The BaseEssayAI Alignment Forum

How I'm thinking about different types of automated alignment

In a 7-minute post linked from Forethought, owencb argues that automating alignment work splits into three kinds, empirical, conceptual and direct, with different strategic consequences, and sets out which look worth speeding up and the two uncertainties that decide priorities.

Open the source ↗3 quotations checked, word for word ↓

Why it matters

The same thing, for different readers.

For lawmakers

owencb says AI safety funders should not prioritise speeding up automated empirical alignment research, since the capabilities needed overlap heavily with automating AI capabilities research.

For researchers

owencb says benchmarks for conceptual research, which later automated empirical work could target, are a more robust investment than other ways of automating conceptual alignment research.

The question it raises

owencb suggests a sufficiently aligned AI revising its own values reflectively could resemble meditation more than neurosurgery, while noting the risk of wrong starting conditions.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“The capabilities that are needed for automating empirical alignment research very heavily overlap with the capabilities needed for automating AI capabilities research.”
owencb, Author on the AI Alignment Forum, linking to ForethoughtExplaining why accelerating automated empirical alignment research is not a priority.
Found word for word in the source
“If you can do a good job of automating conceptual alignment research, you should be approximately done”
owencb, Author on the AI Alignment Forum, linking to ForethoughtOn how the three types of automated alignment work interact.
Found word for word in the source
“My personal guesses on these uncertainties are that most of the loss-of-control risk comes later (after AI is superhuman even at tasks that don’t have verifiable rewards), and that it is possible to directly approach decent automated conceptual research.”
owencb, Author on the AI Alignment Forum, linking to ForethoughtClosing the section on priorities, with the author noting high uncertainty.
Found word for word in the source

The record

Questions
What happens when AI helps build more capable AI? · Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
People
owencb (Wrote it, Author on the AI Alignment Forum, linking to Forethought)
What we read
We read the whole source.
Found by
The Radar, run of November 12, 2026, via ai-alignment-forum (tier 1: the source’s own channel or an official record)
Cite this
AI Alignment Forum, “How I'm thinking about different types of automated alignment”, November 11, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/alignmentforum-2026-11-11-types-of-automated-alignment
Our notesA linkpost; the original is on forethought.org, the saved text is the Alignment Forum copy. The author's real name is not given on the page.