Exploration · as of October 2026
AI safety compendium
The fate of humanity may hinge on whether we're able to safely align superintelligent AI. Over a hundred unique ideas have been proposed for how to achieve this. The prospect of contributing to such a consequential project is alluring to many, but I myself spent too much time rederiving ideas that had already been offered by others before me. Therefore, I set out to map every idea in AI safety, so that I could start from the frontier and build upon it. This is that map.
Thus far, my attempted mapping has covered 227 sources, producing 150 unique ideas across 12 category groups.
Has this been tried?
Describe an idea in plain words. Below are the closest approaches on this page. Each one shows how many of the 7 source groups name it and how much work it gets.
Or try one of these:
24 sources on the field, in 7 groups
Expert panels, governments, labs, funders and researchers have each published their own overview of ways to make AI safe. This chart lines up 24 of them, sorted into 7 groups by who wrote them.
Each column is one group of sources. Each row is one approach to making AI safe. A filled dot means at least one source in the group names the approach. A hollow dot means none of them names it, but at least one implies it or names something close.
Tap a row to read what each source says about it. Tap a group name to see its sources.
The 24 sources, by group, with links
How many sources cover each group
The approaches on one map
Each point is one of the 150 approaches, placed by judged scores on two of seven scales. Colour shows the family; shape shows the group within it. Tap a point or a label to read about the approach.
Approaches with the same scores are packed side by side; position inside a cluster has no meaning.
Families and groups
Colour = family; shape = group within the family. Bold labels mark open opportunities: work the field calls neglected, or that funders are asking for now.
Where the 150 approaches sit on each scale
Kind of tool is not a true scale. Its categories run from most formal to least formal: maths and code, then measurement, then law and markets, then strategy, then philosophy and norms. Read it as categories, not as a number.
Positions are judgments, not measured quantities. The families do not form clean regions on the map.
150 approaches, by group
This list draws on 227 published sources, the 24 in the chart among them, and on scans of research agendas, lab programmes and funder calls. It names who works on each approach and how much attention it gets.
People and organisations
A who's who of people and organisations that shaped the field's ideas, money and institutions. Open a card for key ideas, works and where to start reading.