They Build Language Models and Can’t Use Words

The US midterms also said something about artificial intelligence. According to NBC’s exit poll, 56% of voters named AI or data centres as a crucial or major factor in their choice for Congress: the lowest of the five issues tested, yet still a majority. At least thirty-three towns voted to restrict data centres. Jon Ossoff, re-elected senator for Georgia, used his victory speech to attack the “titans of Silicon Valley” who “dig bunkers for themselves”, inside a broader charge against concentrated wealth. AI was one cause among several, and I see some risks worth addressing here.

For years the public conversation about AI has been shaped by people I know well: brilliant nerds, swots, with a score to settle with the world. They talked about power, supremacy, a race to be won, and “superintelligence”, a word that in Washington had become a political programme. There are exceptions and, I have to admit, one of them is Peter Thiel, someone I could hardly be further from, who nonetheless argues from genuine learning and makes you think. One of very few exceptions, though.

That language did a job. If AI is a race that cannot be lost, rules become a handicap and resources belong with whoever runs fastest: the rhetoric of supremacy was the argument for concentration. It also bred polarisation. Citizens worried about AI, about the power bill rising next to a data centre, about their jobs, about their children, were pushed towards the other camp, and the other camp took them in. Buying Twitter may not have been one of Musk’s wiser moves: it gave him a stage for a self-referential story and set his critics on fire. Trump, at least, tells his story to himself on his own network, Truth, or Pravda in Russian. Both built closed rooms where the message reaches only those already convinced. The risk now runs the other way: blanket bans where targeted ones or verification would do, laws that damage the machine without governing it, probably written to answer an angry electorate more than a real problem.

Concentration has a concrete face, and the latest one is in schools. From 2027 Google will switch off Search, Gemini and NotebookLM by default on student accounts, and schools will decide whether to switch them back on. A school that wants Google Search without its AI features has no such option: according to Google’s head of education, “search and AI are deeply integrated”. A private company sets the terms on which schools reach one of the most basic tools there is, and those terms are all or nothing.

A society that rejects the nanny state, and my idea of freedom points that way, only works if its citizens can tell what they are being sold, and that starts in the classroom. And if a school has to give up web search to keep AI out of the classroom, the freedom to choose has already been exercised by someone else.