What New Jobs?

Against the Dark Forest Theory of the Labor Market

Everyone admits that new technologies can eliminate specific jobs,¹ but a much smaller fraction of thinkers believe that technology can reduce the overall demand for human labor. Technology might replace some workers or indirectly undermine their usefulness by supplanting entire industries, but those same new technologies invariably create novel employment demand, too, and those newfangled careers will necessarily offset whatever satisfaction of labor demand is achieved by rolling out the new tech: it’s tough to get a job as an elevator operator today, but also you couldn’t land a job as a software developer a hundred years ago. This turnover balances out and self-corrects over the long run. So the theory goes.

The backbone of the techno-optimist rebuttal concerns this promised influx of jobs that new inventions will transmute from nonexistence into economic reality. However, whether these inventions will stimulate some new labor demand is uninteresting: the vital question is whether, as the techno-optimists presume, the laws of economics guarantee that the labor demand generated by new tech will outpace the expiration of current employment opportunities (at least, supposing a flexible time horizon). If so, maybe the fearless techno-vanguard should offer their logic up for some more exacting analysis before we rely so unhesitatingly on its comfort.

The Dark Forest Theory of the Labor Market

Whenever people minimize fears about durable tech-based unemployment, they partially blame our temptation to panic on an asymmetry between the obviousness of jobs that are ripe for elimination and the unavoidably more mysterious character of the future jobs that new tech will create. This asymmetry, they suggest, engenders too much negativity and economic myopia, and history is stuffed with misguided freak-outs over stuff like looms and electric lights eradicating work, and because passing up those technological advances would have been so injurious and dumb, now any worries about long-term tech unemployment are immediately classified as dangerous and ahistorical fear-mongering. Even if some individuals were negatively affected by new inventions rewiring the labor market, society plainly benefited in the long run.

The techno-optimist case against the job market going haywire isn’t just that people underestimate the magnitude of future labor demand, though—it’s that the jobs aren’t even really ascertainable in theory: you have to trust that they’re out there, populating the ethereal realm of untapped economic potential. This is The Dark Forest Theory of the labor market: a popular, unflinching conviction that innovation cannot permanently erode the demand for human labor because there’s always an inexhaustible (but also invisible) variety of newly profitable use cases for human effort beyond the technological boundaries of society's present conditions.

Techno-optimists could likewise argue that trying to outline the workforce of the tomorrow involves a communist-like hubris that overestimates the predictability of economies and their participants and that anticipating the labor market ramifications and myriad side effects of technological leaps is just too challenging—it’s an undertaking best reserved for Laplacian Demons. Combine this overarching limitation on economic forecasting with a tendency to overlook the potential demand from jobs that haven’t materialized yet, and it’s unsurprising that doomsayers systematically misjudge the balance of opportunities that naturally arise from labor-supplanting tech.

But this Dark Forest view overstates the discrepancy between identifying the outmoded jobs and futuristic jobs in both directions. Firstly, we can’t appreciate and delimit the scope of endangered work with such great precision. While some careers have plainly been jeopardized (I wouldn’t be too upbeat about matching at a radiology residency program right now), other professions will be crushed by AI that currently appear insulated from it. The chaotic subtleties of higher-order effects obfuscate the implications for both the supply and demand for labor. It wasn’t obvious that smartphones could eventually threaten taxi services, for example. Moreover, the superior efficiency of AI creates motivation to reconfigure workloads so that it can infiltrate more of the business than is immediately accessible to it. The bulk of automation is actually achieved by figuring out how to reshape tasks to make them conducive to automation rather than upgrading the automated toolset to be super flexible.²

Secondly, it’s untrue that future labor demand is perfectly opaque. Even if trying to anticipate the gamut of businesses (and specific corresponding jobs) made possible by AI is a fool’s errand, we can fruitfully survey the routes toward generating more labor demand and hazard some coarse-grained conjecture about the labor market’s future conditions. Fortune-telling is impossible, but if some pedestrian economic reasoning and stylized history can discredit fears about the labor market collapsing, it could also provide a usable foothold for revivifying them.

It wouldn’t have been mere guesswork to expect that replacing horses with cars would create jobs in the auto sector, and it seems obvious in retrospect that computers would boost the demand for software developers, right? Deciphering the likelihood of overarching labor market trajectories doesn’t require pinpointing specific companies that will thrive or anything, so why can’t the techno-optimists furnish convincing examples from this stockpile of new use cases for human labor that AI will unleash, even though everyone is so confident about their existence?

Development and Infrastructure

One reason could be that AI won’t generate many first-order labor opportunities. If we survey the mechanisms by which innovation can drive labor demand, the most obvious/direct route is via the labor inputs required to produce, deploy, and maintain those new technologies. But will producing, deploying and maintaining AI require a lot of well-paid human effort? Not nearly enough: there’s no way the legions of desk-jockeys who are going to be out-competed by LLMs will be conveniently offset by breakneck hiring at places like Anthropic.

Mind-bending sums of investment are feeding into AI ventures right now, but presumably that’s channeling into hardware and compensation wars for superstar talent, as bleeding-edge firms jostle for the most uber-elite cadre of specialists to propel their research rather than conduct a scatter-shot hiring bonanza; businesses like that will gladly overwork a small platoon of standouts in exchange for eye-popping salaries before hiring a bigger staff for more pedestrian pay (you see this same strategy in other high-prestige outfits like big law firms). And even this handful of coveted super-nerds are sealing their own fate, ultimately, since the AI will eventually become self-improving and is already on the verge of functioning like a software dev. Those super-nerds behind AI will continue growing rich from the appreciation of their equity (their hard work will literally pay dividends), but they’ll no longer survive by drawing a salary.

What about building the requisite infrastructure for AI—shouldn’t that elicit a sizable spike in labor demand from the markets? For example, workers might be employed to operate dedicated nuclear facilities for the premier LLMs in order to churn out the unprecedented wattage burned up by these data-crunching armies of ultra-pricey GPUs. Well, that’s… something, but again, I expect this to be a pitiful fraction compared with the sweeping graveyard of careers that AI is positioned to bring about.

Technological and Economic Limits on Business Ventures

AI doesn’t straightforwardly promise to offer humanity an array of useful new powers, either. Unlike some other inventions, AI’s innovative payout is mainly concerned with making extant tasks more automatic rather than accomplishing unimaginable feats. Before the airplane, commercial air travel wasn’t just unprofitable; it was downright impossible. Surmounting that technological boundary was needed for people to have careers as commercial airline pilots, air-traffic controllers, etc. So far, the magic of LLMs has more to do with replicating human effort than doing the impossible, like back when Deep Blue beat Kasparov—the shock was primarily about who beat Kasparov rather than about someone managing to beat him.

The way for AI to increase labor demand without really changing what’s technologically feasible would be to make more businesses economically feasible. This is a somewhat wobbly division, but efficiency windfalls can convert dormant tech into profitable enterprises by making the business models more defensible. It wasn’t obvious that smartphones would yield the monstrous economic gains that mobile software companies eventually accumulated, and computer game companies are nonsensical ventures in a world wherein people don’t own personal computers. Can AI jump-start a new ecosystem of business ventures?

Plenty of start-ups are trying to leverage AI by providing different setups and points of ingress (sometimes called “wrappers”) to the big LLMs, and established businesses are strangely ornamenting their products with totally superfluous AI hangers-on, but people still rely on phones and laptops, and the salad of applications we use has ossified. Even if, by some miracle, the tech guys—who can’t seem to read the fucking room—coax everyone into interacting with AI by wearing giant computer goggles or whatever, we’d still be using (mostly) the same kinds of apps, except it’ll be the one developed to work on goggles instead of the desktop version, the web app version, the mobile app version, or the version inexplicably designed to run on your microwave or washing machine. Instead of fostering a lively arena for businesses (composed of humans) to offer previously unimaginable products and services, it’s more obvious how AI could be substituted for software users themselves or serve as a monopolistic choke point between humans and online activity; viz., it looks more probably that AI mutates into the mythical, winner-take-all Super App that VCs drool over than another new platform.

Even if the equities markets become littered with software companies predicated on widespread AI usage, however, the human effort fueling those companies is itself a prime target for replacement; so again, this mechanism isn’t going to motivate longstanding and robust labor demand. Even without AI devs, there’s already too much software and a glut of new CS grads. Sure, if AI can positively alter the scope of what’s commercially viable, then that could inspire a flowering of business endeavors and entrepreneurial verve (and therefore potentially some hiring and employment), but presumably AI will redraw the boundaries of economic feasibility mostly by reducing headcount. If more businesses are launching or surviving only because they’re operating with skeleton crews, then that should temper everyone’s sunny outlook about how this all nets out for the labor market.

Technology Trees

So, AI is unlikely to generate many jobs directly—it’s an invention specifically geared toward magnifying the efficiency of firms by shrinking staff and outperforming humans at office work rather than by enabling people to accomplish wholly novel activities like soaring through the air or communicating instantaneously across the globe. But maybe AI is a stepping stone to other technologies that really do augment the capabilities of human beings. In that case, AI would be a vital node in humanity’s tech tree, and some newly possible businesses could—unlike AI—actually require a bunch of human labor to maintain, deploy, etc.

Harnessing electricity, for example, induced a need for workers to produce the electricity and deploy the accompanying infrastructure, but electrification primarily boosted labor demand by unlocking other technologies and efficiencies: usage of the electricity supplied was a greater labor-market tailwind than the process of supplying it. Is AI a necessary building block for other inventions? As discussed supra, because AI is primarily emulating human activities, it’s not super obvious that AI would function like a missing puzzle piece or jumping-off point in engineering serious atoms-not-bits advancements. If the goal of AI right now is capturing the ability to mimic domain experts, and those experts already have powerful computing resources at their disposal, then the key question is what new capacities can AI provide that were unobtainable using human-computer pairs.

The more promising tech-tree hypothesis is that AI will simply catalyze innovation by aiding or guiding the discovery process in general. Unfortunately, science has diminishing returns, and the majority of low-hanging fruit has been plucked already—there are simply much fewer world-changing innovations still outstanding. Something like room-temp superconductors would plainly confer new abilities and important engineering options, and it’s imaginable that AI helps fabricate such a thing, but it could also just be physically impossible. Alternatively, AI-assisted research could simply achieve marginal advancements in a speedier and less costly way than humans alone. AI could facilitate a bunch of medical improvements and accelerate drug discovery, for example, but even though that undoubtedly betters the human experience, it doesn’t straightforwardly culminate in a massive uptick in hiring.

Conclusion

Per the Dark Forest Theory of the labor market, no one can perceive the entire jumble of business opportunities presented by upcoming innovation with enough accuracy to justify pessimism—history is too densely riddled with misfires from the most sophisticated investors and celebrated thinkers about this stuff, like when Paul Krugman famously prognosticated in 1998 that the internet’s economic impact wouldn’t surpass the fax machine. Hence, they think it’s all just too speculative and nonobvious right now to warrant an alarmist mentality.

However, the asymmetry between the visibility of endangered jobs and the (in)visibility unrealized jobs is overstated and too casually relied on. We can claw more detail from the future about potential work opportunities than techno-optimists assume, particularly by specifying the economic mechanisms through which AI could unleash new utility from humans that might inspire offers for employment. Surveying the pathways for AI to spin up new labor demand isn’t very heartening, and so counting on this trite stuff about the Luddites is just too glib. Upon scrutinizing the pathways to boost labor demand, AI appears poorly suited to foster enough of an uptick in the usefulness of human labor to manifestly outweigh its inexorable downward pressure on salaries and hiring, especially by comparison with the technological marvels of the past.

More crucially, we can justify panic by simply thinking about what careers are insulated from disruption and working backwards (viz., manual labor—stuff that is accomplished in the physical world rather than online and would require better robotics for AI to effectuate). Coach-builders and farriers were presented with a chance to redeploy their efforts as carmakers. AI is unlikely to present the modern workforce with a similarly straightforward and innocuous exchange. For AI to boost labor demand, the work generated can’t also be doable by other AIs. Everyone whose work product is digital or who completes their job using computers (or who could conceivably do so) is eliminable, economically speaking, in comparison to agentic AI. So, the only tasks that AI could engender that would require human participation will necessarily be offline, and ultimately the vital question isn’t just whether AI will create some new jobs, but whether the side effects from deploying AI will stir up enough manual labor to offset its impending dominion over office work.

Even some techno-optimists are openly contemplating a gloomy destiny for the labor market and endorsing UBI as a remedy for widespread joblessness. Others channel their optimism into views other than the Dark Forest Theory (e.g., that fears about the usefulness of humans spring from a failure to understand the principle of comparative advantage). Some believe that the rising tide of economic progress will necessarily lift all boats. I’ll have to reserve my analysis of this stuff for some future posts, along with further reasoning about why this time is plainly divergent from the historical episodes that are typically cited, but the convenient history of false alarms about tech-based unemployment has inspired too much confidence in an inexhaustible Dark Forest Theory of the labor market that deserves rethinking.

1. Well, nearly everyone: sometimes people muse about how improving technology in an industry results in a greater number of workers in that industry. Of course, this is possible, but it’s hardly destiny.

2. See Frey, Carl Benedikt, The Technology Trap, Princeton University Press (2019) at 311.