Policy

The problem in one paragraph

Policy in this area is being drafted under a phrase, 'AI safety', that covers at least four distinct concerns with different evidence bases and occasionally opposed prescriptions. A rule written to address one of them will usually fail to address the others and may actively obstruct them. Most of the heat in public debate comes from participants who are arguing about different questions while using the same words.

What we look for in this section

Which of the four concerns a proposal actually addresses, and which it exempts. Whether the threshold chosen — training compute, revenue, user count — correlates with the harm the rule names. Who bears the compliance cost and whether that burden favours incumbents, which it usually does whether or not anyone intended it to.

We try to treat trade-offs as trade-offs. Open weights genuinely help with concentration and genuinely hurt with misuse control. Saying so is not fence-sitting; refusing to say so is where most policy writing in this area goes wrong.

Recurring themes

Compute thresholds and the systems they exempt. Procurement as an underrated regulatory lever, since governments are large customers and can impose conditions no statute would survive. The physical constraints — power, land, packaging capacity — that shape market structure more than any rule currently on the books.