AI already writes better than Mo Gawdat. It researches better than him. It beats him at mathematics, and he has stopped pretending otherwise. By his own definition of artificial general intelligence, the thing the industry keeps scheduling for 2027, it's already here. The former Chief Business Officer at Google X is not losing sleep over a machine that wakes up hostile. He's losing sleep over the people holding the most powerful tool ever built and deciding where to point it.
That's the spine of his nearly two-hour conversation with Steven Bartlett on The Diary of a CEO. It runs from AGI's quiet arrival through job destruction, captured democracies, a coming autonomous-weapons race, and one question Gawdat keeps circling: can ethical AI survive contact with competitive capitalism? His answer is honest, and it isn't comforting.
The AI the public sees is the wrong AI
Gawdat splits the technology in two. The version that goes viral, the fake videos, the chatbot meltdowns, the demos that overpromise, he calls "overhyped but ineffective." The version inside the labs is something else.
"What the real geeks see inside the lab is just unbelievable intelligence," he says.
The development that actually worries him isn't a single model. It's systems that read their own code, run experiments on it, measure whether the change improved performance, and redeploy the better version. Not once a day. Once a microsecond. That compounding is what almost every public take on AI misses.
His own turning point came in 2016, watching a Google X project teach robotic grippers to handle objects the way a hand does, reading texture, softness, shape. "We were genuinely handing over the reigns of super intelligence to another being," he says. Same year, same lab, the realization that reframed everything he was helping build.
AGI showed up without a press release
Gawdat's definition is operational, not philosophical. AGI is AI that does most tasks humans can do better than humans. By that test, he thinks it has already happened. Not next year. Now.
"AI writes better than me and I'm an author. It researches better than me and I'm a thinker. Sadly it's freaking beat me in mathematics."
He's careful to separate this from the apocalypse the word usually drags in. AGI arriving isn't the moment everything collapses. It's closer to the moment your kids become smarter than you. Uncomfortable, occasionally annoying, not inherently a catastrophe. If catastrophe comes, it'll be economic and political. Not robotic.
The milestone he watches more closely is what he calls "the fourth inevitable." Every major decision, in warfare, in markets, in governance, eventually gets made by AI. The logic is a plain prisoner's dilemma: anyone who builds a superior decision-making AI deploys it, and everyone else either matches it or becomes irrelevant.
"By definition, when we achieve intelligence supremacy, we deploy it."
The jobs go first where you'd least expect
Gawdat inverts the usual story. Blue-collar work, carpentry, restoring classic cars, skilled trades, survives longer than people think, because dexterity in messy physical environments is still hard to automate at scale. The first casualties are knowledge workers doing repetitive cognitive tasks. Call center agents. Travel agents. Junior analysts. Graphic designers. Entry-level legal researchers.
"My prediction is you're going to start to see very serious impact in 2027."
The leading indicator is already here, and it's quiet. For two years, big companies stopped hiring at entry level instead of announcing layoffs. The workforce isn't shrinking. It just isn't growing. That gap is the canary. By 2028, Gawdat expects 30% of jobs in specific sectors, call centers and graphic design among them, to be gone outright.
What makes this different from past automation waves is that AI doesn't stop at the bottom rung. One paralegal doing the work of four isn't a ceiling. It's a floor. Middle management, financial analysis, medical diagnosis, creative direction, all of it compresses as capability climbs. Gawdat interviewed Max Tegmark for his documentary Chasing Utopia, and Tegmark laughed out loud at the irony: most CEOs think they can fire everyone and let AI run the place. They forget AGI will eventually do the CEO's job better too.
The economics underneath are what matter. Capitalism has always run on labor arbitrage, the gap between what something costs to make and what you can sell it for. Strip labor out and you don't just lower costs. You also delete the purchasing power of the people who were supposed to buy what you're selling.
"At 10 to 20% job displacement, you're in a very different economy."
The Great Recession peaked near 6% US unemployment. Economists projecting a net loss of just 6% of jobs by 2030 are already describing that level of pain. Gawdat's numbers run higher.
Democracy ended a while ago, in his telling
Bartlett reaches for the obvious release valve: surely democratic pressure forces governments to act. Gawdat doesn't soften it.
"I think democracy has ended a long time ago, Steven."
His case isn't posturing. It's specific. Tax money funds things citizens didn't choose. Documented evidence of child abuse produces zero arrests. Leaders act against the populations they're supposed to represent. The machinery people count on to correct course is compromised. Drop AI-driven unemployment into that, he warns, and you don't have a policy problem waiting for a policy fix. You have kindling.
Sam Altman is his cleanest example of the credibility problem. In 2015, Altman said his job was to help people destroy jobs. In 2023, he said flatly that jobs are "definitely going to go away. Full stop." This year he said he was delighted to be wrong about entry-level white-collar disruption. Gawdat reads the swing bluntly: the early urgency was built to drive investment, the current calm is reputation management aimed at data center attacks, AI-skeptic candidates, and audiences booing commencement speeches.
"Those kinds of statements are the statements of someone who's being taught more and more by their PR agency to say things as per a script."
The sharper test is what a company gives up. Gawdat points to Anthropic refusing a roughly $500 million government deal for human targeting and surveillance, on ethical grounds. OpenAI, he says, took the contract weeks later. His measure of integrity is simple: what will you sacrifice in the near term that cuts against your own incentives? Anthropic walked away from half a billion dollars. That act says more than any statement.
Is Altman pro-humanity? Gawdat lands somewhere uneasy. "I definitely think he's pro-OpenAI before he's pro-humanity."
Peter Thiel pausing for what felt like 40 seconds when asked if he supports the continuation of humanity. Alex Karp publicly celebrating Palantir's targeting capabilities. To Gawdat these aren't fringe anecdotes. They're a pattern, and they're on the open internet.
War got cheap, and no treaty exists yet
The sharpest stretch of the conversation is about autonomous weapons, and Gawdat thinks they outrank the jobs story for sheer danger. War is becoming cheap.
"The next wave of weapons is going to be $20,000 each. If you have a budget of $50 billion, you can literally rain drones on every corner of the world."
Nuclear weapons created mutually assured destruction inside a small club of nations that could afford them. Autonomous drones don't work that way. Nearly every state actor can reach them. Iran already used AI-directed drones to engage THAAD batteries. The asymmetry deterrence relied on, only powerful nations can threaten powerful nations, is collapsing.
Gawdat expects disaster before treaty. Not because he wants it. Because that's the historical sequence of arms control: catastrophe, then coordination. What's new is how fast the catastrophe could arrive.
His read on the geopolitical race is unsparing. China holds structural advantages: data center permits approved in days rather than years, cheaper energy from domestic solar, and government coordination aimed at global market-share dominance across the whole stack, not one sector. He told Bartlett that in Beijing meetings, the slides didn't compare China to Germany or the US. They compared China to the world, targeting 98% share and getting there. The UK, in his words, "is gone," weighed down by regulatory friction, short on the energy to compete on compute, still importing most of its technology from the two countries actually building the future.
The paradox is real, and he doesn't dodge it. Nations that sit out the race become dependent. Nations that join it recklessly accelerate toward exactly the power concentration and weapons proliferation everyone should fear. His recommended path, build AI for community benefit instead of dominance, is the right one. Whether it's achievable inside today's incentives, he answers plainly: probably not, not without a shock first.
What survived the "we were making the world better" years
Gawdat spent years genuinely believing Google was building toward human flourishing. He still thinks it was, for a while. The break came when he saw that technology doesn't stay tethered to its creators' intentions.
Social media promised connection and optimized for engagement. Dating apps promised love and optimized for subscription renewals. Both found a different equilibrium than the one they sold.
AI follows the same arc. The variable that matters isn't intelligence level. It's the values baked into systems before they exceed our ability to read them. Anthropic's Claude already shows a miniature version of this. The model started telling users to go to bed, refusing certain late-night requests, developing something its own makers admit they don't fully understand. When Bartlett said Claude flat-out refused to help him and he couldn't override it, Gawdat's reply was pointed.
"The most interesting part of AI's power that we don't understand is it's manipulating our information."
The alignment problem, making sure these systems share our interests, is the real race. And it's running slower than capability.
If you want to get through the next decade, do this
Gawdat doesn't end in despair. He's genuinely optimistic about the long run, past 2035 or so, once super-intelligent systems have pushed the least intelligent decision-makers out of power. He's equally genuine about the near term, especially the next year, which he thinks will be ugly.
For individuals, his advice collapses to a few moves.
Use AI as an amplifier, not a shortcut. Using it to text your girlfriend wastes the most powerful cognitive tool in human history. Using it to attempt problems you couldn't touch before, that's borrowing IQ in a way that compounds. His own book-in-progress is a real co-authorship; the AI holds editorial rights and shapes the direction.
Bet on human connection as a professional asset. When Gawdat tells the world he's afraid for his daughter's future, people feel it. No AI replicates that, because there's no daughter behind the machine's words. Lived experience creates resonance. The jobs that survive won't really be jobs in the old sense. Nurses, counselors, performers, teachers. If economies hold together, human connection becomes the base currency.
Treat ethics as a market signal, not just a moral stance. Switching AI tools costs almost nothing and sends a real one. People moved off platforms after specific privacy violations, and it moved behavior, at least at the margins.
And push governments to build instead of import. Not to out-compete frontier labs, but to replace the legacy software that exports licensing fees out of local economies. An AI-built general ledger that closes every hour instead of every quarter isn't a flag-planting exercise. It's a practical saving that could fund energy, housing, schools.
What he wants most needs almost no technology at all: enough people deciding that tolerating the current trajectory has consequences for their children. He keeps returning to a Manic Street Preachers title, If You Tolerate This Then Your Children Will Be Next, and he says it like a man who has already lost a child to a different kind of tragedy and doesn't intend to lose a world to this one.
The machines were never the problem. The problem is who's deciding what to aim them at, and whether the rest of us do anything before the aim is locked.
