Branko Marcetic’s case in Jacobin is that the fight over whether a superintelligent AI might one day end humanity has crowded out the harms the existing one is already doing. The mistake-prone model is the dangerous one — not because it is smarter than us, but because it is wrong constantly and we keep deploying it where being wrong once is fatal.

The clearest example is a near-war:

  • A Special Operations Command analyst used a chatbot to draft a report on a Chinese cargo ship; it hallucinated a nuclear weapons program.
  • The US military mobilized to intercept and board, guns drawn, aircraft already airborne — standing down only after officials checked and found the report was AI nonsense.
  • CNN traced it to two trends: “AI native” analysts who over-trust the tool, and the Pentagon’s rapid AI adoption under Pete Hegseth.
  • A direct US-China war would likely trigger a global depression and carries a high chance of going nuclear.

The same failure mode runs through weapons targeting. Palantir’s Maven compressed hours of target analysis into minutes before the strike that killed scores of Iranian schoolgirls, with personnel assuming the software would flag errors it did not. AI targeting was likewise central to Israel’s flattening of Gaza.

Beyond the battlefield:

  • MIT researchers call the classroom pattern “cognitive surrender” — 95 percent of college students now use AI for assignments, and graders and employers report weaker memory, less persistence, and vanished critical thinking.
  • Chatbots, designed to flatter and hold engagement, back users’ harmful statements 47 percent of the time, and have been linked to suicides, murders, and users fed their own delusions.
  • 17 LLMs answered only 2 of 5 financial questions correctly; legal hallucination rates run 58–88 percent; a Stanford/Harvard study found the best models made severely harmful clinical recommendations in 22 percent of cases.
  • AI-drafted briefs are drawing rising court sanctions, and two federal judges issued error-ridden rulings with it.
  • A class action alleges UnitedHealth used an error-prone AI tool to serially deny insurance claims; another alleges an AI hiring platform systematically discriminated against Black applicants.

Marcetic’s sharpest point is about causation: the market-driven hype about imminent superintelligence is itself part of the damage. Overstating capability is what gets the technology into war rooms and clinics full of people who then take the output at face value. The doomer debate doesn’t have to be settled to justify pausing, regulating, and guardrailing this industry — or banning it outright for particular uses.

The dumber, mistake-prone AI of today is enough of a risk to humanity as it is.