Hélène Landemore’s essay in Noema proposes three tiers for giving ordinary people a say in AI governance: an LLM-mediated global consultation, then real-time online deliberation, then citizens’ assemblies. The first tier is already buildable: Anthropic interviewed 80,508 people across 159 countries and 70 languages in a single week.
The Global Assembly convened for COP26 seated 100 people chosen by lottery, speaking over 30 languages, and it took stipends, translation and 68 hours. Paolo Spada and I have argued that a hundred people is not a representative sample of anything, because small samples, biased population lists and low acceptance rates defeat the claim before anyone deliberates.
The objection everyone reaches for is the weaker one. Anthropic’s respondents were “active Claude users who’d already found enough value to keep using AI”, but my work with colleagues found that skew in who participates does not reliably transmit to who benefits, because it has to survive four links, and in our cases at least one always broke.
In Iceland’s crowdsourced constitution the last link broke, and the process had no bearing on the decision. A consultation nobody is obliged to answer is safe from bias the way a suggestion box is safe.
That first tier costs an API key and a week, and Anthropic’s stated next steps are a further study and deployment partnerships. So here is the opportunity for anyone funding this: the consultations will happen regardless. What still needs paying for is somebody on the other end obliged to reply.
Assorted links
- Stanford HAI makes the case that open weights are not open source: you can download the model and still not check the work.
- Clem Delangue would regulate weights, APIs and applications as three different things, because “we don’t regulate steel, we crash-test cars.”
- Chinese technology firms are recruiting AI trainees out of high school, against McKinsey’s speculated five-million-worker shortfall by 2030, which existing pipelines would fill about a third of.
- Generative AI floods the market for books: across 14,419 Amazon titles, the number selling in a quarter grew 19.2-fold and quarterly revenue only 8.9-fold.