Quartz 4

Effective Altruism — The Mathematics of Doing Good

9 min read

Effective Altruism — The Mathematics of Doing Good

Effective altruism (EA) asks a question that feels obvious once stated: given that you want to do good, how do you do the most good? The answer, EA argues, requires the same tools used to solve any optimization problem — evidence, quantification, expected value, and systematic prioritization. The result is a movement that is simultaneously one of the most mathematically rigorous attempts at ethics ever made, and one of the most contested.

Confidence level: established (the methodology and movement are real and documented; individual cause prioritization claims are emerging to theoretical)

Key Facts

  • Founded on Peter Singer’s 1972 essay “Famine, Affluence, and Morality”: affluent individuals have a moral obligation to donate until the marginal utility of further giving equals the marginal utility of keeping the money
  • William MacAskill and Toby Ord co-founded the Centre for Effective Altruism (2011); MacAskill coined “effective altruism” as a label
  • GiveWell (2006): most rigorous charity evaluator; top charities (anti-malaria nets, direct cash transfers, vitamin A supplementation) save lives at 5,000 per death averted
  • Toby Ord’s The Precipice (2020): 10% probability of existential catastrophe from unaligned AI this century; 3% from engineered pandemics; 0.1% each for asteroid impact, supervolcano, nuclear war
  • FTX collapse (November 2022): Sam Bankman-Fried’s ~439M → $216M)
  • Open Philanthropy: >$330M donated to AI safety research organizations through 2025
  • EA-affiliated organizations employ ~5,000 people globally; annual EA-directed giving estimated at 1B (post-FTX)

The Core Methodology

The Utilitarian Foundation

EA traces directly to Peter Singer’s 1972 thought experiment: a drowning child ten meters from you and a drowning child in a distant country have equal moral weight; physical proximity does not create moral priority. If spending X surplus, not spending it is equivalent to allowing a preventable death.

This is impartialist consequentialism — outcomes matter, not relationships or identity of those affected. It creates the moral engine of EA: maximize expected welfare (or prevent expected suffering) given limited resources.

The Three Prioritization Criteria

MacAskill’s framework for cause selection requires a cause to score highly on:

  1. Scale: How many people are affected, and how severely?
  2. Neglectedness: How much existing attention and funding does this cause receive relative to its importance?
  3. Tractability: How much progress can additional resources actually make?

The combination generates unexpected cause priorities. Global health beats domestic poverty (scale: billions of people in low-income countries; tractability: proven interventions; neglectedness: vastly underfunded relative to domestic aid). Existential risk beats global health by longtermist calculation: even a 0.0001% reduction in extinction risk saves more expected lives than all global health spending combined, because the expected future population is ~10¹⁶ people.

The Quantitative Tools

QALYs (Quality-Adjusted Life Years): one QALY = one year of perfect health. An intervention preventing a year of blindness might be worth 0.6 QALYs. GiveWell’s Against Malaria Foundation analysis uses probabilistic models integrating RCT data, epidemiology, and QALY weights to produce a cost-per-QALY-gained estimate.

DALYs (Disability-Adjusted Life Years): one DALY = one lost year of healthy life. Used primarily in WHO and academic public health literature; equivalent in logic to QALY but inverted in direction.

Expected value calculations: EA reasoning routinely multiplies small probabilities by enormous stakes. Toby Ord: reducing AI extinction risk by 0.00001% × 10¹⁶ future people × 80 expected life-years = 80 billion expected QALYs — larger than all human welfare in recorded history. This reasoning requires trusting expected value under deep uncertainty — an assumption many philosophers reject.

Cause Areas

Near-Term: Global Health and Animal Welfare

GiveWell’s top charities (2025) distribute anti-malaria bed nets (Against Malaria Foundation), vitamin A supplementation, direct cash transfers (GiveDirectly), and deworming programs. Cost per life saved: 5,000. These are established interventions with strong RCT evidence.

Animal welfare is the second major near-term EA cause: factory farming affects ~80 billion land animals per year; welfare improvement per dollar may be extremely high due to neglectedness.

Long-Term: Existential Risk and the Longtermist Turn

William MacAskill’s What We Owe the Future (2022) and Toby Ord’s The Precipice (2020) argue that existential risk reduction dominates all near-term cause areas on expected value grounds.

Longtermism: the position that the long-run future of humanity is the most important moral priority, because the vast majority of people who will ever live have not yet been born.

Ord’s risk estimates (theoretical, not empirical — based on expert surveys and structural reasoning):

  • Unaligned AI: ~10% this century
  • Engineered pandemics: ~3%
  • Nuclear war: ~0.1%
  • Asteroid impact: ~0.01%
  • All other: ~5%
  • Total: ~18.5% existential catastrophe this century

The logic: if any of these risks could be reduced by even 1% for 1B spent on malaria nets ($5,000 per life saved). This is the mathematical case for AI safety funding.

The FTX Crisis and Its Aftermath

Sam Bankman-Fried (SBF) was the most prominent EA longtermist, worth ~190M to AI safety, pandemic preparedness, and civilizational resilience projects in 2022.

In November 2022, FTX collapsed amid fraud allegations; SBF was convicted on all charges (November 2023). The FTX Future Fund collapsed, clawing back commitments. MacAskill, who had publicly praised SBF, distanced himself immediately.

The structural critique the crisis revealed: EA’s utilitarian cost-benefit framework — “large expected gains justify rule violations if the expected value calculation comes out positive” — provided a coherent rationalization for the fraud. “Galaxy-brained” reasoning (long chains of expected-value logic leading to counterintuitive conclusions) is a known failure mode of consequentialist reasoning, and SBF embodied it.

Post-FTX, EA has:

  • Emphasized institutional robustness and anti-corruption norms more explicitly
  • Seen significant criticism from within the movement about longtermist epistemic overconfidence
  • Continued funding AI safety at scale through Open Philanthropy (~$330M through 2025)

The Major Philosophical Objections

The Fanaticism Objection (Pascal’s Mugging)

EA reasoning requires trusting expected value under astronomical uncertainty. If there is a 10⁻¹⁰ probability that an action saves 10¹⁰ lives, its expected value (1 life) should dominate all other considerations — but this allows tiny probabilities of huge impacts to override all other moral considerations indefinitely. Philosophers call this the fanaticism problem: expected value reasoning with extreme numbers produces recommendations that seem morally absurd. The formal analog is Pascal’s Mugging (Nick Bostrom): a mugger demands money in exchange for not destroying 10¹⁰ lives via implausible future tech; expected value reasoning says you should pay.

Moral Uncertainty

EA assumes expected-value calculations can be made across different ethical frameworks, but moral uncertainty (uncertainty about which ethical theory is correct) does not straightforwardly aggregate. MacAskill’s response is “maximize expected moral value across a probability distribution over ethical theories” — but this requires knowing the probability of utilitarianism being correct, which is itself a metaethical question with no established answer.

The Neglect of Justice and Systemic Change

Critics from the political left (notably Amia Srinivasan, Émile P. Torres) argue that EA’s individualist, quantitative approach ignores the structural causes of suffering (economic exploitation, colonial history, political power). Distributing malaria nets while supporting financial deregulation that causes the poverty requiring the nets is, on this view, not effective altruism — it is harm laundering.

Power Concentration Risk

A movement directing hundreds of millions of dollars on the basis of a single utilitarian framework, controlled by a small techno-philanthropist class (Gates Foundation, Open Philanthropy, formerly FTX), embodies the structural risks it claims to address. The concentration of moral authority in a small community creates echo chambers, institutional capture, and reduced epistemic diversity — the exact failure modes that cause civilizational catastrophe.

Cross-Realm Connections

  • concept-deep-time: Longtermism is applied deep time ethics — it takes seriously the moral weight of the far future. But deep time reveals a tension: the most consequential changes in Earth’s biosphere (Great Oxygenation Event, Cambrian explosion) were not performed by moral agents at all. True deep time ethics confronts the fact that the most important future events may be beyond any agent’s control or prediction

  • concept-halting-problem: EA’s top cause — AI alignment — is formally undecidable (Rice’s Theorem: any non-trivial property of programs is undecidable). EA may be spending billions of dollars attempting to solve a problem that is mathematically unsolvable in the general case. This does not mean specific alignment approaches are impossible, but it means the problem cannot be “solved” in the sense EA imagines. The expected-value calculation over AI risk implicitly assumes the risk is tractable — an assumption the halting problem undermines

  • concept-free-will: The utilitarian framework requires coherent moral agents making effective choices. But Sapolsky’s hard incompatibilism and the microbiome-free-will connection suggest human decision-making is more constrained by biology and environment than EA’s models assume. “Earning to give” assumes rational control over career and donation decisions that neuroscience increasingly questions

  • concept-quantum-measurement-problem: EA’s expected-value reasoning is logically parallel to quantum probability: collapse a distribution over outcomes weighted by probability × value. But quantum mechanics shows that probability distributions over complex systems can be irreducibly indeterminate (not just epistemically uncertain). Applying the same mathematical structure to moral calculations may be importing false precision

  • concept-soc-civilizations: SOC (self-organized criticality) analysis of civilizations suggests that systemic collapse is a power-law process, not a manageable risk with a well-defined probability. Ord’s 10% AI risk estimate treats existential risk as if it were a frequentist probability — but if civilizational catastrophe is an SOC avalanche, it may not have a well-defined probability at all

  • concept-emergence: EA assumes macro-level social outcomes are predictable from micro-level individual action (donate to the most effective charity, change the world). Emergence theory and Hoel’s Causal Emergence suggest that macro-level causal structures may be irreducible to individual actions — meaning the most important levers for civilizational change may not be accessible to individual donors at all

See Also

Graph View

Backlinks