Can AI automate alignment research?

Could AI systems that are not yet dangerous do the alignment research needed for the systems that are? Alignment research is the work of making AI systems reliably pursue the aims their builders intend; see AI alignment.

Why it matters

Plans that give alignment months rather than decades depend on a yes. A short window is only enough if AI researchers do most of the work. Several frontier developers describe some version of this as their main plan. It is a narrower and more testable question than How hard is alignment.

The case for yes

• AI already speeds up ordinary research and software work, and much alignment research is ordinary empirical work: running experiments, building tools, and checking results.
• For some problems, checking a solution is easier than producing one, so humans can verify work they could not have done.
• Each generation of aligned assistants might help align the next, creating a feedback loop that outpaces capability gains.
• Monitoring and control measures could catch an assistant that sabotages the research.

In his 2026 essay Against SURA, Scott Alexander sketches the hopeful version. Each model generation finds tricks that keep the next one well behaved, until some generation solves alignment properly.

The case for no

• The hardest parts are the ones we cannot check. If humans cannot evaluate a proposed alignment method, rewarding the AI for good proposals rewards proposals that look convincing.
• A system capable of solving alignment may already be capable enough to be dangerous, or to quietly underperform on research that would constrain it.
• Time pressure removes the slow, adversarial review that would catch subtle errors.
• Success at research that can be scored may not transfer to the conceptual problems that cannot.

Simon Lermen, replying to Scott, argues that the default outcome is a last-minute attempt to implement a flawed idea that merely sounds convincing. These objections apply an older argument from AGI Ruin - A List of Lethalities to current plans: humans cannot reliably check the work of a system smarter than they are, so they cannot tell a real solution from a persuasive wrong one.

What would change minds

AI-produced alignment advances that humans can independently verify. Evidence about whether models underperform on safety work while performing well elsewhere. Whether research assistants generalize from scored tasks to open conceptual ones.

Depends on Can we learn from AI failures before they are fatal: a feedback loop between generations requires failures to be visible.