(This is the fifth essay in a series that I’m calling “How do we solve the alignment problem?”. I’m hoping that the individual essays can be read fairly well on their own, but see this introduction for a summary of the essays that have been released thus far, and for a bit more about the series as a whole.

Podcast version (read by the author) here, or search for "Joe Carlsmith Audio" on your podcast app.

See also here for video and transcript of a talk on this topic that I gave at Anthropic in April 2025. And see here for slides.)

1. Introduction

In my last essay, I argued that we should try extremely hard to use AI labor to improve our civilization's capacity to handle the alignment problem – a project I called “AI for AI safety.” In this essay, I want to look in more [...]

---

Outline:

(00:43) 1. Introduction

(03:16) 1.1 Executive summary

(14:11) 2. Why is automating alignment research so important?

(16:14) 3. Alignment MVPs

(19:54) 3.1 What if neither of these approaches are viable?

(21:55) 3.2 Alignment MVPs don't imply hand-off

(23:41) 4. Why might automated alignment research fail?

(29:25) 5. Evaluation failures

(30:46) 5.1 Output-focused and process-focused evaluation

(34:09) 5.2 Human output-focused evaluation

(36:11) 5.3 Scalable oversight

(39:29) 5.4 Process-focused techniques

(43:14) 6 Comparisons with other domains

(44:04) 6.1 Taking comfort in general capabilities problems?

(49:18) 6.2 How does our evaluation ability compare in these different domains?

(49:45) 6.2.1 Number go up

(50:57) 6.2.2 Normal science

(57:55) 6.2.3 Conceptual research

(01:04:10) 7. How much conceptual alignment research do we need?

(01:04:36) 7.1 How much for building superintelligence?

(01:05:25) 7.2 How much for building an alignment MVP?

(01:07:42) 8. Empirical alignment research is extremely helpful for automating conceptual alignment research

(01:09:32) 8.1 Automated empirical research on scalable oversight

(01:14:21) 8.2 Automated empirical research on process-focused evaluation methods

(01:19:24) 8.3 Other ways automated empirical alignment research can be helpful

(01:20:44) 9. What about scheming?

(01:24:20) 9.1 Will AIs capable of top-human-level alignment research be schemers by default?

(01:26:41) 9.2 If these AIs would be schemers by default, can we detect and prevent this scheming?

(01:28:43) 9.3 Can we elicit safe alignment research from scheming AIs?

(01:32:09) 10. Resource problems

(01:35:54) 11. Alternatives to automating alignment research

(01:39:30) 12. Conclusion

(01:41:29) Appendix 1: How do these failure modes apply to other sorts of AI for AI safety?

(01:43:55) Appendix 2: Other practical concerns not discussed in the main text

(01:48:21) Appendix 3: On various arguments for the inadequacy of empirical alignment research

(01:55:50) Appendix 4: Does using AIs for alignment research require that they engage with too many dangerous topics/domains?

The original text contained 64 footnotes which were omitted from this narration.

---

First published:
April 30th, 2025

Source:
https://www.lesswrong.com/posts/nJcuj4rtuefeTRFHp/can-we-safely-automate-alignment-research

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Diagram showing process-focused and output-focused evaluation steps with monitoring elements.

Diagram showing relationships between empirical and conceptual alignment research automation steps.

Diagram showing three research categories along

Diagram showing three research categories across

Diagram showing relationship between empirical and conceptual alignment research automation

The image is a flowchart-style diagram that illustrates the connections between different types of alignment research and their automation paths. The diagram moves from

Diagram showing process-focused and output-focused evaluation of a production system.

Diagram showing AI training progression across 5 difficulty levels with annotations.

The image illustrates a conceptual framework for training and evaluating AI systems, with difficulty increasing from Level 1 to Level 5. The first four levels (shown in green) represent tasks that can be evaluated, while Level 5 (in orange) represents tasks beyond current evaluation capabilities.

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

“Can we safely automate alignment research?” by Joe Carlsmith

LessWrong (30+ Karma)

“Can we safely automate alignment research?” by Joe Carlsmith