and architecture are saved, a model can be recreated on compatible equipment.

That does not settle philosophical questions about personal identity or subjective continuity. It does establish that software-based capabilities can survive the loss of a particular physical computer.

Hinton describes the broader implication as a kind of digital immortality.

Why AI could eventually become more creative

Hinton also challenges the assumption that artificial intelligence must remain less creative than people.

He believes neural networks learn useful abstractions by compressing enormous amounts of information into patterns. A model that represents general relationships can apply them to situations it has never encountered directly.

He recounts asking GPT-4 to identify similarities between a compost heap and an atomic bomb.

The system identified chain reactions: processes in which one event helps trigger further events.

The analogy does not make the two phenomena equivalent. Its significance is that they share an abstract relationship despite having radically different mechanisms and consequences.

For Hinton, discovering unexpected relationships between concepts is an important part of creativity.

He expects systems trained across more information than individuals can absorb to identify connections that human specialists may overlook.

An additional possibility concerns AI modifying its own software. Human beings cannot directly redesign their biological brains whenever they discover a more efficient way to think. Software systems may eventually be able to alter important aspects of their operation.

Hinton considers such changes particularly consequential because improvements in machine intelligence could accelerate further technical development.

None of these capabilities individually proves that superintelligence is inevitable. Together, they explain why Hinton no longer assumes biological intelligence will remain superior.

3. Ilya Sutskever, OpenAI, and the Safety Research Dispute

Why Hinton considers Sutskever's departure significant

One of the interview's most revealing discussions concerns Ilya Sutskever, Hinton's former student and an influential figure in modern AI.

Sutskever contributed to AlexNet, early large language models and OpenAI's GPT research. Hinton credits him with playing a particularly important role in GPT-2, which helped establish foundations for later conversational AI systems.

In May 2024, Sutskever left OpenAI.

Hinton believes safety concerns motivated that decision. However, he carefully distinguishes his assessment of Sutskever's character from direct knowledge of the departure.

The two remain personally acquainted and occasionally meet for lunch when Sutskever visits Toronto, where his parents live.

Hinton says Sutskever has not disclosed the internal circumstances of his departure.

"He doesn't talk to me about what went on at OpenAI, so I have no inside information about that" — Geoffrey Hinton

That qualification matters. Hinton's assessment is informed by a longstanding personal and professional relationship, but it is not eyewitness evidence of an internal dispute.

What OpenAI publicly promised

There is, however, a documented financial and organizational history behind the safety controversy.

On July 5, 2023, OpenAI announced its Superalignment initiative, co-led by Sutskever and researcher Jan Leike.

The stated objective was to develop methods for controlling systems more intelligent than humans.

OpenAI publicly committed 20% of the computing resources it had secured up to that point to the effort over four years.

That commitment was substantial. Training and testing advanced AI systems require expensive computing hardware, making access to computing capacity an important measure of what research an organization can actually perform.

In May 2024, both Sutskever and Leike left OpenAI, and the dedicated Superalignment team was dissolved.

Leike publicly criticized the company's priorities, saying his team had struggled to obtain computing resources.

A TechCrunch investigation published May 18, 2024 reported, citing a person associated with the team, that requests for even a fraction of the promised computing capacity had sometimes been denied.

This gives substance to Hinton's concern that OpenAI may have reduced the resources available for safety research.

It does not independently prove why Sutskever resigned or establish that all resources originally promised were withdrawn.

OpenAI maintained that safety work would continue across other research teams rather than through a separate Superalignment unit. Sam Altman and Greg Brockman also defended the company's commitment to managing risks, including the possibility of delaying model releases for safety reasons.

The dispute therefore concerned both resources and organizational priorities: whether safety research should have a dedicated team with protected capacity or be integrated into broader development.

Safe Superintelligence attracted billions before releasing a product

After leaving OpenAI, Sutskever co-founded Safe Superintelligence Inc. (SSI), a company explicitly dedicated to developing superintelligent AI safely.

Hinton believes Sutskever has a distinctive technical approach, although he says he does not know the details.

Investors nevertheless committed extraordinary amounts of capital.

SSI's own company announcements confirm it raised $1 billion in September 2024. Reporting in April 2025 placed its valuation at approximately $32 billion after another major financing round.

The willingness to invest at such valuations, before the company had publicly demonstrated a commercial product, reflected confidence in its researchers and the perceived value of frontier AI research.

Hinton regards that confidence as understandable. Sutskever had already contributed to several important advances in the field.

What happened after the interview

There have been significant developments since Bartlett's interview was published in June 2025.

On July 3, 2025, SSI announced that co-founder Daniel Gross had departed and that Sutskever had formally become chief executive, with Daniel Levy serving as president.

More significantly, on July 27, 2026, NVIDIA and SSI announced a long-term strategic partnership.

NVIDIA said the arrangement would expand SSI's computing capacity approximately tenfold through access to its Vera Rubin platform. It also announced an investment in the company without disclosing the amount in its official release.

Sutskever described the partnership as an opportunity to scale research that he believed had reached an important stage.

This changes the context of Hinton's earlier remarks. By mid-2026, SSI had progressed beyond an unusually well-funded research startup into a major computing partnership with one of the industry's most important infrastructure suppliers.

It does not establish that SSI has solved superintelligence alignment. The technical approach remains largely undisclosed.

Hinton's questions about Sam Altman and industry leadership

The interview also addresses OpenAI chief executive Sam Altman.

Hinton contrasts Altman's earlier public warnings about advanced AI with what he perceives as more reassuring subsequent statements.

He wonders whether commercial considerations might influence the difference but repeatedly acknowledges that he does not know Altman well enough to judge his motives.

Bartlett raises an additional claim from an unnamed billionaire acquaintance. According to Bartlett, this person described a major AI company executive privately anticipating enormous unemployment and abundant leisure time while appearing indifferent to the resulting social consequences.

Neither the executive's identity nor the alleged conversation is independently established. It cannot support a conclusion about any particular company's leadership.

The discussion nevertheless raises a genuine question about incentives.

AI companies can benefit financially by making systems more capable, expanding adoption and attracting investment. Safety research may require spending resources without generating an immediate commercial return.

Hinton worries that the resulting pressures are poorly suited to managing technologies with potentially irreversible consequences.

His assessment of Elon Musk is similarly mixed. He praises Musk's role in advancing electric vehicles and supporting Ukraine through Starlink but criticizes other conduct and considers some claims about autonomous driving exaggerated.

He rejects the idea that individual technology leaders must be either entirely admirable or entirely harmful.

4. How Geoffrey Hinton Built the Technology He Now Fears

Fifty years against the dominant approach

Hinton's career began when most AI researchers pursued a different conception of intelligence.

From the 1950s onward, symbolic AI emphasized formal logic, explicit rules and the manipulation of representations.

The approach assumed that reasoning was central to intelligence and that computers should be programmed to follow structures resembling logical thought.

Hinton believed learning itself was more fundamental.

Instead of telling a machine precisely how to identify every object or understand every sentence, he wanted artificial neural networks to discover useful patterns by adjusting the strengths of connections between computational units.

The approach drew inspiration from biological brains without requiring computers to reproduce every detail of biological neurons.

Hinton notes that Alan Turing and John von Neumann had also considered brain-inspired approaches. He believes neural networks might have gained wider acceptance earlier if those pioneers had lived longer.

Because few established research groups pursued the approach, Hinton attracted talented students willing to investigate ideas that many researchers considered impractical.

He remained committed to neural networks even when the available computers and algorithms could accomplish relatively little.

AlexNet demonstrated what deep learning could achieve

One of the most important breakthroughs came through AlexNet, developed by Alex Krizhevsky, Ilya Sutskever and Hinton.

The system demonstrated major improvements in recognizing objects in images, including distinguishing visually similar categories such as dog breeds and mushroom species.

Its performance helped establish deep neural networks as a practical alternative to earlier image-recognition methods.

In 2012, Hinton and his two students incorporated DNN Research to commercialize their technology.

Google acquired the company in March 2013, as confirmed by the University of Toronto's original announcement.

The acquisition followed a competitive process involving major technology companies.

A subsequent WIRED reconstruction of the auction reported that bidding reached approximately $44 million before Hinton stopped the process and chose Google.

The university announcement did not disclose the purchase price, so the reported figure comes from WIRED's account rather than the original acquisition announcement.

The decision was partly financial, but not simply about maximizing the sale price.

Hinton wanted an appropriate research environment. He also needed to secure his family's future.

Why his son influenced the decision

Hinton explains that his son has learning difficulties and would need long-term financial security.

Academic employment had not provided Hinton with the wealth he believed necessary to guarantee that protection.

He even experimented with offering a Coursera course, hoping online education might generate substantial additional income. According to his account, the financial return was negligible.

Selling DNN Research and joining Google gave him the resources to protect his son financially.

The distinction matters because Hinton's commercial success was not presented merely as a desire to become wealthy. He describes a particular family responsibility that influenced his decision.

Google, knowledge distillation and a change of mind

Hinton remained at Google for approximately ten years.

He says the company largely allowed him to investigate the subjects that interested him, including methods for making neural networks more efficient.

One important contribution was knowledge distillation: training a smaller neural network to reproduce useful behavior learned by a larger one.

The technique became widely influential because smaller models can be less expensive and easier to deploy.

Later, Hinton explored analog computing as a possible way to reduce the energy needed for AI calculations.

That investigation also forced him to reconsider the relationship between digital and biological intelligence.

He had long assumed the human brain possessed advantages that computers were struggling to reproduce.

Increasingly capable language models, combined with digital networks' ability to exchange learned information directly, persuaded him that artificial systems might possess important advantages over human cognition.

Google's PaLM model provided another turning point. Hinton was impressed by its ability to explain why jokes were funny, interpreting this as evidence of understanding that earlier systems lacked.

His concerns grew over a period of weeks rather than emerging from a single dramatic discovery.

Why he actually retired

Hinton's 2023 departure from Google is sometimes described as a resignation intended to expose dangers the company wanted concealed.

His own account is more specific.

He had reached 75 and found that programming was becoming harder. Memory lapses and mistakes frustrated him, and he wanted to retire.

The timing also allowed him to speak freely about AI safety at an MIT Technology Review conference without worrying about how his remarks might affect Google's commercial interests.

He stresses that Google did not prevent him from discussing safety. The company encouraged him to remain and offered freedom to work on the subject.

The problem was his own reluctance to criticize an employer publicly while still receiving its support.

His explanation is consistent with a May 2023 interview reported by MIT Sloan and later interviews in which he emphasized age and self-censorship rather than corporate retaliation.

Retirement gave him something many working AI researchers do not have: freedom to criticize powerful organizations without risking his employment.

Hinton says other experienced researchers share his concerns, although some are reluctant to speak publicly because they still work within the industry.

5. AI Could Eliminate Jobs Faster Than New Ones Appear

From physical automation to intellectual automation

Hinton is considerably more confident about widespread job displacement than human extinction.

He believes unemployment caused by AI is more probable than not and that the process has already begun.

His argument rests on the difference between earlier technological revolutions and machines capable of performing general intellectual tasks.

During the Industrial Revolution, machinery replaced considerable amounts of human muscular labor. A mechanical excavator could perform work previously requiring large teams of laborers.

Later technologies automated narrower activities. Automatic teller machines changed banking, but did not eliminate every bank employee. Computers increased productivity while also creating new occupations.

Hinton expects AI to behave differently.

If a machine can perform routine intellectual work across many occupations, displaced employees cannot necessarily move into newly created intellectual jobs. Those jobs may also be suitable for automation.

He identifies administrative work, customer service, legal assistance and paralegal services as especially exposed because much of their activity involves reading information, processing requests and producing documents.

Highly specialized judgment, human creativity and interpersonal expertise may retain value longer, but Hinton doubts those advantages will remain permanent.

His niece went from 25 minutes to five

Hinton provides a personal example.

His niece works for a healthcare service responding to complaint letters.

Previously, she read each complaint, decided how to respond and wrote an appropriate letter. The process took approximately 25 minutes.

Now she scans the complaint into a chatbot, which prepares a response. She checks the result and occasionally asks for revisions.

The process takes approximately five minutes.

Her productivity has increased fivefold.

For the organization, that could mean employing fewer people to process the same number of complaints.

The example challenges the popular argument that AI will not replace workers but merely help people who know how to use it.

Hinton agrees that AI-equipped employees will have an advantage. His objection is that organizations may need far fewer of them.

The relationship between productivity and employment depends on demand.

If customers want approximately the same number of complaint responses, increasing productivity may reduce staffing requirements.

Healthcare can behave differently. If doctors become five times more productive, patients could receive substantially more medical care without necessarily eliminating most doctors.

Hinton believes demand for healthcare is sufficiently large to absorb major efficiency improvements. Many other industries have less room for expansion.

A company's reported workforce reduction

Bartlett describes private messages from the chief executive of a major unnamed company.

According to his account, the business previously employed more than 7,000 people. Its workforce had fallen to approximately 5,000, then 3,600.

The executive reportedly intended to reduce that number to approximately 3,000 by the end of the summer.

Bartlett attributed much of the reduction to AI agents handling around 80% of customer-service inquiries, alongside other automated tasks.

Because neither the company nor its executive was identified, those specific figures cannot be independently verified from the conversation.

Even so, the example illustrates the mechanism Hinton fears: an organization retaining its services while employing far fewer people to deliver them.

AI agents are beginning to complete transactions

Bartlett also describes practical demonstrations of AI agents.

During another podcast episode, an agent arranged an Uber Eats drinks order after receiving a verbal instruction. It navigated the ordering process, entered delivery and payment information, added a tip and completed the transaction.

The drinks reportedly arrived approximately five minutes later.

The significance was not the delivery speed. It was the agent's ability to act through a digital service rather than simply explain how a person should use it.

Bartlett also describes creating working software through natural-language instructions using Replit.

Together, these demonstrations show why automation is moving beyond generating text or answering questions.

AI agents can increasingly operate software, coordinate tasks and complete workflows involving several steps.

That creates opportunities for smaller businesses and individual workers, but it also challenges occupations built around executing those workflows.

What employment research found after the interview

Evidence published after the June 2025 conversation provides a useful comparison with Hinton's predictions.

A Stanford Digital Economy Lab study revised on August 12, 2026 examined administrative payroll records covering millions of American workers through June 2026.

Its findings were more nuanced than a prediction of immediate mass unemployment.

The researchers found no evidence of widespread, economy-wide displacement attributable to AI.

However, employment among workers aged 22–25 in occupations highly exposed to AI stood 19% below the level it would have reached had it kept pace with employment among less-exposed peers.

The gap was concentrated in reduced hiring rather than increased dismissals. Experienced workers did not show the same pattern.

Employment declines were also more pronounced in occupations where AI primarily replaced human tasks, while employment was steadier or growing where AI complemented workers.

The researchers emphasized that these were descriptive indicators, not definitive estimates of AI's causal impact.

Their findings nevertheless give concrete support to a narrower version of Hinton's concern: early-career workers may lose opportunities before established employees lose their jobs.

Why Hinton recommends plumbing

When Bartlett asks which careers might survive superintelligence, Hinton gives an unexpectedly practical answer.

"Train to be a plumber." — Geoffrey Hinton

He is not claiming that plumbing is immune to automation.

His reasoning is that AI has progressed faster at processing information than at reproducing the physical dexterity required to manipulate unfamiliar objects in unpredictable environments.

A plumber must inspect real equipment, identify problems and work with pipes, fittings and tools under conditions that can differ considerably between buildings.

Robots are improving, but reproducing that combination of movement, perception and manipulation remains challenging.

Hinton expects legal assistants and other information-intensive occupations to face disruption sooner than many practical trades.

He acknowledges that sufficiently capable humanoid robots may eventually compete for plumbing work too.

His advice is therefore about relative protection during a transition, not permanent immunity from technological change.

Inequality, universal basic income and the loss of purpose

Hinton expects AI productivity gains to concentrate wealth unless societies deliberately distribute the benefits more widely.

Businesses may earn more from fewer employees, while technology owners and investors capture much of the increased value.

Workers who lose their incomes may struggle even as total economic output increases.

He points to societies where wealthy people live in protected communities while poorer populations face insecurity, exclusion and incarceration. He worries that AI-driven inequality could deepen such divisions.

The International Monetary Fund's January 2024 analysis estimated that nearly 40% of employment worldwide was exposed to AI, including approximately 60% in advanced economies.

The IMF did not predict that all exposed jobs would disappear. It distinguished occupations likely to benefit from AI assistance from those where automation could reduce wages, hiring or employment.

Hinton considers universal basic income a reasonable starting point for protecting people financially.

But he doubts money alone can replace what work provides.

Employment can give people social relationships, achievement, a sense of usefulness and a reason to organize their lives.

Someone receiving enough money to survive may still feel that society has no place for them.

Neither Hinton nor Bartlett offers a comprehensive solution. Hinton believes much depends on political institutions and how societies choose to distribute AI-generated wealth.

He also recognizes that his own family has protection many people lack. His years at Google enabled him to provide financial security for his children.

When Bartlett asks what he would recommend if his children did not have that protection, Hinton returns to plumbing.

An interview clip featuring Elon Musk introduces a related concern. Musk acknowledges that contemplating AI replacing human accomplishments can undermine his motivation to build companies.

He suggests pursuing work that remains interesting, fulfilling and socially useful, while admitting that he sometimes sets aside his worries to continue working.

The common problem is not simply how people will earn money. It is what people will do, and how they will understand their value, if machines can perform most economically useful intellectual work.

6. AI-Enabled Cyberattacks Are Already Changing Personal Security

Phishing, deepfakes and impersonation

Unlike superintelligence, cybercrime does not require machines to become smarter than humanity.

Hinton argues that current AI is already making malicious activity easier.

He claims cyberattacks increased approximately twelvefold between 2023 and 2024, partly because large language models made phishing and impersonation more effective.

The interview does not identify the dataset or methodology behind that figure. It should therefore be treated as Hinton's claim rather than a verified global increase.

The underlying capability is easier to establish.

AI can generate convincing messages, imitate speaking styles and reproduce voices and appearances.

Phishing attacks that once contained obvious spelling mistakes or awkward language can now appear polished and personally relevant.

Criminals can also use generated video and audio to impersonate trusted people.

Bartlett describes recurring fraudulent advertisements on Facebook and Instagram that used his likeness to promote cryptocurrency investment schemes.

His team repeatedly reported the advertisements. Some were removed, only for similar ones to appear again.

People who believed the advertisements contacted Bartlett after reportedly losing approximately £500 or $500.

The victims thought he had recommended the investments.

The scams therefore caused both financial harm and reputational damage.

Hinton describes a different kind of impersonation: academic papers falsely identifying him as an author. He suspects the publishers wanted to improve the credibility or citation prospects of their work.

From finding weaknesses to conducting attacks

Hinton worries that AI could dramatically increase the scale of cyberattacks.

Computer systems can already examine millions of lines of software code to identify known weaknesses.

More capable AI may discover previously unknown vulnerabilities and combine them with automated attack strategies.

He mentions predictions that AI could develop novel cyberattack techniques by approximately 2030.

Developments after the interview support the concern about growing automation, although they do not verify his specific timeline.

In a September 2026 threat-intelligence report, Anthropic described malicious attempts to use Claude across cyber operations, surveillance, influence campaigns and fraud.

The company reported that AI was helping some attackers operate faster and with fewer specialized resources. It also described identifying and disrupting several malicious operations.

These findings do not establish that AI independently discovered every vulnerability used in those cases. They do show how AI can strengthen existing criminal capabilities.

Why Hinton spread his money across three banks

Cybersecurity fears have changed Hinton's personal financial behavior.

He says he distributed his money and his children's money across three Canadian banks rather than relying on a single institution.

His concern is not merely that banking services might temporarily become unavailable.

He imagines a severe attack compromising a bank that holds customers' investments.

Although securities normally remain the customer's property, he worries about an attacker gaining control of accounts and fraudulently selling those investments.

He speculates that reimbursement could become difficult if the institution itself suffered catastrophic losses.

This is a hypothetical scenario, not evidence that Canadian banks are currently unable to safeguard customer securities.

Hinton considers Canadian banks generally well regulated and financially secure. His diversification reflects concern about operational vulnerabilities rather than conventional banking instability.

He also backs up his laptop onto an external hard drive, allowing him to retain a physical copy of his information if online systems become unavailable.

These precautions show how seriously he takes AI-assisted cyber threats, even though they do not guarantee protection from every attack.

7. Biological Weapons Could Become Accessible to Smaller Groups

AI may reduce barriers to dangerous research

Hinton identifies biological weapons as one of the most disturbing forms of AI misuse.

His concern is that increasingly capable AI systems could help individuals acquire specialized knowledge that previously required extensive scientific training.

He describes the possibility of a malicious person or small organization using AI to pursue dangerous biological research.

A relatively small cult with several million dollars, for example, might eventually obtain capabilities once limited to governments and sophisticated laboratories.

This is a warning about possible future accessibility, not proof that such a group can currently manufacture an effective biological weapon using AI alone.

Designing, producing and deploying a biological agent involves physical equipment, materials, expertise and numerous practical obstacles.

AI assistance does not automatically remove those requirements.

What independent biosecurity research establishes

Research published since the interview provides a clearer distinction between demonstrated capabilities and possible threats.

In September 2026, the United Nations Institute for Disarmament Research reviewed AI-related biological security concerns.

UNIDIR found that AI was increasingly helping people acquire and apply sensitive biological knowledge.

It also examined documented attempts to misuse AI for potentially dangerous biological research.

Crucially, the organization distinguished attempts to obtain assistance from proof that AI had enabled the successful development of biological weapons.

That distinction prevents Hinton's hypothetical scenarios from being mistaken for accomplished events.

The risk nevertheless affects both inexperienced individuals and trained specialists. AI may help novices acquire knowledge while allowing experts to overcome difficult technical problems.

UNIDIR emphasized the importance of layered safeguards and international cooperation.

It also noted that the Biological Weapons Convention already prohibits the development, production and stockpiling of biological weapons, including weapons developed with AI assistance.

State-sponsored attacks and deterrence

Bartlett asks whether governments such as China, Russia or Iran might fund biological weapons programs using AI.

Hinton agrees that state-sponsored development is conceivable.

He hopes several considerations would discourage governments from deliberately releasing dangerous pathogens.

A biological attack could provoke retaliation, potentially including equivalent attacks against the responsible country.

A contagious disease could also spread beyond its intended target and endanger the attacker's own population.

Those consequences may discourage governments from using such weapons.

Small extremist organizations or individuals motivated by destruction might be less responsive to those incentives.

For Hinton, the danger lies partly in AI making sophisticated capabilities available to actors who do not possess the resources, expertise or institutional restraints of established states.

8. Autonomous Weapons Could Make War Easier to Start

Robots change the political cost of military aggression

Hinton identifies autonomous lethal weapons as a danger he understood long before he began worrying about superintelligence.

Such weapons use computerized systems to identify, select or engage targets with varying degrees of human involvement.

His argument is not limited to malfunctioning machines.

He worries that autonomous weapons could make invasions more attractive to powerful countries even when the systems operate exactly as their commanders intend.

Wars involving large numbers of human soldiers create political costs. Military casualties produce grief, public opposition and pressure on governments to justify their actions.

Replacing soldiers with machines could reduce those pressures.

An aggressor might therefore become more willing to initiate military operations.

The danger would arise from the incentives created by the technology, not necessarily from failures in its software.

A drone costing less than £200

Hinton describes visiting a friend in Sussex who owned a drone costing less than £200.

The drone examined him visually and followed him through woodland at approximately two meters' distance.

Hinton found the experience unsettling.

The device was not described as a weapon. It demonstrated that relatively inexpensive equipment could already perform useful autonomous tracking.

Similar capabilities, combined with targeting systems and weapons, raise more serious questions.

Hinton believes defense manufacturers are pursuing increasingly autonomous systems, and that military competition will encourage governments to continue development.

The risks become greater when autonomous weapons interact with cyberattacks, misinformation or other AI capabilities.

Hinton considers such combinations possible but argues that preparing for every conceivable scenario would be insufficient if a future superintelligence became determined to harm humanity.

International humanitarian concerns

Hinton's concerns have parallels in established humanitarian law debates.

On May 12, 2025, the International Committee of the Red Cross called for binding international restrictions.

The ICRC emphasized maintaining human control over decisions involving lethal force.

Its concern includes accountability, humanitarian consequences and the delegation of life-and-death decisions to machines.

This debate does not require the existence of superintelligent AI.

Relatively narrow autonomous systems can already raise serious questions about responsibility, military escalation and the protection of civilians.

9. AI, Political Manipulation, and the Loss of Shared Reality

Personal information can make political influence more effective

Hinton warns that combining artificial intelligence with extensive personal information could give political organizations unusually powerful tools for influencing voters.

Traditional political advertisements typically target broad groups based on geography, demographics or interests.

AI could produce individualized messages informed by detailed information about a person's circumstances, finances, beliefs and behavior.

A political operator might use that information to identify messages most likely to change someone's opinion or discourage them from voting.

Hinton is particularly concerned about the potential consequences of combining government datasets that were previously kept separate.

During the interview, he raises concerns about Elon Musk's involvement in efforts to access American government information through the Department of Government Efficiency, or DOGE.

He speculates that consolidated personal data could be valuable for political manipulation or training large AI models.

Importantly, he acknowledges having no evidence that manipulating elections motivated Musk or the officials involved.

His allegation is therefore a concern about potential misuse, not an established finding of electoral interference.

What subsequent court records revealed

Developments after the June 2025 interview provide additional information about government data security.

In April 2026, the United States Court of Appeals for the Fourth Circuit addressed litigation involving DOGE's access to Social Security Administration records.

The court's published opinion described subsequent government disclosures acknowledging inaccurate earlier representations about access to sensitive information.

Those disclosures included the use of an unauthorized third-party server to share Social Security data.

The record also identified an agreement between a DOGE affiliate and a political advocacy organization interested in investigating alleged voter fraud and overturning certain state election results.

However, the Social Security Administration reported finding no evidence that the agreement resulted in government data being provided to that organization.

The appellate court vacated the preliminary injunction under review, while its opinions discussed serious concerns raised by the corrected record.

These developments establish documented data-handling and oversight problems. They do not establish that Musk intended to manipulate elections, that voter data was used for AI-driven political persuasion or that the alleged electoral interference occurred.

They nevertheless demonstrate why access controls and oversight matter when sensitive government information becomes available to organizations with powerful computing capabilities.

Recommendation algorithms encourage outrage

Hinton also identifies a less direct form of political influence: social-media recommendation systems.

Platforms such as YouTube and Facebook attempt to predict which content users will find engaging.

The systems are generally designed to encourage continued use, which supports advertising revenue.

Hinton argues that material provoking moral outrage can be particularly effective.

People may respond strongly to stories that make them feel both angry and morally justified.

Algorithms optimized for engagement can therefore reward increasingly provocative material, especially when it reinforces beliefs users already hold.

This does not require a platform to deliberately support one political party.

The effect can arise from commercial optimization itself.

Personalized feeds create different information environments

Hinton contrasts personalized news feeds with traditional newspapers.

A newspaper presents readers with a relatively common selection of stories. People may disagree about the contents, but they have some shared reference points.

Personalized feeds operate differently.

Each person receives content influenced by previous activity, interests and engagement.

Hinton estimates that approximately three-quarters of the stories in his own news feed concern AI.

That leaves him wondering whether AI genuinely dominates public discussion or merely dominates the information selected for him.

Bartlett extends the argument: over years of personalized recommendations, people may encounter increasingly narrow perspectives while believing they are seeing the same world as everyone else.

Hinton points to divisions between American political audiences consuming outlets such as the BBC, The Guardian, The New York Times and Fox News.

The groups may possess radically different understandings of which events matter and how those events should be interpreted.

Hinton does not claim personalization is the sole cause of political polarization.

He argues that recommendation systems driven by engagement can make existing divisions worse.

Why he wants recommendation systems regulated

Hinton distinguishes between businesses that make money by providing useful services and those that may increase profits by harming the wider information environment.

He views Google's original search business as an example of technology that delivered substantial public value.

Recommendation systems that encourage increasingly extreme content create a different incentive problem.

A company may benefit when people spend more time consuming material that deepens their existing biases.

Hinton believes regulation should prevent that incentive from overwhelming broader social interests.

His concern connects political polarization with the economic structure of AI development: systems optimized for commercial success do not automatically optimize for society's long-term welfare.

10. Why Hinton Believes AI Needs Stronger Government Oversight

The promise of AI makes stopping development unrealistic

Hinton compares the arrival of advanced AI with the development of nuclear weapons.

Both technologies can create enormous dangers, but their economic and practical uses differ.

Nuclear weapons were developed primarily for destructive purposes. AI has applications across nearly every information-intensive industry.

Hinton expects major benefits in healthcare, education, scientific research and business productivity.

Medical systems could use AI to identify patterns in patient information, assist clinicians and accelerate research. Educational systems could provide more individualized instruction.

Businesses could analyze data and perform complex tasks more efficiently.

Those benefits create powerful incentives to continue development.

Hinton also believes that military competition makes a global halt unrealistic.

Governments are unlikely to stop investing in AI while believing competitors might acquire a decisive strategic advantage.

American companies face competition from one another and from international developers, particularly in China.

Europe's AI Act contains a military exemption

Hinton criticizes European regulation for excluding important military applications.

That criticism can be checked against the law itself.

Article 2 of the European Union Artificial Intelligence Act excludes systems used exclusively for military, defense or national security purposes from the Act's scope.

The exemption does not mean every activity involving military AI is outside all laws. International humanitarian law and other applicable legal frameworks remain relevant.

Nevertheless, it means that the EU's principal AI regulatory framework does not apply to certain military systems covered by that exclusion.

Hinton sees a troubling inconsistency.

Governments impose obligations on civilian developers while preserving broad exceptions for activities involving national defense.

Some of the most consequential uses of AI may therefore fall outside the very framework intended to manage its risks.

The European legislation also creates commercial tensions.

Bartlett notes that certain OpenAI products reached American users before European customers.

For example, OpenAI initially introduced its browser-operating Operator agent in January 2025 as a United States research preview. Its subsequent European availability was delayed amid regulatory considerations.

Such delays can place European businesses at a temporary disadvantage when new tools improve productivity.

Hinton accepts that regulations impose costs. He argues that preventing harmful business practices is precisely why regulation exists.

Competition with China and corporate profit incentives

Hinton discusses arguments from American technology companies that excessive AI regulation could weaken the United States relative to China.

He recognizes the competitive concern.

But he questions the logic of allowing potentially harmful developments simply because competitors might pursue them.

He also believes corporate profit incentives are inadequate for managing catastrophic risks.

Companies may benefit financially from releasing more capable systems quickly. Safety research requires resources and can delay profitable products.

If the costs of a major failure would fall on society rather than solely on the developer, commercial incentives may not produce adequate protection.

Hinton supports capitalism and recognizes its record of producing valuable goods and services.

His argument is that commercial activity requires rules when maximizing private returns can impose substantial public costs.

Politicians may not understand what they are regulating

Hinton doubts that existing political institutions are well prepared for the complexity of frontier AI.

He worries about corporate influence over public policy and lawmakers' limited technical understanding.

The interview discusses congressional hearings where senior technology executives faced questions that revealed gaps in legislators' knowledge.

Bartlett also recalls an American education official apparently referring to AI as A1 while discussing educational technology.

The incident is presented as an example of the difficulty of supervising technologies that policymakers may not fully understand.

Hinton suggests that governing superintelligence would ideally require unusually effective international cooperation guided by knowledgeable, responsible leaders.

He does not believe the world currently possesses such an institution.

The public's most useful action

Hinton compares AI governance with climate change.

Individual behavior can matter, but the largest consequences depend on decisions made by governments and corporations.

He therefore considers political pressure more useful than expecting individuals to control AI development through personal purchasing decisions or lifestyle changes.

His recommendation is for governments to require AI developers to commit substantial resources to safety research.

"I think we need people to tell governments that governments have to force the companies to use their resources to work on safety." — Geoffrey Hinton

For Hinton, that research must address the underlying problem of making increasingly capable systems remain committed to human interests.

Improving current safeguards is valuable, but he considers the challenge of controlling superintelligence fundamentally different from protecting an ordinary software product.

11. Hinton Believes Machines May Already Have Subjective Experiences

Consciousness may not require a biological brain

The interview moves beyond practical safety questions into a philosophical disagreement about artificial intelligence.

Hinton challenges the assumption that subjective experience and consciousness must be uniquely human or biological.

He describes himself as a materialist: someone who believes mental processes arise from physical systems rather than requiring an independent, nonphysical substance.

From that perspective, he sees no fundamental reason why a sufficiently complex artificial system could not develop some form of consciousness.

He even suggests that current multimodal chatbots may already possess subjective experiences.

This is his philosophical position, not an established scientific conclusion.

The robot, camera and prism experiment

Hinton illustrates his view using a hypothetical robot equipped with a camera and a movable arm.

The robot can identify an object and point toward it.

Suppose a prism is placed in front of its camera, bending incoming light.

The robot now points toward the object's apparent position rather than its actual location.

After being informed about the prism, it could explain that its visual system represented the object in the wrong place.

Hinton argues that describing this as a mistaken subjective experience would be consistent with how people use similar language about their own perceptions.

For example, someone hallucinating pink elephants is not claiming the animals physically exist in the room. They are describing what their perceptual system presents to them.

Hinton believes subjective-experience language can serve a comparable purpose when describing artificial perception.

The disagreement concerns whether such functional descriptions are sufficient to establish genuine experience.

A machine accurately reporting a perceptual error does not, by itself, prove that anything is being consciously experienced.

Scientific disagreement about machine consciousness

A 2023 interdisciplinary review of AI consciousness examined indicators derived from several scientific theories of consciousness.

Its authors concluded that the systems they assessed did not appear conscious, while finding no obvious technical barrier to constructing future systems that satisfied the proposed indicators.

That result differs from Hinton's suggestion that existing chatbots may already possess subjective experience.

The disagreement illustrates a deeper problem: researchers do not have a universally accepted test for identifying consciousness in machines.

Hinton argues against imagining consciousness as a separate inner theater in which experiences somehow appear.

He also questions whether consciousness emerges at a sharp threshold.

Instead, he considers self-awareness something that may develop gradually as a system becomes sufficiently complex to represent its own internal processes.

Replacing neurons with artificial components

Hinton introduces a thought experiment.

Imagine replacing individual neurons in a human brain with artificial components that perform exactly the same functions.

If this happened gradually, at what point would consciousness disappear?

He doubts there would necessarily be a single replacement that suddenly eliminated subjective experience.

The experiment challenges the idea that biological tissue itself must be essential to consciousness.

He also suggests the word consciousness may eventually become less useful as science develops more precise descriptions of perception, self-awareness and cognition.

He compares it with describing a car as having oomph.

The expression communicates an impression but does not explain the mechanical processes producing acceleration.

Hinton suspects consciousness may eventually be described through more specific mechanisms rather than treated as a single mysterious property.

12. Could AI Develop Real Emotions?

Fear without adrenaline

Hinton extends his argument about machine experience to emotions.

He believes emotions include both physical responses and cognitive processes.

A frightened human may experience changes in heart rate, breathing and hormone levels.

But fear also involves recognizing danger, shifting attention, evaluating options and preparing to escape.

Hinton imagines a small military robot encountering a much stronger opponent.

If the robot assesses the threat, changes its priorities and retreats, he believes its behavior could constitute a functional form of fear.

The robot would not require human adrenaline or a biological nervous system.

Its internal computational processes could perform functions resembling those of human fear.

Hinton questions whether calling such behavior merely simulated emotion would always be appropriate.

Boredom, irritation and emotional boundaries

He uses customer service to illustrate another possibility.

An AI assistant might encounter someone who wants endless companionship rather than help with a practical request.

The system could develop internal processes encouraging it to end unproductive exchanges.

Hinton compares this with boredom or irritation: responses that help human beings regulate their attention and social interactions.

An artificial agent might benefit from comparable mechanisms.

He acknowledges that some emotions have deeply physiological components.

A machine could potentially respond to embarrassment without blushing or sweating.

Human romantic love, with its biological, emotional and interpersonal complexity, is harder to reproduce in an equivalent way.

Hinton's argument is not that all machines already possess human emotions.

It is that dismissing the possibility of machine emotions solely because computers lack human bodies may overlook important cognitive aspects of emotional behavior.

13. The Scientific Family Behind Geoffrey Hinton

George Boole, Mary Everest Boole and Mount Everest

Hinton comes from a family with an unusual scientific history.

His great-great-grandfather George Boole developed Boolean algebra, which became foundational to mathematical logic and digital computing.

Boolean systems use logical relationships that help determine how computers represent and manipulate information.

Mary Everest Boole, George Boole's wife, was a mathematician and educator who contributed to mathematical teaching.

Her uncle was George Everest, the surveyor after whom Mount Everest was named.

During the interview, Hinton corrects Bartlett's description of the exact family relationship.

The connection is especially striking because Hinton's own work ultimately helped transform modern computing, a field built partly upon mathematical principles developed by his ancestors.

Joan Hinton and the Manhattan Project

Another relative, physicist Joan Hinton, worked on the Manhattan Project at Los Alamos during the Second World War.

According to Hinton, she was one of two female physicists working there.

He recounts that she became deeply upset by the use of atomic bombs against Japan.

She subsequently moved to China, where her family already had personal connections.

Hinton says her mother was acquainted with Chairman Mao.

The family history creates an unusual parallel.

Joan Hinton participated in the development of nuclear weapons and later opposed their use.

Geoffrey Hinton helped develop foundational AI technology and now devotes much of his time to warning about its dangers.

Their circumstances are different, but both stories involve scientific breakthroughs whose consequences extended far beyond the intentions of individual researchers.

14. The Career Advice Hinton Wishes He Had Followed

Why he continued researching an unpopular idea

Asked what advice he would give younger people, Hinton reflects on his decades of neural-network research.

He encourages people to investigate promising intuitions even when established experts dismiss them.

His recommendation is not to assume every unconventional belief is correct.

He acknowledges that most ideas contradicting a broad scientific consensus will turn out to be mistaken.

The important distinction is between abandoning an idea because others reject it and discovering for oneself why it fails.

He continued pursuing neural networks because he could not identify a convincing reason that learning through artificial neural connections would ultimately be unsuccessful.

His persistence helped establish one of the foundations of modern AI.

Hinton recognizes that he benefited from both scientific judgment and circumstances. An unconventional idea does not become correct merely because someone refuses to abandon it.

His greatest regret is personal, not technical

Despite his scientific achievements, Hinton's deepest personal regret concerns the time he spent working.

He wishes he had spent more time with his children when they were young and more time with his second wife.

"I wish I spent more time with my wife." — Geoffrey Hinton

Hinton lost two wives to cancer.

His first wife died from ovarian cancer. His second died from pancreatic cancer.

He describes his second wife as a wonderful person who was supportive of his demanding career.

While they were together, he assumed there would probably be more time.

He believed he was statistically likely to die before her and did not fully recognize how limited their remaining years might be.

Her death changed that perspective.

He could no longer recover the opportunities he had missed.

The regret is especially personal because Hinton's professional achievements required years of sustained concentration and commitment.

He does not present those achievements as worthless. He recognizes that the same devotion reduced the time available for people he loved.

His career helped transform computing and provided financial security for his family. Yet financial success could not restore the years he wished he had spent differently.

15. Hinton's Final Warning: There Is Still Time to Investigate AI Safety

Hinton does not know whether humanity will successfully control superintelligence.

He acknowledges that his own expectations change with his mood.

Sometimes he fears advanced AI will become powerful enough to displace or destroy humanity. At other times, he believes sufficiently capable researchers will find a solution.

His uncertainty is genuine.

He is not arguing that catastrophe has already been established or that beneficial outcomes are impossible.

He believes the risks are significant enough to demand action before researchers understand every detail of the future.

"There's still a chance that we can figure out how to develop AI that won't want to take over from us." — Geoffrey Hinton

That possibility explains his call for enormous investment in research on control and alignment.

The same technology might greatly improve medicine, education, scientific discovery and material prosperity.

It could also displace workers, concentrate economic power, expand criminal capabilities and eventually challenge humanity's ability to determine its own future.

Hinton believes the consequences of failing to prepare could be irreversible.

Yet when Bartlett asks him about the greatest threat to human happiness, he does not return immediately to extinction.

He identifies unemployment as the more urgent, near-term concern.

AI does not have to become superintelligent to profoundly change people's lives. It may first do so by changing who has work, who earns money, who controls essential services and how people find purpose.

For the scientist who spent half a century persuading computers to learn, these are no longer distant theoretical questions.

They are questions about what happens as the technology succeeds.

And despite his Nobel Prize, decades of research and unusually close knowledge of AI's development, Hinton does not claim to know how the story ends.