Did MIRI Ever Aim to 'Take Over the World'? AI Safety Community Feuds Over FAI History and CEV

On August 15-16, the AI safety community erupted in intense debate surrounding MIRI's historical mission. JD Pressman accused MIRI affiliates of denying the organization once planned to build Friendly AI (FAI), calling it "gaslighting" of those involved; Habryka insisted MIRI never attempted to "take over the world" and cited old articles to refute the claim. The debate extended to whether CEV (Coherent Extrapolated Volition) equates to a power grab, with both sides sticking to their guns without reaching a consensus.

Confirmed

• The focus of the debate is whether MIRI truly committed to building FAI between 2011-2013; Pressman claimed they believed they could achieve AGI first through mathematical purity.
• Regarding evidence: Pressman cited a 2013 article by MIRI co-founder Luke Muehlhauser; Habryka stated Luke explicitly expressed no desire to build FAI themselves, and Eliezer discussed the harms of attempting to take over the world as early as 2004.
• jessicata cited MIRI's 2017 fundraising and strategy update, stating their strategy was to retain the option of "which projects to help" and maintain pre-emptive secrecy on the AGI critical path, rather than "one organization should execute mission AGI".
• Wei Dai clarified that Pressman was initially just relaying others' "take over the world" remarks, not directly advocating the view.

Unconfirmed

• Whether MIRI seriously plotted to take over the world remains inconclusive: Pressman argued the mathematical object defined by MIRI was the way to "ethically take over the world", while sarcastically noting this is too flattering—the reality being MIRI let fans imagine this while never taking it seriously internally; Habryka denied gaslighting, threatening to "post 50 more times" to refute the claim.
• There is also disagreement on whether CEV equals a power grab: Habryka argued that if the US unilaterally extended sovereignty globally, even promoting democracy would constitute a takeover, and that successful CEV would empower hypothetical or simulated humans rather than real ones; Wei Dai countered that the core of a power grab is depriving others of power, whereas CEV aims to maximize human empowerment, so implementing CEV is not taking over the world.
• Wei Dai also cited ThomasCederborg's article on LessWrong: Parliamentary CEV (PCEV) would give extra influence to those who essentially enjoy harming others; if implemented, the consequences would be worse than extinction, highlighting the importance of Alignment Target Analysis (ATA).

Why It Matters

• The controversy points directly to divergences in historical narratives within the AI safety circle, relating to the assessment of MIRI's path and credibility; the debate over CEV's empowerment vs. power grab and the ethical flaws of PCEV also indicate that the choice of alignment target remains a pending safety issue.

2026-08-15 ~ 2026-08-16 · 25 related posts

• Debate over MIRI's mission statement and Friendly AI creation — jd_pressman · 2026-08-15
• Habryka denies gaslighting in MIRI takeover debate, cites LessWrong discussion — ohabryka · 2026-08-15
• Habryka stands firm: will repeat MIRI didn't try to take over world 50 times — ohabryka · 2026-08-15
• JD Pressman: MIRI Never Seriously Tried to Take Over the World, Just Allowed Fans to Imagine — jd_pressman · 2026-08-15
• [source] Habryka argues MIRI didn't try to take over world, citing Luke and Eliezer — ohabryka · 2026-08-15
• jd_pressman questions: is tinkering with CEV during takeover legitimate? — jd_pressman · 2026-08-15
• Habryka clarifies: denies MIRI takeover, admits CEV brain-in-box plan — ohabryka · 2026-08-15
• jd_pressman insists: MIRI defined math object to take over world — jd_pressman · 2026-08-15
• JD Pressman: MIRI's Math Object Is an Ethical Way to Take Over the World — jd_pressman · 2026-08-15
• Habryka: 'take over the world' is misleading, use other terms — ohabryka · 2026-08-15
• CEV Ethical Flaw: PCEV Gives Extra Influence to Sadists, Outcome Worse Than Extinction — weidai11 · 2026-08-15
• Habryka: improving world isn't takeover; CEV empowers others — ohabryka · 2026-08-15
• CEV: Empowerment or Power Grab? AI Safety Debate Continues — weidai11 · 2026-08-15
• MIRI Debate Intensifies: Is CEV Equivalent to 'Taking Over the World'? — weidai11 · 2026-08-16
• MIRI's Early FAI Ambitions Spark Heated AI Safety Debate — jd_pressman · 2026-08-16
• JD Pressman accuses MIRI members of rewriting history and gaslighting — jd_pressman · 2026-08-16
• [source] JD Pressman accuses MIRI members of rewriting history and gaslighting — jd_pressman · 2026-08-16
• [source] MIRI's 2017 Strategy: Retaining Optionality on Alignment-Capabilities Insights — jessi_cata · 2026-08-16

7 near-duplicate retellings: jd_pressman · ohabryka · jd_pressman · jd_pressman · jd_pressman · jd_pressman · jd_pressman