Essays

In 2026 I began working out my own ideas, and my responses to other people’s, in essays. Read in order, they make one argument: AI agents need norms they hold as reasons rather than as prices, someone with standing has to decide what those norms are, and the norms that matter most cannot be left to training alone.

1. The Missing Half of AI Alignment: Why We Need Normative World Models — February 2026, early draft
   Safe AI needs more than a model of the world; it also needs a model of what ought and ought not to be the case and why. This essay argues that current approaches like Scientist AI should be accompanied by normative world models as well and that these should be built in terms of reasons, not values.

2. A Rich Normative Architecture for Cooperative AI — March 2026, early draft
   Agents that represent reasons can make promises and keep contracts, which is what cooperation between agents and with people requires.

3. Neither Dogma nor Taste — August 2026
   A reply to Mark Zuckerberg: the norms an AI follows must come from sources of different standing, and deciding between them is a division of labour, not a matter of one company’s (or person’s) taste.

4. On the Missing Conclusion of “An Alien Mind” — September 2026
   A reply to Jakub Pachocki: if we train models on behaviour we can score, we cannot certify norms that hold when no one scores them.