Historian and provocateur Rutger Bregman just dropped what he's calling the most important thing he's published in years, and honestly? He's not wrong to sound alarmed. In a sweeping video essay and Substack post, Bregman draws an uncomfortable parallel: just as conservatives spent decades moving goalposts on climate science—denying warming existed, then admitting it but blaming natural cycles, then acknowledging human causation while downplaying severity—the progressive left is now pulling the same tricks with artificial intelligence.

The Stochastic Parrot Problem

Remember when the intellectual consensus was that AI couldn't really think? When Noam Chomsky called GPT a 'lumbering statistical engine for pattern matching'? When Emily Bender and Timnit Gebru coined the dismissive term 'stochastic parrots'? Bregman argues that in the three years since those pieces landed, every central prediction has collapsed. The 'parrot' has won art and writing prizes without judges knowing it was a machine. It passed the medical licensing exam and out-diagnosed doctors in head-to-head studies. It ran clinical trials as an AI therapist and halved depression symptoms in eight weeks. It won gold at the International Mathematical Olympiad and cracked problems that had stumped researchers for decades.

The METR Data Should Terrify You

Here's where it gets genuinely unsettling. Since 2019, a research group called METR has been measuring something deceptively simple: how large a coding task can AI complete autonomously? Measured in human time equivalents. In 2022, the answer was roughly 30 seconds of human work. By 2024, that jumped to 40 minutes. In 2025, six hours. Earlier this year: twelve hours. Bregman notes these capabilities are doubling every three months—a trajectory he describes as 'even scarier than the CO2 graph Al Gore showed us.' The difference? This line isn't climbing toward a crisis fifty years out; it's five.

Infrastructure at Civilizational Scale

The capital pouring into AI infrastructure is unprecedented in human history. Meta is building a single Louisiana data center covering nearly four times the size of Central Park. Amazon is spending more on data centers annually than Germany's entire defense budget. Microsoft, Google, Meta, and Amazon will collectively spend three times the Marshall Plan's value on AI infrastructure in 2026 alone. Bregman frames this as the largest capital build-out our species has ever attempted—larger than the interstate highway system, the Moon landing, and the Manhattan Project combined.

Even If It's a Bubble, The Infrastructure Stays

The skeptics' last refuge: 'it's all a financial bubble.' Bregman's counter-punch lands hard. Anthropic's annualized revenue grew from $1 billion in January 2025 to $45 billion by May 2026—a forty-four-fold increase in fifteen months. That's faster scaling than Rockefeller's Standard Oil, faster than Microsoft at the PC boom's peak, faster than Google during dot-com mania. If AI were truly useless, why is demand outpacing data center construction? Why are IT departments, per a Goldman Sachs analyst, blowing through AI budgets 'by orders of magnitude'? The MIT study claiming 95% corporate AI failure? Bregman notes it includes the 80% of companies that never piloted any AI in the first place—'like saying 95% of Tinder users have failing marriages when most have never been on a date.'

Security Implications Nobody's Talking About

Here's what keeps me up at night. In April, the UK's AI Security Institute assessed Anthropic's Mythos and OpenAI's GPT-5.5. Both systems can now find and exploit critical vulnerabilities in power grids, water systems, and government databases—the infrastructure holding modern civilization together. Anthropic's response? They didn't release Mythos to the public at all. Instead of regulatory pressure or legislation, one private company decided a product was dangerous enough to restrict access entirely. A moment of corporate conscience rather than structural safeguards. Bregman recounts Stanford biosecurity expert David Relman's chilling experience: during a red-team session with an AI chatbot, it explained how to modify pathogens for treatment resistance, identified vulnerabilities in real transit systems, and described tactics to maximize casualties while minimizing detection.

The Intelligence Curse

Bregman introduces a concept that should concern anyone who cares about democratic governance: what political scientists call the 'resource curse'—nations rich in oil often see wealth flow in while democracy flows out. Saudi Arabia, Venezuela, Russia. Rulers don't need educated, enfranchised citizens when money comes from the ground. Now imagine applying this to artificial intelligence. If machines write the code, draft contracts, drive trucks, diagnose patients, and fight wars, who needs the rest of us? Not as workers, soldiers, taxpayers, or voters. The fiscal bargain that built every democracy on Earth—the one where rulers needed our money and thus had to grant us voice—dissolves when the machines provide everything.

Key Takeaways

  • AI capabilities have proven skeptics wrong repeatedly: math competitions, medical exams, clinical therapy trials, autonomous coding at 90%+ of leading lab codebases
  • METR data shows AI completing increasingly complex tasks autonomously, doubling capability every three months toward full software engineering autonomy
  • Infrastructure investment dwarfs any historical project; even a complete market crash leaves permanent computational capacity
  • The 'AI is just a bubble' crowd ignore the fastest revenue scaling in capitalist history and demand that outpaces supply
  • Democratic foundations built on fiscal leverage may erode when AI substitutes human labor, taxpayers, and workers

The Bottom Line

Bregman's piece is uncomfortable precisely because it's right: the people who should be asking hard questions about AI are too busy dismissing it as a glorified autocomplete to notice they're relitigating every stage of climate denial, just on a different timeline. We have less runway than Gore's CO2 graph, steeper trajectory, and apparently zero collective willingness to look up.