NOTE-1302: The Case for Strong Longtermism

Contribution

It states strong longtermism precisely and argues for it. Axiological strong longtermism (ASL) says that in the most important decision situations, far-future effects fix which options are near-best. Deontic strong longtermism (DSL) says one ought to choose accordingly. The paper claims more rigour than Bostrom (2003) and Beckstead (2013), a wider base of assumptions, and a new argument for the deontic version (p. 3).

Key insight

Expected value is driven by the largest scenario in which one has any credence. So a tiny credence in space settlement or digital minds sets the expected size of the future, and any option that shifts the probability of a persistent state (extinction, or a lasting AI-controlled order) carries that size with it. Effects that change which persistent state the world ends up in do not “wash out” (§4.1).

Assumptions

• Expected total utilitarianism, at first. Ex ante value is expected value, value is total welfare, and value is time-separable (p. 5). Section 6 relaxes the first two.
• A zero rate of pure time preference. Called “near-essential” (p. 18). “With even a modest positive rate of pure time preference … the argument would not go through” (p. 18).
• Lives are on average worth living (p. 10), needed for the extinction examples, not for the AI one.
• Digital minds would have comparable moral status (p. 8).
• The status quo is wasted resources (p. 4). Benefits are measured against it.
• Neglectedness. The claim is about the current margin. Whether it would hold if 10% of world GDP already went to the far future, “We aren’t sure” (p. 14, n. 14).

Key results

• The thesis (p. 3).
  • ASL(i): every option that is near-best overall is near-best for the far future.
  • ASL(ii): every such option delivers much larger benefits in the far future than in the near future.
  • “Far future” means from some time t on, say 100 years (p. 4).
• The benefit ratio (BR, p. 4). The highest far-future ex ante benefit attainable without net near-future harm is many times the highest attainable near-future benefit. The appendix proves that BR(n) gives ASL(ii), and ASL(i) once options with net near-term harm are removed (pp. 30–32). With n = 10, a 90%-best option gets at least 80% of the far-future best, and at least 8 times as much far-future as near-future benefit.
• The near-future benchmark (p. 5). Bednets save a life for about $4,000,so0.025livesper$100 (GiveWell).
• The size of the future (pp. 6–9).
  • Natural-risk extrapolation gives at least 87,000 years; mammalian species last 0.5 to 6 million years; Earth stays habitable about a billion years.
  • Newberry’s six scenarios run from 10^14 future lives (Earth, mammalian reference class) to 10^45 (Milky Way, digital life) (p. 8).
  • A 0.01% credence in biological settlement of the Milky Way adds at least 10^32 (p. 9).
  • Estimates used: main 10^24, low 10^18, restricted 10^14 (p. 9).
• Tractability, per $100 (pp. 11–15). Read with main / low / restricted estimates.
  • Spaceguard asteroid survey ($71 million by 2012): 500 million / 500 / 0.05 lives. Further asteroid detection: 300,000 / 0.3 / 0.00003.
  • Pandemic preparedness ($250 billion cutting a 1-in-22,000 century risk by 1%): 200 million / 200 / 0.02.
  • AI safety (at least 0.1% risk; $1 billion averting 1% of it, so 0.001% absolute): one trillion / one million / 100. Only this example beats bednets on the restricted estimate.
• Meta-options (§4.4). Research and a longtermist fund that saves for later raise the bar for intractability objections. A 2% annual failure rate for such funds is called “conservatively high” (p. 16, n. 16).
• Individuals (§5). Donors and career choosers are in the same position, given diminishing rather than increasing returns.
• Robustness to axiology (§6, pp. 17–19).
  • Risk aversion and prioritarianism strengthen the case for mitigating catastrophe.
  • Person-affecting views: “premature extinction is in itself at worst neutral” (p. 18). The extinction examples may then fail. The AI example does not, since a large population exists either way. “Narrow” person-affecting views are set aside as implausible; “wide” ones agree with totalism on the AI cases (p. 18, n. 20).
  • A future worse than nothing defeats the case for avoiding extinction but not the AI case.
  • Discounting: conceded as above. A zero rate is endorsed by Sidgwick, Ramsey, Stern and others; 38% of surveyed experts on social discounting agreed (p. 19, n. 21). Models of partiality that weight contemporaries more are judged unlikely to discount enough (n. 21).
• Cluelessness (§7). Simple cluelessness cancels out for trivial choices but not for deliberate attempts to make AI safer. Conscious unawareness, arbitrariness, imprecision and ambiguity aversion are each answered; ambiguity aversion is called “anyway irrational” (p. 24).
• Fanaticism (§8). “One of the most plausible ways in which the argument for strong longtermism might fail” (p. 25). Replies:
  • Beckstead and Thomas show that avoiding it forces “timidity”.
  • At the societal level the probabilities (one in 100,000 for $1 billion on AI) are above the one-in-a-million risks people act on.
  • At the individual level ($10,000 for one in ten billion) the authors concede the probabilities may be problematically small (p. 26).
• Deontic strong longtermism (§9, pp. 26–29).
  • Stakes-sensitivity argument: when the stakes are very high, no serious side-constraint applies and the prerogatives at issue are minor, one ought to choose a near-best option (P1). The most important decisions today are like that (P2).
  • Against non-aggregation: voting, fighting in WWII, protest.
  • Against a non-identity restriction: it would deny that governments must act on climate change, “clearly incorrect” (p. 28).
  • Demandingness is left to the literature. Society is nowhere near any cap on required sacrifice: “We ought to be doing a lot more for the far future than we currently are” (p. 29).
  • Side-constraints rule out atrocities; there is usually a permissible alternative (p. 29, n. 38).

Claims

id	claim	strength	support
C1	Any reasonable estimate of the expected number of future beings is at least 10^24	weak to moderate: driven by small credences in settlement and digital minds, by the authors’ own description	§3
C2	Some available interventions have far-future expected benefits per $100 many times the 0.025 lives of bednets	moderate on the main estimate, weak on the restricted one, where only the least quantified (AI) case passes	§4
C3	ASL holds in society’s and individuals’ resource and career decisions	moderate, conditional on C1–C2, expected-value reasoning and zero pure time preference	§§2–5, appendix
C4	The case survives risk aversion, prioritarianism, person-affecting views and a negative-value future	moderate: by switching to the AI example, which C2 supports least	§6
C5	A positive pure rate of time preference would defeat it	strong: conceded	p. 18
C6	Cluelessness and ambiguity aversion do not undermine it; fanaticism may	moderate: arguments, with fanaticism left open	§§7–8
C7	DSL follows from ASL on any plausible non-consequentialism	moderate: a valid argument whose second premise is argued briefly	§9

Concepts

• axiological / deontic strong longtermism: far-future effects are the most important determinant of the value of options / of what one ought to do (p. 3).
• benefit ratio (BR): far-future attainable benefit without near-term harm is many times near-future attainable benefit (p. 4).
• persistent state: a set of world states that, once entered, the world is expected to stay in for an extremely long time (p. 10).
• washing-out hypothesis: expected value differences between options decay with time fast enough that near-future effects dominate (p. 10).
• stakes-sensitivity: minor constraints and prerogatives give way when the axiological stakes are very high (p. 27).
• fanaticism / timidity: sacrificing arbitrarily much for a tiny chance of a vastly larger payoff / refusing to (p. 25).

Connections

• Bostrom 2003 and 2013 (LIT-1739, LIT-1773). The extinction-risk arithmetic is Bostrom’s (cited as 2013:18, p. 11). The paper extends it beyond total utilitarianism and adds non-extinction trajectory changes.
• Ramsey (LIT-1733). Named among the endorsers of zero pure time preference (n. 21), the assumption the case needs.
• Person-affecting views and the non-identity problem (Narveson, LIT-1769; LIT-1768; Parfit, LIT-469). The paper concedes the extinction case to them and rejects a non-identity restriction on obligation by the climate-change case.
• Tarsney (LIT-1753) and Thorstad (LIT-1688). Tarsney is cited on washing-out (n. 22). Thorstad’s challenge postdates this version.
• Scheffler (LIT-1712). Not cited. The paper has no ground of obligation in what the living value.

Bearing on the record

• The owner’s thesis (CLAIM-491). The paper is the clearest case in the record of an obligation grounded in an impersonal good. These are inferences; the authors say nothing about the self.
  • The ground is the total good, not anyone’s. Value is total welfare, counted the same whoever and whenever it falls on. The thesis has no quarrel with that as a view. It bears only on what could move someone to act on it.
  • Nothing is said about motivation. The agent is “a rational altruistic actor” (p. 17) with an “altruistic budget” (p. 29). Being moved is presupposed. The “ought” is subjective but normative, and demandingness is left to the literature. Even the one sentence on prudence treats it as a cost: devoting resources to the far future “at the large expense of our own prudential interests” (p. 29).
  • The governing image is prudential. “Humanity, today, is like an imprudent teenager” (p. 2), whose most important choices matter for the rest of its life. That models the obligation on one agent’s concern for its own future. The thesis can read this as humanity as a self extended in time. The authors use it only as an analogy.
  • Partiality is the nearest the paper comes to a self. Models that weight one’s contemporaries more (Setiya, Mogensen) are mentioned and judged not to discount enough (n. 21). A thesis that grounds concern in the self predicts some such discount. The paper does not dispute that people have one, only that they ought to.
• The triviality objection (CLAIM-492). Not addressed. But the paper sharpens what a counterexample to the thesis must be: someone who acts on strong longtermism at a real cost, with no identification with humanity or its future. The paper says nothing about whether such people exist.
• Proposed dated note on CLAIM-491: “Greaves and MacAskill (NOTE of 2026-10-29) ground obligations to the far future in total welfare, undiscounted and owed to no one’s self. They say nothing about why anyone would act on them; their agent is simply ‘a rational altruistic actor’. So the thesis is not contradicted by the view, only challenged to explain its adherents. Their one image, humanity as ‘an imprudent teenager’, models the obligation on prudence across one agent’s life, which is the thesis’s shape, offered as analogy.”
• No THEORY filed.
• No ML instruction. AI safety is an example, not advice on practice.

Limitations

• Not the published chapter. The 2025 text may differ.
• The decisive numbers are judgement. The AI estimates (0.1% risk, 1% averted by $1 billion) have, by the authors’ account, “no hard quantitative evidence” (p. 14). Only that example passes on the restricted estimate.
• Person-affecting robustness rests on the weakest example. C4 succeeds by moving to AI, the least quantified case.
• The individual-level fanaticism problem is conceded, not answered (p. 26), yet §5 extends the thesis to individuals.
• An inconsistent illustration. Section 5 supposes $1 billion could cut the chance of an asteroid collision this century by 1 in 120,000 (p. 16). Page 11 gives the remaining risk of a 10km+ collision as 1 in 150 million, so the cut exceeds the risk. It is an illustration, but it does not match the paper’s own figures.
• Typos and slips.
  • “miniscule” for “minuscule” (p. 10).
  • “John Hopkins Center” for “Johns Hopkins” (p. 13); “Center for New American Security” for “Center for a New American Security” (p. 14).
  • “Sandberg and Armstrong 2013” in the text (p. 7); the reference list has Armstrong and Sandberg.
  • “near-future benefits of bednet mitigation” for “distribution” (p. 23).
  • “several centuries’ hence” (p. 20); “500 additional million lives” (p. 11).
  • “The expected value approach we assumed in section 3” (p. 20); it was set out in section 2.2.

Open questions

• Who acts on strong longtermism, and why? A study of committed longtermists that measured identification with humanity or with future generations, alongside their stated reasons, would show whether the impersonal ground is the operative motive.
• Does the 2025 chapter keep “near-essential” for zero pure time preference, and the same person-affecting reply?