Who Builds 2100? The Demographic Engine of Technological Evolution

The disagreement is 1.4 billion people, not a rounding error

Technology does not invent itself. Every chip generation, every trial-and-error loop in a materials lab, every incremental redesign of a hip joint or a hydraulic valve is the output of some finite number of people who spend some finite number of hours doing it, and that number has a size, a shape, and — this is the part usually left out of both techno-optimist and techno-pessimist writing — a trajectory that is now bending toward its own historical maximum. This article is not a population-panic piece in either direction. It does not argue that more people is straightforwardly better, and it does not argue that fewer people means collapse. It argues something narrower and more checkable: that technological evolution is an input-output process with population on the input side, that the size and age structure of that input are genuinely disputed by billions of people among the demographers who study it, and that the dispute matters for who ends up inventing, building, and staffing the technology of 2100.

Start with the dispute itself, stated plainly, because it is usually flattened into a single scary or reassuring number when it is actually a live disagreement between two well-resourced institutions using different models. The United Nations Population Division’s 2024 revision of the World Population Prospects — the world’s standard demographic reference, produced by the body every national statistical office ultimately reports into — projects that global population will keep growing for another fifty to sixty years, reaching a peak of around 10.3 billion in the mid-2080s, up from 8.2 billion in 2024, before declining gradually to about 10.2 billion by 2100 ^[1]. The UN puts the probability that the peak occurs within this century at 80 percent, with the peak itself falling somewhere between the mid-2060s and 2100 ^[1]. That 10.2 billion figure carries real uncertainty even within the UN’s own accounting: a 95 percent prediction interval running from 9.0 billion to 11.4 billion ^[1].

Set against this is the Institute for Health Metrics and Evaluation’s 2020 forecast, published in The Lancet by Stein Emil Vollset and the Global Burden of Disease collaborators, which projects a global population peaking earlier, in 2064, at 9.73 billion, and declining faster and further, to 8.79 billion by 2100 ^[2]. The UN’s own 2024 report does the arithmetic on the gap directly: IHME’s 2100 figure sits 1.4 billion people, almost 14 percent, below the UN’s ^[1]. A third major projection, from the Wittgenstein Centre for Demography and Human Capital, lands in between at 9.9 billion for 2100 ^[1]. Three serious institutions, using the same historical data and the same cohort-component method, land on three different central estimates for the same single year, and the two extremes of that range do not overlap until well past a billion people apart.

The reason the estimates diverge is not sloppiness on anyone’s part; it is a genuine methodological fork over how to model the one variable that dominates the long-run outcome, fertility in the countries that have not yet finished their demographic transition. The UN’s approach projects future fertility from historical levels and trends, treating past progress in development, education, and contraceptive access as already embedded in the trend line, and its medium scenario further assumes continued progress toward gender equity lets people better realize existing childbearing intentions ^[1]. IHME’s model instead builds fertility forward from assumptions about women’s future educational attainment and the “satisfied demand for contraception” — a completed-cohort-fertility model in which those two variables alone are reported to explain 80.5 percent of the historical variance the model is fit to ^[2]. Put simply: the UN extrapolates a trend that already contains development’s past effects, while IHME re-derives fertility from a forward-looking model of how fast development itself will proceed, and it expects that proceeding faster. Both approaches are legitimate; they simply encode different bets about how quickly girls now being born in the highest-fertility regions will get more schooling and better contraceptive access than the historical trend line assumes. Because more than 80 percent of the UN’s higher end-of-century population sits in the group of countries still growing past 2054, and because sub-Saharan Africa alone is projected by the UN to reach 3.3 billion by 2100 versus a much lower trajectory under faster-development assumptions, that one regional bet is doing most of the work in the 1.4-billion-person gap ^[1].

The honest reading of this dispute is not that one institution is right and the other is spreading misinformation. It is that 2100’s population is unknown by an amount most public discussion of demography does not register: a spread wide enough to contain the difference between a world with a slowly declining but still-enormous 10.2 billion people and one nearly 1.4 billion smaller. Both trajectories agree on the more important structural fact this article is built on — that the number of people on the planet this century, for the first time in the whole of recorded history, is going to stop rising and start falling — and it is that shared structural fact, not the exact digit at the end of it, that the rest of this article treats as load-bearing.

Low fertility spreads like a cultural trait, and coercion does not reverse it

Both forecasts agree that fertility decline is the dominant driver, which raises the prior question: why does fertility fall, and does it keep falling once it starts. The demographic transition — fertility and mortality both dropping as a society industrializes — used to be told as a mechanical story: children stop being economically useful, contraception becomes available, fertility falls as a rational response to changed incentives. That story is not wrong so much as incomplete, because it does not explain why the transition, once it starts in one subpopulation, spreads to neighboring subpopulations that have not experienced the same economic change, nor why fertility keeps falling well past the point any income-and-incentives model would predict it should stabilize.

Cultural evolutionary demography treats the transition as exactly what an evolutionary biologist would recognize: the spread of a transmissible trait through a population via social learning rather than genetic inheritance, propagating along an S-shaped adoption curve the way any successful cultural innovation does. Heidi Colleran’s 2016 review in Philosophical Transactions of the Royal Society B lays out the mechanism explicitly, distinguishing three separable phases — the origin of low fertility in some initial subpopulation, its spread to others, and its subsequent maintenance — each plausibly driven by different processes at different scales ^[9]. Spread happens disproportionately through prestige-biased transmission: high-status individuals who adopt smaller-family norms are disproportionately imitated, giving a small number of early adopters outsized, one-to-many influence over the trait’s frequency in the wider population, the same asymmetric-copying dynamic that drives fashion, language change, and technology adoption generally ^[9]. This is not incidental to the transition; it is most of why fertility decline moves through a population faster than any change in the underlying economic incentives could explain on its own, and why it moves through subpopulations — urban professionals before rural laborers, the highly educated before the less educated — that share status networks rather than merely similar incomes.

Nowhere has the trait’s frequency fallen further or faster than East Asia. South Korea’s total fertility rate — averaging roughly 0.8 children per woman across the 2018–2023 period tracked by the World Bank, and bottoming at 0.721 in 2023 before a modest recovery ^[7] — is the most extreme sustained sub-replacement rate ever recorded for a large national population outside wartime, at less than a third of the roughly 2.1 needed for a population to hold its size without migration. Statistics Korea’s own reporting shows the rate ticking up again to 0.799 in its most recent annual figures, a genuine and worth-watching reversal, though still deep in territory no advanced economy had previously occupied for a sustained period ^[8]. The transition, once fully underway, does not gently approach replacement level and stop; it overshoots, and Colleran’s framework predicts exactly this kind of overshoot as a maintenance-phase dynamic distinct from the forces that caused the initial decline.

The natural next question for anyone who thinks a below-replacement trait spreading through a population is a problem worth intervening on is whether the state can simply reverse it by decree, and the twentieth century ran that experiment to failure in Romania. Nicolae Ceaușescu’s regime banned abortion by decree in 1966, at a time when abortion had been the primary method of birth control in the country; Cristian Pop-Eleches’s study of the policy, published in the Journal of Political Economy, documents that the birth rate doubled the following year ^[10]. That immediate spike looks, at first glance, like proof that coercive pronatalism works. It is not: Pop-Eleches’s central finding is that the ban’s effect operated overwhelmingly through composition rather than through a genuine increase in desired family size — educated, urban women had disproportionately relied on abortion before the ban and continued to seek ways around fertility control afterward, so the additional births were concentrated among women with fewer resources to invest in each child, and the children born into that compositional shift went on, once analysis accounted for who was actually being born, to have worse educational and labor-market outcomes as adults than the cohorts before and after them ^[10]. A policy that moves the birth rate by suppressing the means of birth control, rather than by changing anyone’s underlying fertility preference, does not touch the cultural trait Colleran’s framework describes; it only distorts, temporarily and at a measured human cost, who ends up bearing the children that trait would otherwise have prevented.

That said, cultural selection on fertility itself is not a one-way ratchet toward zero, and treating it as one is the mirror-image error to population panic. Colleran’s own framework makes selection’s counter-pressure explicit: under high-fidelity vertical transmission — parents successfully passing their own fertility preferences and behaviors to their children — high-fertility lineages simply produce more of the next generation than low-fertility lineages do, so left alone, evolutionary selection continually pulls back toward higher fertility, and the low-fertility trait’s persistence in a population requires continual cultural innovation and lateral re-transmission to keep outrunning that pull rather than reflecting some settled endpoint ^[9]. Pew Research Center’s demographic projections give this dynamic a concrete, sober empirical anchor rather than a rhetorical one: fertility differences correlated with religious affiliation — 3.1 children per woman among Muslims and 2.7 among Christians in the 2010 baseline, against 1.7 among the religiously unaffiliated and 1.6 among Buddhists — are projected to drive the Muslim share of world population up 73 percent between 2010 and 2050, more than double the 35 percent growth rate of the global population as a whole ^[12]. None of this is a claim about any specific religion’s future dominance; the honest reading is narrower and more mechanical, the same reading a population geneticist would give a selection differential — whichever subpopulations retain higher fertility, for whatever combination of cultural, religious, or economic reasons, mechanically make up a growing share of every subsequent generation, which is exactly the long-run dynamic that keeps the demographic-transition literature from treating “fertility converges to some low universal number and stays there” as a safe assumption.

Growth is research effort times research productivity, and effort is running out of people

Set the dispute over the exact 2100 headcount aside for a moment and ask the question this article actually cares about: what does a population that stops growing, and starts aging, do to the rate at which new technology gets invented. The economics of long-run growth has a specific, quantitative answer to this question, developed originally by Charles Jones and tested empirically by Nicholas Bloom, Charles Jones, John Van Reenen, and Michael Webb in a 2020 American Economic Review paper whose opening line states the whole argument of this section: in the standard class of growth models, “the long-run growth rate is the product of two terms: the effective number of researchers and their research productivity” ^[3].

Semi-endogenous growth theory, the framework Jones developed and that his 2022 Annual Review of Economics piece surveys, formalizes this into a relationship between population growth and the long-run growth rate of ideas ^[4]. Population enters the idea-production function as the stock of people available to search the technological fitness landscape for improvements, and because ideas are non-rival — one firm’s discovery of a better transistor layout does not use up the possibility for another firm to discover a different one — expanding that search population raises the rate of discovery roughly in proportion to its own growth, discounted by how much duplicated effort dilutes the return to any one searcher and by how much each new discovery makes the next one to find. The canonical form of the relationship is

where $n$ is the growth rate of the population of researchers, $\phi$ measures how much easier or harder each new idea makes the next one to find (values below 1 are required for the model’s growth path to be stable at all — the discipline’s own name for this is diminishing returns to knowledge accumulation), and $\lambda$ measures how much duplicated research effort — different teams stumbling onto the same idea — dilutes each additional researcher’s marginal contribution. The formula says something that should be intuitive from the fitness-landscape framing: if the population of people searching for new technology stops growing, $n$ falls toward zero, and the long-run growth rate of ideas falls with it, one for one, unless something else changes $\lambda$ or $\phi$.

Bloom, Jones, Van Reenen, and Webb’s contribution was to show this is not merely a theoretical curiosity but the empirically dominant pattern across every domain they examined. Their headline case is Moore’s Law: the doubling of transistor density on a computer chip every two years, a remarkably steady 35 percent annual growth rate sustained for half a century, has been achieved only by continuously enlarging the population of semiconductor researchers, to the point that the number of researchers required today to sustain the same doubling is more than 18 times larger than the number required in the early 1970s, with research productivity in the field declining at an average rate of 6.8 percent per year even as the output — chip density — kept its steady exponential pace ^[3]. This is diminishing returns behaving exactly as an evolutionary biologist would expect a fitness landscape to behave: the easy improvements, the ones a small number of researchers can find quickly, get found and exploited first, and each subsequent generation of the same technological lineage requires searching a larger, more thoroughly picked-over space for a smaller remaining gain. The pattern is not specific to semiconductors. Research productivity for crop-seed yields — corn, soybeans, cotton, wheat — declines at roughly 5 percent per year even as agricultural R&D spending rises; mortality-improvement research for cancer and heart disease shows a similar rate of decline; a Compustat-based analysis of research productivity within individual firms finds it declining in more than 85 percent of the sampled companies, at an average rate of about 10 percent per year; and even at the level of the aggregate US economy, taking the most extreme possible view of where new ideas could hide, research productivity has fallen by a factor of 41 since the 1930s, a 5.1 percent average annual decline, while the effective number of researchers rose by a factor of 23 over the same period just to keep aggregate growth from collapsing ^[3]. The paper’s authors trace the origin of this specific aggregate-economy observation to Jones’s own 1995 growth model, of which their empirical exercise is, in their words, simply another way of looking at the same underlying point ^[3].

None of this says technological progress stops. It says progress has been purchased, for decades, by feeding an ever-larger population of researchers into a search problem with steadily worsening returns — which is precisely the arrangement that a population no longer growing, and in some read of the evidence above shrinking after mid-century, cannot keep doing indefinitely. If $n$ falls and $\lambda$ and $\phi$ stay where the historical data put them, the semi-endogenous growth relation says the rate of technological advance falls with it. That is not a prediction this article is making outright — Sections 4 and 5 below identify two separate mechanisms that could push back against exactly that conclusion — but it is the default, unmitigated arithmetic, and any claim that a shrinking population is technologically neutral has to explain why one of those mitigating mechanisms will be strong enough to matter.

An aging population is already buying the machines a shrinking one cannot staff

The demographic transition does not just shrink tomorrow’s population of researchers; it inverts the age structure of every population it passes through, and that inversion creates its own selection pressure on which technologies get built — not because researchers vanish, but because the market that used to demand young workers now demands something to substitute for the ones who no longer exist. Where the previous section treated population as an input to the search for new ideas, this section treats an aging population as a demand shock that determines which niches in the technological landscape get an evolutionary push to fill.

The UN’s own aging accounting states the mechanism plainly: in the group of countries whose population has already peaked, the share of people aged 65 or older is projected to nearly double, from 17 percent in 2024 to 33 percent by 2054, eventually reaching around 40 percent by 2100, while the working-age share that has traditionally supplied both labor and care shrinks correspondingly — and the UN’s own policy language for what to do about it names the same substitution this section is about: “leveraging technology, including robotics, automation and artificial intelligence” ^[1]. Japan’s Cabinet Office reports the leading edge of this curve directly: 29.0 percent of Japan’s population was aged 65 or older as of its most recent annual accounting, alongside a population aged 65–74 of 16.87 million that has already been overtaken in size by the 75-and-older cohort — meaning Japan’s aging population is not merely large, it is old even by the standards of an aging country ^[14].

Japan is also the best-documented case of what actually happens to robot adoption once that demand pressure hits an industry, and the honest record is neither the sleek robot-caregiver future of consumer press coverage nor a story of nothing happening. Karen Eggleston, Yong Suk Lee, and Toshiaki Iizuka’s National Bureau of Economic Research study of Japanese nursing homes, using prefecture-level government subsidies as a natural experiment, documents genuine and fairly rapid growth: the share of nursing homes reporting use of any type of care robot rose from 17.6 percent in 2016 to 26 percent among the 857 facilities in their 2017 survey ^[11]. But growth from a low base is still a low base — an early phase of the adoption S-curve, not evidence of an already-completed radiation of automation into every corner of the industry. Monitoring robots, which watch for falls or unauthorized bed exits, were the most common single category at 14.9 percent of facilities; transfer-aid robots, the kind that physically lift a resident from bed to wheelchair, were adopted by only 7.7 percent (split between 4.7 percent wearable and 3.3 percent non-wearable designs); mobility robots reached 5.3 percent, and communication robots just 2.8 percent ^[11]. The subsidy program behind this growth was substantial by 2017 — covering almost three-quarters of the nursing homes in the study’s sample — and the paper’s central empirical finding cuts against a simple labor-substitution story: robot adoption increased employment, largely by adding flexible-contract care workers and nurses and by reducing the reported difficulty of staff retention, rather than shedding jobs ^[11]. This is the honest evidentiary basis for treating Japan as a frontier case rather than an already-arrived future: real, subsidized, empirically documented adoption, concentrated in the least physically demanding robot categories, still a minority practice even in the single country with the oldest population and the most aggressive government support for exactly this substitution.

The broader empirical pattern generalizes beyond Japan’s nursing homes. Daron Acemoglu and Pascual Restrepo’s study of demographics and automation, using both cross-country data and variation across US commuting zones, finds that aging — specifically, a rising ratio of older to middle-aged workers — is associated with significantly greater adoption of industrial robots and other automation technologies, concentrated in industries and regions that had previously relied most heavily on the middle-aged workers now becoming scarce ^[5]. The mechanism the two forms of evidence share is the same one an ecologist would recognize: automation is not colonizing every niche uniformly, it is radiating specifically into the niches that population aging has vacated of the workers who used to fill them, at a pace set by how urgently and how completely that vacancy bites — fastest and most completely, so far, in the lowest-physical-and-cognitive-demand categories, slower and shallower everywhere the vacated task is harder to substitute for.

The wildcard: if the researchers are machines, the arithmetic changes

Section 3 established that the standard growth arithmetic ties the rate of invention to the growth rate of the human population doing the inventing. That arithmetic has one obvious escape hatch, and it is worth stating precisely who argues for it and on what terms, because “AI will fix the demographic problem” is exactly the kind of claim this publication’s editorial rules require attributing to a named source rather than asserting as settled fact.

The named argument belongs to Philippe Aghion, Benjamin F. Jones, and Charles I. Jones, whose 2017 National Bureau of Economic Research paper works out formally what happens when artificial intelligence automates tasks inside the idea-production function itself, rather than only inside the production of ordinary goods. Their result, worked through with an idea-production function in which a fraction of research tasks can be automated, is that in the case without automation their model collapses to exactly the semi-endogenous relation from Section 3 above — the growth rate of ideas proportional to the growth rate of the human research population, discounted by diminishing returns — matching what they describe as the original point of Jones’s 1995 model ^[6]. Automation changes the equation by adding a second term:

where $g_S$ is the growth rate of human research labor, exactly as before, and $g_C$ is a new term generated by the continuous automation of research tasks — one that can be strictly positive even when $g_S$ is zero, because a fixed number of human researchers can still be automated onto an exponentially shrinking share of remaining, not-yet-automated tasks, concentrating human attention ever more narrowly on the residual problems machines cannot yet touch while machines handle a growing share of the rest ^[6]. In plain terms: if enough of what a researcher actually does — running the experiment, searching the literature, proposing the next candidate molecule, checking the math — can be handed to a machine, population no longer has to be the thing that’s growing for the idea-production function to keep growing too. Aghion, Jones, and Jones are explicit that this is one channel among several they model, and that full automation of the idea-production function is a mathematical limiting case — one route, among the several “singularity” scenarios their paper works through formally, to an unbounded growth rate, which they treat as a theoretical possibility to be checked against evidence rather than a forecast ^[6].

This is a live argument, not a settled one, and it deserves the same falsifier discipline this article has applied to the population dispute. The claim, stated as a prediction: if AI-driven automation of research tasks is making a material, positive contribution to $g_A$ through the $g_C$ channel described above, that contribution should show up as a rising, not merely stable, count of significant scientific or technical discoveries directly and substantially attributable to AI systems performing research tasks rather than merely accelerating human-directed computation, sustained over multiple years rather than a single showcase result. The clearest indicator available so far is a vendor claim that deserves exactly that label: Google DeepMind reports that its GNoME system identified 2.2 million candidate crystal structures, of which 380,000 were assessed as sufficiently stable for further investigation, a scale the company describes as equivalent to “nearly 800 years” of prior human-paced materials discovery, against a baseline of roughly 20,000 stable crystals found by all human experimentation to date and roughly 48,000 more from prior computational screening methods, with 736 of the predicted materials reportedly synthesized independently by external researchers in concurrent work ^[13]. Horizon: through 2032. The prediction is disconfirmed if AI-attributed contributions of this kind fail to appear in additional, independently verified domains beyond materials screening — drug candidate discovery, mathematical proof search, novel experimental design — over the intervening years, or if the rate of independently verified, practically consequential AI-originated discoveries plateaus rather than continuing to rise; it is supported if the pattern generalizes and compounds. Nothing in the evidence gathered for this article settles the question either way today, and Aghion, Jones, and Jones’s own paper is a model of what the arithmetic would look like if the mechanism holds, not proof that it does.

Four boxes for 2100, built by crossing who’s alive with who’s inventing

Put the two live disputes together — where the population lands, and whether research productivity keeps declining on its historical path or gets a machine-augmented boost — and the honest picture of 2100 is not one forecast but a small matrix of scenarios, each internally consistent, each resting on assumptions this article has tried to state rather than bury.

Reading the table left to right along the top row: under the UN’s higher population trajectory and a continuation of Bloom, Jones, Van Reenen, and Webb’s documented historical pattern of declining human research productivity, the growth rate of ideas keeps decelerating, but more slowly than under any lower-population path, because a larger absolute pool of researchers, even at declining individual productivity, still generates more total research effort than a smaller one — the same logic that let the aggregate US economy sustain steady growth for decades by tripling its effective researcher count even as per-researcher productivity collapsed ^[3]. Assumptions: the UN’s medium-fertility scenario materializes; no AI-driven automation of research tasks reaches the scale Aghion, Jones, and Jones’s $g_C$ term describes. Indicator: aggregate research-effort growth (R&D spending deflated by researcher wages, or head-count of researchers in national science statistics) continuing to rise in step with population, without an accompanying discontinuity in AI-attributed discovery counts. Horizon: 2050. Falsifier: either the UN’s own population trajectory is revised down toward IHME’s in a subsequent World Population Prospects revision, or AI-attributed discovery counts of the kind Section 5 describes rise sharply rather than staying flat — either would move this article’s classification of 2100 out of this box.

The bottom-left box is the one most population-panic commentary implicitly assumes without saying so: IHME’s lower population path combined with no material AI offset to declining research productivity. This is the box in which Section 3’s arithmetic bites hardest and Section 4’s automation response is a partial, not full, substitute — Eggleston, Lee, and Iizuka’s own Japanese data shows robot adoption rising but still a minority practice even under strong subsidy, and Acemoglu and Restrepo’s cross-country evidence shows automation responding to aging with a lag rather than pre-empting it ^{[11, 5]}. Indicator: falling absolute (not merely per-capita) counts of active researchers in national science-workforce statistics for major research-producing countries, sustained for at least a decade, without a compensating rise in AI-attributed research output. Horizon: 2060. Falsifier: absolute researcher counts stabilize or rise despite a declining population — plausible if a larger share of a smaller population moves into research occupations, which Section 3’s arithmetic does not rule out — or AI-attributed output rises enough to move this scenario into the bottom-right box instead.

The two right-hand boxes both depend on the Aghion-Jones-Jones mechanism actually materializing at the scale their formal model allows for, which Section 5 was careful to label a named, live argument rather than a documented fact. If it does materialize, the difference between the top-right and bottom-right boxes — whether population is also growing or is instead the IHME-scale decline — matters much less than in the human-bound cases, because a large enough $g_C$ term dominates the sum in the numerator of the growth relation regardless of what $g_S$ is doing. That is precisely Aghion, Jones, and Jones’s point about their own model: sufficiently complete automation of the idea-production function decouples the growth rate of ideas from the growth rate of the human population searching for them, which is the specific, named mechanism by which a smaller, older 2100 population could still preside over accelerating rather than decelerating invention ^[6].

None of the four boxes should be read as this article’s forecast, for the same reason the population dispute in Section 1 resists a single number: the evidence assembled here is compatible with elements of more than one box holding in different technological domains simultaneously, materials science and drug discovery moving fastest toward the machine-augmented boxes while fields with less digitizable experimental loops stay closer to the human-bound ones. What the matrix is for is narrower and more useful than a forecast: it names, in advance, the two independent axes — population path and research-productivity path — that a reader checking back over the coming decades should track separately, rather than collapsing into a single vague sense of whether “technology is still progressing,” because each axis has its own named forecasters, its own indicators, and its own way of being wrong.

The sentence this article has been building toward

The technosphere of 2100 will be built by a population smaller, older, and more machine-augmented than today’s — and every term in that sentence now carries a named source rather than an intuition. Smaller relative to the trajectory humanity has been on for two centuries: even the UN’s higher-population path has global population declining after the mid-2080s peak, a first in the whole of recorded demographic history ^[1], and IHME’s lower path puts the 2100 figure 1.4 billion below that ^[2]. Older: the share of the population past 65 in the demographically leading countries is set to nearly double by 2054 and to approach 40 percent by 2100 in the group of nations whose population has already peaked, a shift the UN’s own report treats as the defining structural fact facilities like Japan’s are already living through ^{[1, 14]}. And more machine-augmented, in a strictly dual sense this article has kept separate throughout rather than collapsing into one story: machines built to substitute for a labor force that aging has thinned, as Acemoglu and Restrepo’s demographics-and-automation evidence and Eggleston, Lee, and Iizuka’s documented, still-modest Japanese nursing-home adoption both show happening today ^{[5, 11]}, and machines potentially substituting for the human researchers a shrinking population can no longer supply in growing numbers, a mechanism Aghion, Jones, and Jones have modeled formally and named honestly as one channel among several rather than a certainty ^[6].

What should not survive this article is the framing that made it necessary to write: that population size is either straightforwardly good, because more people means more inventors, or straightforwardly bad, because a “graying” world is written off in advance as stagnant. The semi-endogenous growth arithmetic in Section 3 says a shrinking research population is a real drag on innovation, not a rounding error — Bloom, Jones, Van Reenen, and Webb’s own decomposition makes that the default case that has to be argued away, not assumed away ^[3]. But Sections 4 and 5 identify two separate, independently documented mechanisms — demand-driven automation filling labor niches aging has vacated, and a live, formally modeled but unproven possibility that automation reaches into the research process itself — that could each partially or fully offset that drag, on timelines and through channels this article has tried to make checkable rather than rhetorical. Cultural evolution set the size and shape of the population that will do the inventing; economic growth theory says what that population’s size implies for the rate of invention; and the honest 2100 forecast is the four-box matrix above, not a single confident number in either direction.