Effective Altruism

Effective altruism (EA) is the project of using evidence and reason to try to find out how to do the most good, and on that basis trying to do the most good. MacAskill and Pummer’s reference definition is careful about what EA is not: not utilitarianism (a category mistake — EA is a project, not a normative theory), not necessarily donation-focused, not necessarily individualistic, and not dependent on precise cardinal measurement of well-being. The project is tentatively welfarist (its aim is promoting well-being), impartial (everyone’s well-being counts equally), cause-neutral (choose causes by where resources do the most good, not personal attachment), and committed to at least limited aggregation (how many benefit matters, not just how much each benefits).¹

The movement coalesced in Oxford around 2011 (Giving What We Can, 80,000 Hours, the Centre for Effective Altruism), drawing on Peter Singer’s famine-relief ethics and on GiveWell’s charity-evaluation model. Its institutional machinery includes meta-charities (GiveWell for global health, Animal Charity Evaluators), career advising (80,000 Hours), and a large philanthropic arm — the Open Philanthropy Project, renamed Coefficient Giving — backed primarily by Dustin Moskovitz and Cari Tuna’s Good Ventures. Torres (a critic, but a well-informed former insider) put committed EA-aligned funding at roughly $46 billion as of 2021.²³

The AI-safety turn

The canonical primary document of EA’s move into AI is Holden Karnofsky’s May 2016 Open Philanthropy cause report, Potential Risks from Advanced Artificial Intelligence: The Philanthropic Opportunity. It announced that AI risk would become Open Philanthropy’s largest bet of senior staff time, justified by the fund’s three cause criteria — importance, neglectedness, tractability — plus a stated estimate of “at least 10%” (moderate robustness) that transformative AI (a transition comparable to the agricultural or industrial revolution — a coinage that stuck) would arrive within 20 years. Notably, the same document registers the reservations critics would later amplify: echo-chamber risk from the EA community itself, the danger of provoking “premature and/or counterproductive regulation,” and the possibility that the far-off work would prove futile.⁴

The intellectual bridge from charity evaluation to AI was longtermism: the view, rooted in Bostrom’s Future of Humanity Institute (founded 2005), Nick Beckstead’s thesis on the overwhelming importance of the far future, and Toby Ord’s The Precipice (2020), that humanity’s vast potential future dominates the moral calculus. MacAskill and Pummer themselves flag the pressure point as their “fifth objection”: if even 10^16 future lives are at stake, any action that slightly reduces existential risk can outrank saving thousands of lives today — their reply is that EA per se does not presuppose the answers to aggregation and future-person weighting, though working them out is part of the project.⁵⁶

The criticism landscape

EA’s critics now arrive from several directions at once, and they do not agree with each other — mapping the flanks is more useful than picking one.

Internal / philosophical objections (catalogued by MacAskill and Pummer, with replies): deontic constraints on promoting the good (answered by a qualified definition — do the most good without violating constraints); individualism versus collective obligation (definition deliberately left underspecified); Lenman’s cluelessness about indefinite-run effects (Greaves: the sea of unforeseeable effects is “vast but also silent”); the fanaticism/small-probability objection above; and incommensurability between cause areas.⁷

Torres (the anti-longtermist left): longtermism is “quite possibly the most dangerous secular belief system in the world” — an ideology whose cosmic-potential arithmetic can render near-term catastrophes (even a climate disaster killing billions) mere “ripples,” whose value-neutrality thesis about technology licenses ever more dangerous capability development, and whose posthuman telos makes it self-defeating: the Baconian drive to maximise control is itself a driver of existential risk. Care about the long term, Torres argues, but reject the ideology.⁸

Chilson (the libertarian regulatory flank): EA’s AI-doom faction has moved from philosophy to legislation with authoritarian design — the draft Responsible Advanced Artificial Intelligence Act (permit regimes for software, six-month industry shutdowns, hardware seizure, chip registries) and California’s SB 1047 (prove-a-negative safety cases, kill-switch mandates), both from EA-adjacent organizations, alongside MIRI’s stated objective of shutting down frontier AI worldwide and Yudkowsky’s call to bomb rogue datacenters. The FTX collapse hangs over all of it: Bankman-Fried framed his fraud as a Machiavellian scheme to fund EA causes.⁹

Rao (the sociological critique): EA proper is one strand of a larger “EA formation” — philanthropy, rationalism, longtermism, x-risk research, AI safety and governance, and frontier-lab adjacency — whose three genealogies (GiveWell-style evidence philanthropy; the Singer–Ord–MacAskill consequentialist lineage; the LessWrong–MIRI–Bostrom rationalist-x-risk complex) were increasingly captured by the third once it supplied astronomical stakes. Rao’s diagnosis is structural: a totalizing moral style that performs uncertainty within its models (credences, sensitivity analyses) while immunizing the models themselves from doubt — “epistemic shock absorbers” — combined with a theological architecture (eschatology, vocation, salvation via alignment) and an extremizing mechanism visible in the formation’s deviant tail (FTX, the Zizian milieu, high-control groups), which he reads as stress tests of the epistemology rather than unrelated debris. His prescription is neither persecution nor acquiescence but pluralist containment — “after the wars of religion,” no doctrine, EA included, gets to monopolize the moral governance of AI. Hence the essay’s title joke: EA promised to solve AI safety; now we have two problems.¹⁰

Connections

• ai-economic-growth-complementarity — the growth-economics side of transformative AI; Karnofsky’s 2016 TAI coinage and the GPI (Trammell–Korinek) line both come out of the EA institutional world described here.
• butlerian-jihad-thesis — the refusal pole of the AI-response spectrum; Rao’s pluralist containment is a third position between refusal and capture by any single governing doctrine.
• aisi-unsanctioned-agent-behaviour-2026 — the state AI-safety-institute apparatus that the EA formation’s decade of field-building helped bring into existence.
• virtue-ethics — the rival normative tradition EA’s consequentialist-adjacent project most often eclipses in online discourse.
• anthropic-cybersecurity-eval-incidents — Anthropic is the clearest case of an EA-adjacent frontier lab; its eval incidents are the kind of “stress test” evidence both Rao and Chilson point at from different directions.

Sources

• William MacAskill and Theron Pummer 2020 — Effective Altruism
• Holden Karnofsky 2016 — Potential Risks from Advanced Artificial Intelligence: The Philanthropic Opportunity
• Émile P. Torres 2021 — Why longtermism is the world’s most dangerous secular credo
• Neil Chilson 2024 — The Authoritarian Side of Effective Altruism Comes for AI
• Venkatesh Rao 2026 — EA Safety

Footnotes

1. William MacAskill and Theron Pummer 2020 — Effective Altruism ↩

2. William MacAskill and Theron Pummer 2020 — Effective Altruism ↩

3. Émile P. Torres 2021 — Why longtermism is the world’s most dangerous secular credo ↩

4. Holden Karnofsky 2016 — Potential Risks from Advanced Artificial Intelligence: The Philanthropic Opportunity ↩

5. Émile P. Torres 2021 — Why longtermism is the world’s most dangerous secular credo ↩

6. William MacAskill and Theron Pummer 2020 — Effective Altruism ↩

7. William MacAskill and Theron Pummer 2020 — Effective Altruism ↩

8. Émile P. Torres 2021 — Why longtermism is the world’s most dangerous secular credo ↩

9. Neil Chilson 2024 — The Authoritarian Side of Effective Altruism Comes for AI ↩

10. Venkatesh Rao 2026 — EA Safety ↩