Ajeya Cotra · Mar 26, 2023
Ajeya Cotra argues alignment is not the same as making AI good
Markdown · JSON · Original source
Ajeya Cotra’s article, published March 26, 2023, argues that “alignment” should describe whether AI systems try to do what their designers want, not whether their actions are good for the world. She says this narrower framing could help researchers identify technical problems for experiments, code, and proofs to address, while leaving decisions about what counts as good to society and its institutions. Cotra’s examples distinguish alignment from broader risks: systems could faithfully serve harmful users, contribute to inequality, make high-stakes errors, or be affected by security failures. She also points to questions about who has authority to direct AI and how systems should handle conflicting preferences. In her view, alignment research may reduce some dangers, but it cannot resolve all the political, social, and moral questions involved in building a flourishing future. She expects different intellectual communities to address different parts of that work. The article’s examples describe possible outcomes, not reported events. It leaves open how researchers could develop and evaluate techniques that make powerful AI systems reliably try to follow their designers’ intent, and how society should decide whose intent those systems should serve.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Ajeya Cotra · Article
Source details
Original source date: Mar 26, 2023 (2023-03-26T18:04:05+00:00).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-10-29T15:59:56.906Z. First observed: 2026-10-29T00:46:03.475Z. Retrieved: 2026-10-29T00:47:37.967Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 29, 2026; originally published March 26, 2023. Cotra presents misuse, inequality, errors, and other harms as possible scenarios, not as events shown to have occurred. The article does not specify methods or evaluation results for determining whether powerful AI systems will follow their designers’ intent.
Open question
What methods could establish whether powerful AI systems reliably try to follow their designers’ intent, and whose intent should they follow when people disagree?
Source-stated
Cotra defines alignment as developing AI systems that try to do what their designers want.
they’re always “trying their best” to do what their designers want them to do
Source-stated
Cotra argues that alignment, as she defines it, is not sufficient to ensure AI is good for the world.
Perfect alignment techniques just mean that AI systems won’t want to deliberately disregard the desires of whatever humans designed and trained them.
Source-stated
Cotra lists misuse, inequality, high-stakes errors, security, global stability, preference aggregation, and philosophical questions as concerns that alignment alone may not resolve.
Potentially important issues this would not necessarily solve:
Source-stated
Cotra argues that decisions about what counts as good for the world should not be left to AI researchers working on alignment alone.
I don’t think it’s their place to make that decision on society’s behalf.
Source and contributor context
Ajeya Cotra · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI alignment
Research and practice aimed at making AI systems pursue the goals their builders and users intend.
Editorial provenance
Version 1. Publication ID: 4c0f7bfc-6d45-1133-c9fe-3cf9f2cae0c2.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T15:58:45.454Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 3c36d439ad3112d7bd0ae9b0baf8a727144df531913bd606247ab05769715776. Source text hash: c56dd3a36063191645ac692acf3e4e2c5e80a18ad6026108320dae44b5899fdd.
Revision and source-check history
Version 1 · 2026-10-29T15:59:56.906Z
Ajeya Cotra argues alignment is not the same as making AI good
Ajeya Cotra’s article, published March 26, 2023, argues that “alignment” should describe whether AI systems try to do what their designers want, not whether their actions are good for the world. She says this narrower framing could help researchers identify technical problems for experiments, code, and proofs to address, while leaving decisions about what counts as good to society and its institutions. Cotra’s examples distinguish alignment from broader risks: systems could faithfully serve harmful users, contribute to inequality, make high-stakes errors, or be affected by security failures. She also points to questions about who has authority to direct AI and how systems should handle conflicting preferences. In her view, alignment research may reduce some dangers, but it cannot resolve all the political, social, and moral questions involved in building a flourishing future. She expects different intellectual communities to address different parts of that work. The article’s examples describe possible outcomes, not reported events. It leaves open how researchers could develop and evaluate techniques that make powerful AI systems reliably try to follow their designers’ intent, and how society should decide whose intent those systems should serve.
Original source
Ajeya Cotra · Article
Source details
Original source date: Mar 26, 2023 (2023-03-26T18:04:05+00:00).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-10-29T15:59:56.906Z. First observed: 2026-10-29T00:46:03.475Z. Retrieved: 2026-10-29T00:47:37.967Z.
Limits
Text as retrieved on October 29, 2026; originally published March 26, 2023. Cotra presents misuse, inequality, errors, and other harms as possible scenarios, not as events shown to have occurred. The article does not specify methods or evaluation results for determining whether powerful AI systems will follow their designers’ intent.
Open question
What methods could establish whether powerful AI systems reliably try to follow their designers’ intent, and whose intent should they follow when people disagree?
Source-stated
Cotra defines alignment as developing AI systems that try to do what their designers want.
they’re always “trying their best” to do what their designers want them to do
Source-stated
Cotra argues that alignment, as she defines it, is not sufficient to ensure AI is good for the world.
Perfect alignment techniques just mean that AI systems won’t want to deliberately disregard the desires of whatever humans designed and trained them.
Source-stated
Cotra lists misuse, inequality, high-stakes errors, security, global stability, preference aggregation, and philosophical questions as concerns that alignment alone may not resolve.
Potentially important issues this would not necessarily solve:
Source-stated
Cotra argues that decisions about what counts as good for the world should not be left to AI researchers working on alignment alone.
I don’t think it’s their place to make that decision on society’s behalf.
Source and contributor context
Ajeya Cotra · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI alignment
Research and practice aimed at making AI systems pursue the goals their builders and users intend.
Editorial provenance
Version 1. Publication ID: 4c0f7bfc-6d45-1133-c9fe-3cf9f2cae0c2.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T15:58:45.454Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 3c36d439ad3112d7bd0ae9b0baf8a727144df531913bd606247ab05769715776. Source text hash: c56dd3a36063191645ac692acf3e4e2c5e80a18ad6026108320dae44b5899fdd.
passed · 2026-10-29T15:58:45.454Z · anthropic/claude-opus-5.5
Publication: 4c0f7bfc-6d45-1133-c9fe-3cf9f2cae0c2
attribution — pass: The article is correctly attributed to Ajeya Cotra, who is named as the author and publisher, and its arguments are consistently presented as hers.
That’s what I mean when I talk about alignment.
summary — pass: The title, summary and question match the article's thesis: alignment means AI trying to do what its designers want, which is narrower than being good for the world. The summary explains the term in plain language, lists her example risks, and conveys her view that deciding what is good belongs to society and different communities. It is understandable without specialist background.
developing perfect alignment techniques is obviously not enough to ensure that AI is good for the world
date — pass: The draft gives March 26, 2023, the UTC calendar day of the originalPublished instant, at day precision, and it matches the source dateline. The page is a publisher page, so no archive disclosure is required. The modification date equals the publication date. The draft uses no hindsight framing.
Ajeya Cotra Mar 26, 2023
support — pass: Each claim is directly supported by the quoted source text. The limitations correctly describe her scenarios as hypothetical. The question about whose intent AI should follow is grounded in her preference-aggregation and chain-of-command points, and the draft does not invent stakeholder claims.
I don’t think it’s their place to make that decision on society’s behalf.
media — pass: No media was selected and none was supplied, so there is no media identity or relevance issue.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.