TLDR; 

Minimal longtermism only provides guidance for those working directly on existential risks, but not everyone can do that. This essay argues that another effective longtermist intervention may be systematically documenting biological, linguistic, and cultural information that once lost, could never be reconstructed. I try to show that even under pessimistic assumptions, such interventions generate substantial longterm utility. 

Minimal vs. Expansive Longtermism

In "Minimal and Expansive Longtermism," Hilary Greaves and Christian Tarsney distinguish between two interpretations of longtermist theory that differ in scope and robustness. "Minimal" longtermism focuses narrowly on mitigating technological existential risks through targeted interventions like AI safety research, pandemic preparedness, and nuclear risk reduction. This approach enjoys strong evidential support because the causal mechanisms are clear: certain technologies increase extinction risk, and mitigating that risk obviously increases future expected utility.

Expansive longtermism, by contrast, claims there are many ways to improve the far future and that far-future effects should guide nearly all current decisions. Greaves and Tarsney assiduously avoid giving a definitive opinion about expansive longtermism's ultimate correctness, but they conclude that arguments in favor of it are “significantly less robust and significantly more speculative than the arguments for minimal longtermism."

The Allocation Gap in Longtermism Theory

While I agree that expansive longtermism is less robust than minimal longtermism (it necessarily incorporates additional empirical claims about long-term causation that are harder to verify), this methodological point may circumvent a practically important question. If longtermism seeks to influence how people allocate their efforts, then determining which "expansive" interventions might nonetheless be defensible becomes crucial for reasons beyond theoretical completeness. It would offer guidance for those whose careers do not involve mitigating existential risk. 

Many people would benefit from such broadened guidance. Minimal longtermism offers clear direction for philanthropists, policymakers, and those capable of pursuing careers in relevant technical fields. However, it provides limited guidance for the substantial population who cannot (or will not) work directly on technological x-risks. Even assuming universal conversion to longtermist views, diminishing returns would quickly limit how many people could productively work on AI safety, biosecurity, or nuclear policy. Moreover, as frequently acknowledged within effective altruism, individuals tend to perform better in roles aligned with their interests and capabilities. How should someone with artistic gifts and inclinations incorporate longtermism into their moral framework? How should an ecologist? An anthropologist?

This creates an "allocation gap" in longtermist theory. If someone accepts longtermist premises but cannot contribute directly to x-risk mitigation beyond financial donations, how should they structure their professional life? Minimal longtermism offers no answer; expansive longtermism might, but only if its more speculative claims can be substantiated.

Rather than accept this theoretical gap, I suggest looking at a class of expansive longtermist interventions that I believe are both concrete and defensible. I will call them now-or-never interventions. The general principle is this: longtermism implies we should prioritize interventions whose tractability is uniquely time-sensitive, i.e. cases where opportunities available now will probably be permanently closed, even to the most advanced future societies.

This essay focuses on one such set of interventions: the systematic documentation of biological, linguistic, and cultural information that, once lost, can probably never be reconstructed. Such information includes genomes of undiscovered species facing extinction, ethnographic records of undocumented cultures, and descriptions of endangered languages.

My argument relies on three points. First, if this information is not preserved now, it will be lost forever. Second, information has unique preservation advantages that make it particularly suited for longtermist timescales. Third, this information provides value for future generations.

If these interventions can withstand scrutiny, it would suggest that there are interventions available to individuals that would positively affect the distant future which are not linked to x-factor mitigation, and thus not included in the current minimal longtermist framework.

Some information cannot be recovered once lost.

Consider a physics researcher at CERN investigating quarks. Assuming the laws of physics remain constant over time, future researchers will likely be equally capable of investigating quarks as researchers today. Furthermore, they will likely have superior technology enabling faster, more productive inquiry. An hour of physics research in 2025 may be significantly less valuable, therefore, than an hour of the same research in 3000. The same logic applies to materials science, climatology, and similar fields.

However, certain inquiries are only possible today. Some information, once lost, may become permanently inaccessible regardless of technological advancement.

Consider the genomes of undiscovered species facing extinction. Current estimates suggest that over 80% of Earth's species remain formally undescribed. Many will go extinct before they are discovered by humans, and that genetic information will vanish forever. As Powell notes in "Taking the Long View: Paleobiological Perspectives on Longtermism",^[1] evolutionary history is deeply contingent, stochastic, and fundamentally resistant to causal reconstruction. A future society trying to reconstruct the genome of a species that we never knew existed would require either 1) time travel or 2) a perfect deterministic simulation of Earth's entire biosphere at the molecular level across millennia. It is plausible that both technologies are literally impossible, making genome recovery equally impossible.

By contrast, sequencing these genomes is very possible today. Complete genome sequencing costs are already relatively cheap, and they are declining exponentially. A cost-effective strategy for preserving unrecoverable genomes might involve systematic sample collection in biodiversity hotspots and specimen preservation until sequencing becomes easier, followed by comprehensive genomic documentation. With bacteria and insects, this could be as simple as taking soil samples and freezing them until funds are available (or sequencing becomes sufficiently cheap) for processing.

The temporal asymmetry here is stark. Information easily obtained today is practically (or literally) impossible to obtain in the future. For much of the genomic information currently being lost, it is plausibly now or never.

The same logic applies to linguistic and cultural documentation. Numerous languages face imminent extinction, and many lack written records or digital footprints needed for future reconstruction. Once a language community disappears, many aspects of that language's phonology and grammar become impossible to reconstruct with 100% certainty. Like extinct, undiscovered species' genomes, recreating unrecorded spoken languages with no written records or extant descendants would likely prove impossible even for advanced civilizations. Many such threatened languages exist in Nigeria, the Amazon Basin, and Papua New Guinea.

This applies even more strongly to individual cultures, since each language typically corresponds to at least one distinct culture. Both linguistic and cultural evolution share biological evolution's historical contingency and stochastic principles, making information recovery practically (possibly literally) impossible once lost. Today, however, documenting disappearing cultures and languages is a similarly tractable problem to preserving vanishing genomes. Recording speakers of endangered languages and conducting ethnographies, while less easy than collecting soil samples, remain very tractable interventions requiring modest resources and technology.

This creates a unique opportunity for modern actors. We have the opportunity to collect data today that could never be collected again. 

Information has preservation advantages.

Unlike other potential expansive longtermist interventions, information possesses remarkable preservation advantages, making it ideally suited for long timescales. Information can achieve near-perfect storage fidelity with relatively modest preservation efforts. A genome sequence, linguistic recording, or cultural documentation stored today can theoretically remain accessible millennia hence. Unlike infrastructure, institutions, or cultural capacities, all of which require perpetual maintenance, information needs only the occasional format migration. This fundamentally distinguishes informational interventions as more promising than alternatives.

Consider first the limitations of infrastructure investments. Longtermist timescales potentially span millions of years. Changing geopolitics, plate tectonics, and human needs mean that it is nigh impossible to predict what infrastructure will remain useful over time. Worse, obsolete infrastructure cannot simply be ignored but often requires significant effort and resources to remove, potentially creating net negative value. Historical examples abound. An example from my own research: Roman forts in Syria were once integral to the Silk Road, but eventually faded into such obscurity that many weren't rediscovered until decades after we invented satellites. Their construction would have seemed logical in the 2nd century AD—they developed advanced economies, accelerated global trade, etc.—but would have been an exceptionally poor longtermist intervention.

Interventions that improve cultural capacity or institutional alignment are similarly fragile over extended periods. Sophisticated knowledge requiring coordination is vulnerable to major societal disruptions, even when it's highly valuable. Consider the loss of techniques for creating Roman concrete and Damascus steel, despite their obvious usefulness. Likewise, institutional alignment is vulnerable to many catastrophes, geopolitical and otherwise. For example, a longtermist in the Qing Dynasty promoting better institutions would likely have seen her efforts rendered irrelevant by subsequent political upheavals.

Information demonstrates superior robustness compared to these alternatives. A single genome sequence can be copied across multiple continents, storage media, and institutional repositories at negligible marginal cost. Destroying every copy requires coordinated global effort, while preserving the information requires only that one copy survives. This asymmetry between destruction and preservation costs creates enormous robustness advantages. Even large-scale global catastrophes might spare information while devastating physical infrastructure and institutions.

Consider extreme scenarios: even a pandemic eliminating 99% of the global population would likely leave most digital information intact in storage systems. With modest efforts to create secure information repositories—already undertaken for much of the world's critical data—vast collections could survive devastating meteors, volcanic events, or other natural disasters. Information storage is remarkably resilient. 

The format independence of information provides additional advantages over physical alternatives. While specific storage media become obsolete, underlying information can migrate across technological platforms. From paper to digital, or in a theoretical future where we store data via quantum methods or DNA, the informational content would persist across media. This adaptability contrasts sharply with physical infrastructure locked into particular technological paradigms that may become entirely obsolete.

Finally, information preservation creates substantial option value for future civilizations. Current generations cannot predict which documented species, languages, or cultures will prove most valuable or interesting to future researchers, but preserved information remains available for discovery and application regardless. This differs fundamentally from infrastructure investments that require bets on specific future needs, or institution-building programs that assume particular requirements will persist across cosmic timescales.

Why preserved information generates utility.

The question of what constitutes value (and how to measure it) has occupied moral philosophers for millennia. Rather than rehearse these debates or adopt strong positions on controversial assumptions, I will proceed with a deliberately modest approach that tries to avoid the most contentious theoretical terrain.

Speculative examples struggle to overcome "arbitrariness" objections

Consider a rather speculative example of how this information could provide utility: a researcher preserves substantial genetic data via soil sampling that would otherwise be irrevocably lost. If, in a millennium, Earth's genetic diversity has suffered substantial irreversible losses (very likely), this preserved corpus might prove invaluable for machine learning applications investigating evolution or genetics. The utility gains from enhancing such technologies, integrated across longtermist timescales, could be enormous in biotechnology, medicine, or related fields. Additionally, if de-extinction technologies (already in development) become cheap and widespread, it might even be possible to reconstruct much of the biosphere as it was before the biodiversity crisis, provided the genetic information required is available.

The problem with this hypothetical is that it runs afoul of the "arbitrariness” issue discussed in section 7.3 of "The Case for Strong Longtermism." If we endeavor to assign any probabilities to events like "this otherwise unobtainable genetic data becomes useful in machine learning in 1000 years," we would essentially be inventing numbers from nothing. While I personally believe that preserved genetic data would likely prove valuable in fields like biotechnology and ecology, demonstrating this rigorously while avoiding arbitrariness objections is very difficult, if not impossible.

If you agree that preserved genetic data has a reasonable chance of being useful for future important research—and that this information can't be recovered once lost and can be preserved over time—then you may already find certain expansive longtermist interventions worth pursuing. But if you're skeptical, the argument for this class of now-or-never interventions doesn't depend on such speculation. A more conservative justification based on patterns we can observe today provides adequate reason to take information preservation seriously as an expansive longtermist approach.

Information provides value beyond instrumental uses. 

Many professions provide value primarily by generating interesting information. Consider archaeologists. Their research usually lacks obvious practical applications; it almost certainly won't cure diseases, improve technology, or save lives directly. Much archaeological information is valuable primarily because it is fascinating. Even if you find it unfathomable how 18th-century bottle typologies could be interesting, it's difficult to dispute that researchers engaged in such work find the information valuable.

Or consider an appeal that may be closer to home: readers of moral philosophy (people on the EA Forum, professors of moral philosophy, etc.) are likely familiar with the utility derived from intellectual pursuits that others might find obscure. While I find certain 19th-century philosophers personally unreadable, I suspect many readers of this essay represent precisely the audience who derives genuine satisfaction from such challenging material. For such readers, it should be relatively straightforward to grasp the utility that could flow from, say, a detailed ethnography of a vanished culture that would otherwise be entirely unknown.

Consider how cultural documents providing insights into otherwise lost societies—Beowulf, the Iliad, random cuneiform tablets about sheep transactions in Ancient Sumeria—continue to be valuable to contemporary readers. An ethnographic record of an undocumented Amazonian culture, or a comprehensive linguistic analysis of a language with no written tradition, could plausibly provide similar lasting value. 

I propose that this lasting value could be quite substantial when distributed across the vast populations and timescales that longtermism contemplates. This holds true even under conservative assumptions about how many people find such information interesting and how much they value it.

Quantifying the expected utility of information.

To establish a conservative baseline for how many people might find preserved information interesting, consider that approximately 1 in 50,000 Americans work as archaeologists. These individuals find (often quite arcane) historical information sufficiently valuable to dedicate their careers to its study. This provides an extremely conservative estimate of interest levels in archaeology, since obviously many more people find archaeological discoveries interesting than just those professionally employed in the field. Let this serve as a rough estimate for the proportion of people that might find historical genetics, linguistics, and ethnography interesting.

When estimating how much utility someone might gain from this information, we cannot avoid some speculation. Let's use a deliberately conservative approach and assume that the cumulative output of one researcher's entire career provides utility equivalent to one-millionth of a quality-adjusted life year (QALY) per interested person—about 30 seconds of high-quality life. This 0.000001 QALY takes opportunity cost into account: it represents how much more a future reader would enjoy this information compared to whatever alternative they would have otherwise consumed. This net benefit—the extra satisfaction from reading about an information-preserved Amazonian culture versus, say, a well-documented European one—is what we're calculating as one-millionth of a QALY.

This is obviously a rough approximation. In reality, among those who encounter a single researcher's work, the distribution of utility gained likely follows a power law: a few people might read the work extensively and derive substantial enjoyment, while many others might glance at it briefly or not at all. Some will get significant QALYs from this person's research, others none. However, averaging across all readers to get one-millionth of a QALY per person in the relevant population seems like a reasonable simplification for our purposes. If you disagree, we will use more conservative estimates in a moment. 

Conservative utility estimates across populations.

 In Greaves and MacAskill’s "The Case for Strong Longtermism," the most conservative estimate for total possible future lives is 10^14 people. If preserved information provides the assumed one-millionth of QALY in utility, and reaches the same proportion of this future population as people currently interested in archaeology (1 in 50,000), then this information would yield approximately 2,000 QALYs for one career's worth of information preservation in this population scenario.

These figures scale dramatically with less conservative population projections. The second-most conservative estimate from Greaves and Macaskill (10^18 future lives) produces 20 million QALYs under identical assumptions. Table 1 does these calculations under these assumptions across all population scenarios from "The Case for Strong Longtermism."

Table 1: QALY Estimates Under Standard Assumptions

Assuming 1 in 50,000 people derive 1/1,000,000 QALY from the preserved information.

The objection that these assumed rates of interest and utility are arbitrary and speculative parallels criticisms of longtermism generally. Following the response typically offered for minimal longtermism, we can say that substantially more conservative estimates preserve the argument's essential structure. 

Consider far more pessimistic assumptions: instead of 1 in 50,000 people finding this information valuable, assume only 1 in 50 million will, as historical distance renders most preserved knowledge increasingly obscure. Additionally, assume the net utility this information provides vs. counterfactual entertainment is 1,000 times smaller, about one-billionth of a QALY per interested person affected, or about 3 additional milliseconds of high-quality life. 

Table 2 presents calculations under these extremely conservative parameters.

Table 2: QALY Estimates Under Conservative Assumptions

Assuming 1 in 50,000,000 people derive 1/1,000,000,000 QALY the from preserved information

When you combine the most conservative population estimates with our pessimistic utility assumptions, the QALY totals become very small. This pattern reflects Greaves and Tarsney's central observation: expansive longtermist arguments prove less robust than minimal ones. Consistently selecting the most conservative parameters can render information preservation's impact genuinely negligible. 

However, as population projections increase, even these drastically reduced estimates produce substantial utility calculations. Any movement toward less pessimistic estimates rapidly produces enormous QALY totals.

This is where probability and expected value calculations typically enter into longtermist arguments. While "The Case for Strong Longtermism" explores such reasoning in greater detail (in section 3), the essential logic is straightforward: even when assigning extremely low credences to particular population scenarios, the expected utility can remain enormous. For instance, if one assigns merely a 1-in-billion chance to the Solar System scenario above yielding 10^27 future lives, the expected value of information preservation still reaches 20 million QALYs under our most conservative assumptions.

Table 3: QALY Estimates Under Conservative Assumptions and Low Credences

Assuming 1 in 50,000,000 people derive 1/1,000,000,000 QALY the from preserved information and a 0.0000001% chance of these scenarios happening

Even if you find some of these scenarios wildly unlikely, the potential benefit remains so enormous at large future populations that information preservation as a longtermist intervention still merits serious consideration. 

Objections 

Before addressing two major objections to this framework, it's crucial to note that these high-utility calculations depend entirely on two critical assumptions. If you object to either, my argument fails. 

Assumption 1: this information can be reliably preserved across cosmic timescales. For the reasons discussed above, I argue this is plausible. 

Assumption 2: this information will be permanently lost if we don't preserve it now. If future generations could independently recreate this genetic, linguistic, or cultural knowledge in 1,000 years, then our preservation efforts would only benefit the intervening generations—not the billions across cosmic timescales that longtermists typically consider. The utility would be limited to those 1,000 years, making the expected value much smaller. Only the "now-or-never" nature of these preservation opportunities generates the massive utility calculations that make them compelling from a longtermist perspective.

Two methodological notes also merit attention.

First, a more sophisticated utility calculation might incorporate diminishing returns as preserved information becomes increasingly buried in expanding historical records. In scenarios where humans persist across galactic timescales, the proportion deriving utility from early-era Earth data would likely decline. However, the diminishing curve would need to be extraordinarily steep to reduce, say, 10^34 QALYs to trivial levels. Moreover, information from humanity's technological dawn on its origin planet might retain special significance, partially offsetting such decay.

Second, these calculations treat digital and biological life as comparable in their capacity to derive utility from preserved information. Readers uncomfortable with this assumption may focus on scenarios excluding digital life projections; the argument's essential structure remains intact in purely biological population estimates.

Objection 1: Why not advance knowledge that remains tractable over time?

A natural counterargument is that interventions that remain tractable over time nevertheless advance human knowledge. Consider the particle physics research at CERN discussed above. While quark research will remain possible for cosmic timescales, accelerating the timeline of understanding quarks today might still be valuable. First, it would move forward the moment in time when people might get intrinsic satisfaction from understanding quarks (the same satisfaction they would get from, say, a historical ethnography). Second, it is at least possible that quarks are the key to space travel, fusion energy, or other transformative technologies. They certainly have more potential to accelerate technological progress than recordings of endangered languages.

The best way to address this objection is with Toby Ord's essay "Shaping Humanity's Longterm Trajectory." In it, Ord models humanity's future as a curve where the x-axis represents time and the y-axis represents the universe's total instantaneous value. The total value of the future equals the area under this curve. One can measure interventions' effects by investigating their effects on this area.

Ord identifies several ways to change humanity's trajectory, two of which are relevant to this objection: advancements and gains. "Advancements" shift the entire future trajectory earlier in time—a physicist's work at CERN might accelerate quark discoveries by several months, moving humanity's developmental timeline leftward without altering its shape. "Gains" permanently increase instantaneous value by a fixed amount—preserved genetic data that would otherwise vanish may raise the curve's height across all future time.

Ord's analysis shows that gains scale directly with total time, while advancements only scale with the difference between the universe's current and final instantaneous values. Therefore on very long timescales, gains tend to vastly outperform advancements.

The implications of this are important for now-or-never interventions. Even assuming a CERN physicist could meaningfully accelerate discoveries about quarks (itself questionable given they typically replace rather than supplement existing talent), this advancement would not scale with the timeline’s duration. Meanwhile, preserved information that would otherwise become permanently inaccessible creates fixed value gains extending across the entire future, whether that spans millions or billions of years.

Ord's framework thus supports now-or-never interventions over pure advancement strategies, particularly across large timescales. His math suggests that for advancement-focused interventions to compete with information preservation, they would require extraordinarily large temporal effects, which an individual researcher is very unlikely to achieve. 

If the distinction between advancements and gains is still unclear, I recommend reading Ord's original essay for a more thorough explanation. He explains it with considerably more detail and rigor than this brief summary could offer.

Objection 2: Why not preserve the species, languages, and cultures themselves rather than merely documenting them?

Direct preservation is often impossible. It faces practical and theoretical limitations that, by contrast, do not hinder simple documentation. 

First, consider that a researcher could systematically collect genetic material from thousands of species across biodiversity hotspots, while conserving those same ecosystems would require land acquisition, policy coordination, and resources typically unavailable to individual actors. And even successful conservation cannot address external pressures. Climate change, for example, would cause many extinctions even if conservationists could preserve every biodiversity hotspot on earth.

Similar constraints apply to cultural preservation. Consider indigenous communities in Colombia's Tayrona National Park that face pressure from globalization and tourism. Halting these processes would be extremely difficult, and would likely have ethically dubious second-order effects on the communities in question. Documentation, such as (with the consent of all participants) recording elderly speakers of dying languages, documenting traditions, or preserving oral histories is both more feasible and ethically straightforward.

Beyond practical constraints, direct preservation faces a deeper challenge: genomes, languages, and cultures evolve continuously. Even absent external threats, these phenomena transform over millennia in ways that render permanent "preservation" impossible. Contemporary English bears little resemblance to Anglo-Saxon; human genetics differ substantially from paleolithic populations. On longtermist timescales, documentation represents the only viable strategy for maintaining access to information about particular historical states.

Conclusion

This essay examined whether defensible longtermist interventions exist beyond the "minimal" longtermist focus on technological existential risks. The analysis suggests that targeting certain kinds of information preservation represents a concrete and tractable approach for individuals seeking to positively influence humanity's long-term trajectory.

The argument rests on three key observations: 1) certain information will plausibly be permanently lost if not preserved now, 2) information possesses unique preservation advantages over time, and 3) even conservative utility estimates suggest substantial expected value for preserved information over longtermist timescales.

While these interventions are less robust than minimal longtermist approaches to existential risk reduction, I hope they offer meaningful guidance for the substantial population who cannot contribute directly to technological risk mitigation. They suggest a wider range of professionals (anthropologists, ecologists, linguists) could pursue careers aligned with their capabilities while contributing to humanity's long-term flourishing.

Beyond information preservation, I suspect the "now-or-never" framework may identify other defensible expansive longtermist interventions unrelated to existential risk mitigation. Hopefully, this represents a promising angle of consideration that could further expand practical guidance for longtermist-motivated individuals.

About the author: 

Hi everyone, name is David Goodman. This is my first post on the EA Forum! I’m an American, I’ve been working as a field biologist for the last two years in Africa and South America, and I’m about to start a masters degree at Oxford. I’m interested in many things, including moral philosophy, chess, economics, archaeology, linguistics, evolution, and ornithology :P

This essay was written as part of the Essays on longtermism competition announced here: Announcing: The ‘Essays on Longtermism’ Competition. 

1. ^{^}

   Chapter 11 in Essays on Longtermism