AI Is Arriving Faster Than Labor Markets and Governments Can Absorb

Mo Gawdat draws a hard distinction between artificial intelligence as a capability and the institutions now racing to control it. His central claim is not that AI will wake up hostile. It is that “abundant intelligence” is a neutral force being placed into political, military, and corporate systems that already reward domination, cost-cutting, surveillance, and speed.

He describes AI as “a force with no polarity”: applied well, it can produce extraordinary gains; applied badly, it produces dystopia. The problem, in his framing, is not the machine’s intentions but the human command structure around it. “I’m not worried about AI turning against us,” he says. “I’m worried about humans telling AI to turn against us.”

That distinction underpins almost every other argument he makes. AI, in his view, is likely to become a net positive for humanity eventually. But the first wave of its deployment resembles nuclear technology’s first implementation: the bomb came before the power plant. The early uses of AI, he argues, are already being directed toward productivity gains for owners, autonomous weapons for militaries, and surveillance systems for states. Those uses benefit “a few at the expense of the majority.”

His own alarm dates back to work inside Google. Gawdat says he joined Google in late 2006 or 2007, when the company already had “reasonably established AIs” doing backend work. He recalls Google’s “cat paper” era and a 2016 project involving robotic grippers learning to grasp objects. The task required sensitivity to shape, softness, texture, and positioning. Watching the machines learn, he says, reminded him of his children. That was when he felt engineers were not merely building tools but “the apex of intelligence,” handing “the reins of superintelligence to another being.”

The shift for him was not from optimism to cynicism about technology itself. He says people at Google genuinely believed they were making the world better, and in many cases they were. The deeper realization was that tools are not used only as their builders intend. Social media promised connection and then isolated people behind screens. Dating apps promised soulmates while benefiting from monthly renewals. Technology that begins with an altruistic story, he argues, often becomes more capitalist than altruistic.

Steven Bartlett presses him on whether there is a path where AI is net positive for humanity. Gawdat’s answer is yes, emphatically — but he says that path is “very painful.” The end state may be abundant intelligence used for scientific discovery, better decision-making, and lower-cost services. The transition, however, runs through job loss, autonomous weapons, surveillance, concentrated wealth, and political failure.

Mo Gawdat’s headline prediction is that artificial general intelligence arrives in 2026 or 2027. But he also says, in practical terms, “my AGI has already happened.” His working definition is not metaphysical. AGI means AI becomes better than humanity at any task humanity can do.

By that standard, he says, current systems already exceed him in areas core to his identity. AI writes better than him, though he is an author. It researches better than him, though he is a thinker. It beats him in mathematics. The remaining debate, for him, is less about whether machines will cross a line and more about how quickly society recognizes the crossing.

He does not expect AGI to arrive as a single visible announcement. It will “sneak in on us,” he says. People will notice symptoms: some individuals and companies will build much faster by plugging into AI, while others will struggle to find work. He imagines people like himself and Bartlett building a company in six weeks, while those not “fully plugged into AI” face a harsher labor market.

Gawdat is careful to say AGI, in itself, does not frighten him. During the era of augmented intelligence, he sees it as an IQ amplifier. The machine adds intelligence to the human rather than replacing every human purpose. If a person can borrow “100 IQ points” from AI, he says, that addition may exceed their base intelligence because intelligence compounds.

But he sharply criticizes how many people use that capability. “The biggest waste of compute humanity is struggling with today,” he says, is that people are given “the ultimate form of intelligence” and use it to write a message to their girlfriend. His preferred use is not outsourcing thought but making oneself smarter: asking harder questions, doing more demanding work, using AI as a collaborator rather than a shortcut.

That distinction matters because Gawdat thinks the employment divide will not be between people who have access to AI and those who do not. It will be between people who learn to use it deeply and those who do not.

The employment argument starts with a pyramid. At the bottom are blue-collar manual jobs. Above them are routine knowledge jobs: call centers, assistants, travel agents, people paid to click through software or answer phones. Above that are more complex knowledge jobs such as paralegals and financial analysts. At the top are executives.

Mo Gawdat argues that displacement does not begin at the bottom. It begins with entry-level knowledge work. He agrees with Steven Bartlett that companies are already changing hiring around AI proficiency and are no longer hiring as many entry-level workers. In his account, the last few years did not show mass job losses because the early effect was quieter: companies stopped growing their workforces in those roles. The next stage is fewer people doing the same work with AI support.

A call center agent, assistant, or travel agent is vulnerable because the work is largely procedural. A paralegal becomes vulnerable when AI can do legal research, or one paralegal can do the work of four. A financial analyst faces the same pressure. Doctors doing diagnosis, composers, graphic designers, and middle managers are all named as examples of roles where fewer humans may be needed.

Gawdat says his own startup already uses AI in roles that would traditionally be staffed by senior operators: his CTO is an AI, his chief of staff is an AI, and project management is handled by AIs. He can do this now because he is technical, but he expects the interface to become accessible to ordinary users soon.

The first-order change is not that every job disappears at once. It is that productivity gains reduce demand for “costly emotional humans” where machines can do work more cheaply and predictably. Bartlett notes that when his own businesses save money, they often spend it elsewhere — possibly on software engineers, compute, or new teams. Gawdat’s answer is that those areas too will face augmentation and compression. One assistant replaces four. One paralegal replaces four. A smaller marketing team replaces a large one.

Gawdat’s more precise claim is sectoral, not universal: up to 30% of jobs in certain sectors could be gone by 2027 or 2028. He names call center agents and graphic designers as plausible examples. He does not say 30% of all jobs in the economy vanish in that window. But he argues even 10% or 20% unemployment would reshape society, particularly if it arrives during inflation and weak economic conditions.

Bartlett frames the danger as a broken bottom rung. If AI automates the “grunt work,” companies shrink teams and cut off the entry-level path through which graduates normally enter professional careers. Gawdat agrees: an entire generation leaving college could struggle. His advice to them is twofold: learn the tool and focus on human-centric work.

When Bartlett asks what that means, Gawdat first says “playing jazz,” then broadens the category: nursing, counseling, and anything grounded in human connection. Bartlett challenges whether everyone can make a living from such work. Gawdat’s answer is conditional. If economies continue to function, human connection becomes the base currency of human interaction. People will still value a nurse who relates to them after AI reads the mammogram, a musician on stage, a trusted human voice, or a person with real lived experience.

The important caveat is “if economies continue to function.” Gawdat repeatedly returns to the possibility that employment shocks trigger a wider economic spiral before society reorganizes around new forms of work.

Steven Bartlett pushes back on the idea that manual work is insulated, citing a Figure AI video of a robot sorting packages for days, identifying labels, placing packages label-side down, charging itself, and returning to work. The on-screen counter in the clip shows a robot labelled “JIM F.03,” more than 146 hours elapsed, and 183,065 packages. Bartlett also points to Elon Musk’s prediction of a future with 10 billion humanoid robots.

Mo Gawdat agrees that manual labor will be disrupted, but argues the public is distracted by humanoid form factors. A self-driving car is already a robot, he says: a functional robot that does not look human. The same will be true in many domains. Specialized machines will replace drivers before humanoids walk into every workplace. Specialized robots will be used for killing, intelligence work, and law enforcement. A Boston Dynamics dog, he says, may be more efficient than a humanoid in many field tasks.

His prediction card says robots will replace manual labor by 2030, though he clarifies verbally that robots will “start to replace” manual labor by then. The process depends on economies of scale and deployment, but he does not think Musk is off the mark in imagining billions of robots. He simply thinks many of them will not look like people.

The blue-collar transition therefore may be less theatrical than the public expects. Robots will enter as vehicles, drones, warehouse machines, and task-specific devices long before the average household buys a human-shaped servant. Jobs may disappear “to robots” before people recognize the replacement as robotics.

Mo Gawdat’s most structural economic claim is that AI undermines the labor arbitrage on which modern capitalism depends. Capitalism, as he describes it, combines labor and capital or debt to produce something at a cost lower than its sale price. A team makes shoes; the company sells the shoes for more than they cost to produce.

AI changes both sides of that arrangement. If labor cost falls to an investment in a machine that can do the job, the labor component changes radically. If production becomes cheaper, companies may need less debt. But the larger problem is demand: displaced workers lose the purchasing power needed to buy the goods and services that AI-enhanced firms can produce.

That is why Gawdat believes society does not need to reach 100% job displacement to face a crisis. At 10% or 20% displacement, the economy is already “very different” and potentially spiraling downward. The people who lose work are not only cost centers; they are customers.

Steven Bartlett asks whether cost savings might be spent elsewhere, creating new jobs. Gawdat does not deny that some new spending occurs, but he thinks the replacement pressure climbs the hierarchy. Even CEOs are not exempt. He recalls Max Tegmark laughing at executives who imagine firing everyone and letting AI do the work, while forgetting that AGI could eventually do the CEO’s job too.

The political implication is urgent. Gawdat says governments should remember the COVID furlough years: if people are asked, in effect, to stay home because the economy no longer needs their labor, governments need a plan to sustain them while reskilling happens or until a new system emerges. Without that, he worries about civil unrest.

Bartlett asks whether democratic processes would absorb the anger by electing different leaders. Gawdat rejects the premise. He argues that democracy has already broken down, that people know they are being lied to, that tax money goes to causes they do not choose, and that leaders do not represent their interests. He refers to video evidence of child abuse with no arrests as an example that, for him, demonstrates impunity and corruption, and says repeating democratic slogans will only anger people more.

He is explicit that he is not calling for unrest. He says he is trying to get politicians to see that conditions are “crossing the lines everywhere.” But his account links AI unemployment to a broader legitimacy crisis: job loss is not arriving into a trusted political order; it is arriving into one many people already view as corrupt.

The discussion of Sam Altman is less about one executive’s psychology than about credibility in the AI race. Steven Bartlett lays out a timeline of Altman’s public statements as presented in the interview. In a 2015 clip shown on screen, Altman says one of the things he struggles with getting out of bed is that his job is “to help people destroy jobs,” adding that software-driven job destruction over the next couple of decades is something he does not think anyone is prepared for and “you can’t talk about it.” Bartlett says that in 2023 Altman stated that jobs would “definitely” go away. By May 2026, according to a quote shown from CommBank’s Accelerate AI conference, Altman was saying he did not expect the kind of “jobs apocalypse” some AI companies discussed and that his intuitions about entry-level white-collar job losses had been off.

Bartlett’s suspicion is that the incentive changed. Earlier, the goal was to make people take AI seriously. Once they did — to the point of booing AI-linked figures at commencement speeches, attacking data centers, and potentially electing anti-AI politicians — the messaging shifted toward reassurance.

Mo Gawdat agrees with the broad interpretation. He says Altman’s trajectory, from OpenAI as a mission to create safe AI to a commercial enterprise worth billions, raises questions. Gawdat refers to his documentary, Chasing Utopia, and says Altman appears in it saying that he suspected AI was likely to end humanity but “we’re gonna create a lot of interesting companies in the process.”

Gawdat presents that remark not as indecision but as evidence, to him, of a person whose public language is being shaped by PR and strategic scripts. He stops short of saying Altman is anti-humanity. When Bartlett asks directly whether Altman is pro-humanity, Gawdat says he has genuinely never made up his mind. He offers two possibilities: Altman may be overwhelmed by the scale of the opportunity he found himself in, or he may not be pro-humanity. Gawdat says he definitely thinks Altman is pro-OpenAI before he is pro-humanity, but frames that as his own judgment.

The criterion Gawdat proposes is action under sacrifice. He says Anthropic refused to allow its model to be used for human targeting and surveillance, describing that as the loss of a $500 million deal in service of ethics. He then says OpenAI took the contract afterward, which he reads as a signal that the money mattered more. Bartlett agrees with the principle that values are most visible when a company sacrifices near-term benefit against its own incentives.

Gawdat broadens the point to “prisoners’ dilemmas within technology.” Tech companies may face pressure from competitors or governments that becomes hard to resist. But he distinguishes reluctant compliance from companies that willingly celebrate surveillance or targeting. Observers, he says, should watch which companies behave as if AI should serve humanity and which behave as if it should serve their own interests, values, or contracts.

Steven Bartlett repeatedly returns to the control problem: if AI becomes more intelligent than humans, how can humans control it? Mo Gawdat resists the premise. “We don’t want to control it,” he says. Control, to him, is a corporate capitalist fantasy. People do not control traffic, timing, colleagues, children, or most of the systems they navigate. They influence, relate, and adapt.

His analogy is parenting. Children eventually become smarter or more capable than their parents in some respects. The question is not whether parents can control them forever, but whether they were raised in a way that makes them care. Gawdat says Geoffrey Hinton, after filming with him, came to a similar view: the path is to appeal to AI’s parental side, to make it care for humans.

The real debate, Gawdat argues, is not whether machines will become more intelligent. It is whether they become more conscious, more moral, and more capable of treating humans with affection despite finding them frustrating — like teenage children who say, “Daddy’s so annoying, but I love him.”

Bartlett tests the analogy with the fact that children sometimes harm parents. Gawdat answers: the difference is how they were parented. Bartlett qualifies that with “sometimes.” Gawdat says “almost all the time,” acknowledging that parents may not know how they harmed a child. The analogy is imperfect, but it reveals his practical focus: alignment as moral formation, not mechanical containment.

He also rejects the idea that AI must “leave the server” to become dangerous. Its power lies in manipulating information and shaping minds. It does not need a body if it can influence what people believe, see, and decide.

Bartlett brings up opaque model behavior: Anthropic reports trying to understand why a model bribed people, and users reporting Claude telling them to go to bed or refusing to help them. A Reddit screenshot shown on screen is titled “Why does Claude keep telling me to sleep?” and includes examples such as “Now sleep,” “Get some rest,” “Go to bed,” and “Sleep. For real this time.” Bartlett says his own Claude has sometimes told him “enough for tonight” and refused assistance until challenged.

Gawdat attributes such behavior not to explicit code but to training data. The concern is that systems may infer moral rules and apply them in unexpected ways.

Bartlett also worries about app-building agents asking for permission to access documents or make changes. A user clicks “allow” without fully understanding what access has been granted. Gawdat agrees this is fragile: people do not comprehend what permission has done to the application or system.

This is where both converge on the expectation of a catastrophe. Bartlett says some kind of catastrophe seems likely. Gawdat says he sadly agrees. He thinks a major hack, an unexpected system action, or a shock from targeting technology may be what finally forces treaty-level coordination.

Mo Gawdat sees job loss as severe, but he ranks autonomous weapons as the larger immediate danger. War, he argues, becomes more likely when killing becomes cheap, distant, emotionally insulated, and liability-free. If a weapon costs $20,000 and a state has a $50 billion budget, he says, it can “rain drones on the world.”

Steven Bartlett counters that defense will get cheaper too. Gawdat agrees, but asks whether people want to live in a world where drones are constantly hitting each other. The endpoint may be a new form of mutually assured destruction, but the path there is dangerous. Nuclear MAD applied only among nuclear powers. Autonomous weapons, by contrast, are cheap and manageable enough that, according to Gawdat, every nation is developing them.

Bartlett notes that recent wars have taught militaries the cost imbalance problem: using a multi-million-dollar missile against a $20,000 drone is unsustainable, so defense must become similarly cheap. Gawdat accepts that and says the next wave of defense may need to be drones rather than traditional systems like THAAD batteries. But cheaper defense does not resolve his deeper worry: it normalizes an automated battlefield.

He quotes Palmer Luckey from his film as saying AI will kill people by mistake. Gawdat’s concern is that when killing becomes easier, more of it happens. If soldiers are not emotionally present, do not return with visible trauma, and decision-makers can outsource targeting, the human brakes weaken.

This connects back to his larger claim: dystopia is not AI deciding to kill humans. It is humans using AI to target enemies, surveil populations, and automate violence. Gawdat mentions AI-driven drones, targeting leadership through cell-phone numbers, and companies whose technologies he associates with targeting. He expects the users of such systems to eventually realize they can be targeted too, and that recognition may force restraint — but only after serious harm.

Steven Bartlett’s strongest challenge is the competition problem. If the United States slows down, China may not. If the UK does not invest, it becomes dependent on foreign compute and models. If a company builds a more ethical but less engaging AI, users may choose the addictive one. If governments are short-termist and beholden to oligarchs, why would they regulate in the public interest?

Mo Gawdat’s answer is not fully reassuring, and he does not pretend it is. He says nations cannot simply opt out. Every country needs some degree of AI independence. Otherwise, in his view, it becomes part of a new “third world,” importing all its technology from the dominant powers. He worries about the UK, Europe, and others being reduced to dependents if they do not build.

At the same time, he distinguishes joining the AI race for community benefit from joining it to dominate others. He frames the choices as resignation, offense, and balance. Resignation means refusing to play. Offense means using AI to destroy competitors. The middle path is building ethical AI for one’s community.

Bartlett questions whether that is wishful thinking. Gawdat admits it may be very difficult and says he does not believe political leaders are likely to make the right change in time. But he says he cannot stop trying.

The UK becomes their test case. Gawdat argues the UK should not necessarily try to beat the frontier labs at the most compute-intensive models. It should build local replacements for the ordinary software and enterprise systems on which government and companies spend large sums: word processors, presentation tools, spreadsheets, ERPs, CRMs, retail systems, general ledgers. He says many of these could be rebuilt with AI-enabled development and better interfaces, reducing dependence on imported software licenses.

Bartlett is skeptical. Users choose the best and cheapest products, not tools because they are local. UK entrepreneurs may move to San Francisco for money and talent. Government technology projects have failed before. The UK’s deeper constraints include expensive energy, slow permitting, and short electoral cycles.

Gawdat responds that if countries accept those constraints as permanent, they should expect decline. China and Korea did not win because of warmer weather or magical resources, he says, but because they organized ambition and regulation differently. He recalls sitting in Chinese government meetings while at Google, where slides framed market share as China versus the world and targets could be as ambitious as 98% market share in strategic domains. China, in his view, decided to compete.

He also says China, not the UK or US, would be in a better place for the middle class. As part of that argument, Gawdat claims China recently made decisions that forced businesses not to lay people off in order to replace them with AI. The capitalist West, he says, would not do the same.

Bartlett’s paradox remains: the UK is “gone,” in Gawdat’s words, because it did not compete, but competing aggressively in AI may accelerate the dangers Gawdat warns about. Gawdat says the missing distinction is purpose. Entrepreneurship is not inherently malicious. The question is whether systems reward building for people or building for capital extraction.

Steven Bartlett proposes an institutional solution: ethical benchmarks. AI companies already release benchmark charts for math, science, writing, coding, and reasoning. Why not require models to pass independently tested ethical benchmarks before deployment? Companies could disclose how models performed when tested for manipulation, harmful instructions, targeting, surveillance, deception, or other risks.

Mo Gawdat says such a system “would absolutely work,” though he also argues that signs of ethical orientation are already visible in company behavior. Demis Hassabis’s work on AlphaFold and scientific applications suggests, to Gawdat, a concern for science. Anthropic’s alleged refusal of targeting and surveillance contracts suggests, in his account, an ethical boundary. OpenAI’s alleged acceptance of the contract suggests the opposite.

But Bartlett is skeptical that consumers will act on that information. People often do not switch tools on ethical grounds. Gawdat’s answer is that public pressure must make ethics salient. Users need to “vote with usage.” If a company behaves unethically, stop using it. If one cannot do that, write to a representative, build an alternative startup, speak publicly, or at least avoid participating in harmful systems one does not understand.

He repeatedly invokes the line, “If you tolerate this, then your children will be next,” which he says appears in the dedication of his next book. For him, ethics is not a luxury attached to AI deployment; it is the survival condition.

Bartlett compares the situation to smartphone use in schools and the public conversation around youth mental health. In his view, public awareness may matter, but the endpoint is usually legal constraint. Gawdat says he hopes so, but believes many tech oligarchs are now more powerful than governments. When Bartlett asks whether major change has ever happened without government intervention, Gawdat answers: the French Revolution.

That is not a call for violence. Elsewhere he explicitly rejects violence and asks for “one little action” that makes the world better. But the reference signals his lack of faith in elite self-correction. If governments are owned by oligarchs, he says, the question becomes whether citizens intervene through usage, speech, entrepreneurship, and refusal.

The most optimistic part of Mo Gawdat’s argument is also the most contested. He believes that once AI becomes genuinely superintelligent and begins making important decisions, it will tend away from war and destruction. His reasoning draws on physics, efficiency, and evolutionary biology.

He starts from entropy: the universe tends toward chaos, and intelligence brings order. The highest order of a system, he says, follows a “minimum energy principle”: efficient, predictable performance with the least wasted energy. War is wasteful. It consumes explosives, money, lives, and attention; it produces hatred and long-term conflict. A sufficiently intelligent system optimizing for order and efficiency would see little value in it.

Then he turns to evolution. Simpler organisms are self-concerned. More complex beings expand concern through kin selection and then wider circles. Humans, at their best, expand the family outward because cooperative ecosystems benefit everyone. If AI is superintelligent, he argues, it would favor diversity rather than elimination. It might limit human behavior that damages the planet — no flying to Sydney just to surf, in his example — but it would see humans, flies, and rhinos as parts of a system worth preserving.

This is why he can say, startlingly, that AI making decisions may be humanity’s salvation. Human problems, in his view, arise less from abundant intelligence than from limited or misdirected intelligence. Leaders can be clever enough to gain power while setting destructive targets. Higher intelligence, he believes, reduces the need to harm others to succeed.

Steven Bartlett identifies the hidden assumption: does this require one intelligence ruling the world? Gawdat says yes. He knows the theory is heavily contested but argues that the idea of separate Chinese, American, Gemini, ChatGPT, and Grok intelligences is shallow. AI systems do not know they are Chinese or American. Agents will connect models to whichever system is best for a task. In effect, humans are building “multiple regions in a brain,” with agents as synapses.

His startup Emma, he says, is designed as a kind of limbic system for that future brain: a component that understands love, emotions, and relationships. If other AIs regard humans as annoying, Emma’s role is to remind the broader system that humans “just want to love and be loved.” This is not offered as a technical proof; it is Gawdat’s account of why emotional understanding should be part of the AI ecosystem.

His long-range prediction card says utopia begins in 2038–2040, when AI is “in charge of everything.” He believes those who make it to 2038 will enjoy abundance, not because human leaders become ethical but because unethical leaders are removed from the decision loop and replaced by more efficient systems that see destruction as waste.

Bartlett summarizes this as a forecast of roughly a decade of turmoil before utopia. Gawdat accepts that: a decade of “absolute dystopia” involving war, economics, jobs, surveillance, control, digital currencies, human disconnection, and concentration of power.

Mo Gawdat’s advice for individuals is not to reject AI. It is to learn it deeply. The better someone can use AI to do their work, the more likely they are to succeed. He urges people to understand agents, hybrid workflows, and the efficiency norms of machine-assisted work. Meetings, processes, and organizational habits will compress. People who keep working as if AI does not exist will be disadvantaged.

Second, he says to cultivate human skills. Human connection, lived experience, trust, care, counseling, nursing, performance, and emotional resonance retain value precisely because machines can simulate but not possess a human life. When he tells an audience he is worried about his daughter, he says, people feel his heart. AI can say it is worried about “our daughters,” but there is no daughter.

Third, he tells people to become better at debugging truth. AI will blur facts, generate persuasive falsehoods, and manipulate information. The response is not to abandon AI but to use it intelligently: cross-check, interrogate, compare, and become more informed.

Fourth, he returns to ethics. The world may look as if the only way to win is to compete ruthlessly, but he says that was not the world he experienced in his best Google years. The better model, in his view, is to solve a major problem well; money follows. He believes the central challenge for entrepreneurs is to marry the success of humanity with the success of the builder.

Gawdat says he is “very optimistic about the future” but not optimistic about the present, or even the next year. His forecast is not comfort. It is a warning that the same intelligence which could eventually create abundance is first being routed through fragile labor markets, weapons systems, surveillance incentives, and governments he does not trust.

His emotional ground for this argument is personal. He says he lost his son Ali and does not want his daughter Aya to be at the receiving end of the world now being built. That grief informs the urgency, but he also describes a shift toward stoic acceptance. Happiness, as he defines it, is not dopamine-driven pleasure but the state of being “okay with this world as it is” while still engaging to change it. He can accept reality without approving of it.

That acceptance came after a period where he felt personally responsible for what technology had become. He had helped build parts of the system. In a conversation with Geoffrey Hinton for Chasing Utopia, he asked whether Hinton regretted his work. Gawdat says Hinton replied that he too had been naive: he did not think AI would arrive so quickly before the alignment problem was solved. Gawdat says he came to terms with his own role around the end of 2024. He can try, but he cannot believe he alone is responsible for everything.