Alignment Forum

The AI Alignment Forum is an online hub where researchers discuss how to make transformatively powerful AI systems act in line with human values. It launched in 2018, and by its 2021 FAQ several hundred researchers had contributed about two thousand posts and nine thousand comments. Topics range from technical models of agency to strategy, and well known posts include "What failure looks like" and "Embedded Agents."

The forum is run by the team behind LessWrong, and the two sites share a codebase, a database and user accounts. Every new Alignment Forum post or comment is copied to LessWrong automatically, and forum members can promote LessWrong content to the Alignment Forum. Karma scores are kept separate, so a post's Alignment Forum rating reflects only votes from that community. An independent board with representatives of major alignment research groups oversees major decisions.

Membership is deliberately small. The FAQ says about 50 to 100 members can post and comment without review and can promote other people's work. Members come from groups such as DeepMind, OpenAI, MIRI, CHAI and ARC and their collaborators. Working on alignment professionally is not enough on its own; membership usually follows a record of strong contributions that several existing members value.

Non members can still take part. They can post on LessWrong, where good alignment content is often promoted, or submit directly to the forum, where a decision comes within three days and rejections get at least a one sentence explanation. Either way the post appears on LessWrong right away.

For newcomers, the forum recommends four introductory sequences: Embedded Agency, Iterated Amplification, Value Learning, and AGI Safety from First Principles. For context on how labs approach safety in practice, see our guide to the OpenAI Preparedness Framework and Anthropic's Responsible Scaling Policy.