The Augmentation-Automation Race

The nature of the American labor market in the years to come is best understood by reasoning backwards from the final role of artificial intelligence: will it augment human workers, or replace them altogether? The character of the interventions that policymakers must choose to prevent the disempowerment of American workers, then, is best understood by imagining how these workers might be displaced.

Consider a time in the future, say late 2029, in which it looks like we’ve gotten the question of AI and the labor market right. Following the increasingly dire warnings of industry and policymakers alike, we spent much of 2027 agonizing over the right way to deal with what sure seemed like an impending wipeout of junior white-collar work.¹ To protect these junior workers, we have heeded the advice of economic populists on both sides: we’ve introduced frictions, slowed down the pace of technical progress and adoption, and kept humans “in the loop.” A victory, it seems, for the American worker.

It’s only then that the real shock hits. The frictions, it turns out, merely delayed the disruption. Instead of AI being integrated into American legacy firms and employers, AI-only firms were built on the outside instead. Vertically integrated services firms emerge—in consulting, banking, or marketing—with only an employee or two and a swarm of AI agents to go with them; international competitors in India or the Philippines that adopted AI much more recklessly than America gain an upper hand. Faced with that competitive pressure, American firms are no longer able to compete, scaling down or going out of business as their escalating payroll costs can’t keep up with the ever more efficient AI token spend of their competitors. And so, after only a few years wrapped in comforting, protective regulations, American workers end up just as displaced and disrupted. But the new status quo leaves them disempowered, with much less political recourse than today and far fewer alternatives inside a now AI-powered economy.

If that future comes to pass, it will be because we’ve failed to realize we’re in a race: between augmenting human workers through the rapid, widespread adoption of AI systems that work together with them in tandem, and automating work through developing systems that replace humans wholesale instead. Which side wins this augmentation-automation race is not something we will simply discover. It’s the consequence of the technical choices we make in designing our AI systems and the policy choices we make in designing our labor markets today. They compound over time: with every decision we make that chooses augmentation over automation, we increase the leverage and standing of workers, and in turn, their ability to press for more favorable terms as technology progresses.

To protect workers from the worst of the disruptions that will come with this technology, to insulate the economy against yet another hollowing out driven by the international advent of new technologies, we must make hard choices in the near term. And to make these choices well, we must make them in service of augmentation winning the race.

Fog of War

The actual effects of AI on the labor market still seem shrouded in deep uncertainty. When it comes to predictions on near-term impacts, we can see two camps emerge. A deep bench of established and renowned labor economists continues to expect “small” effect sizes, at least by the standards of the drastic predictions common in San Francisco.² They predict local disruptions, yes, but expect they’ll be offset by broad-based productivity gains and job growth in other fields. Much as Arvind Narayanan and Sayash Kapoor have pioneered the notion of “AI as a Normal Technology,” many economists might be said to support “AI as Normal Labor Economics”: fundamentally, the AI revolution will unfold very similarly to past technological revolutions, which is to say that it might be temporarily and locally disruptive but ultimately end up in a stable, and perhaps better, global equilibrium.³

Yet in much of Silicon Valley, a different view prevails. The creators of the technology, most prominently Anthropic’s CEO, Dario Amodei, predict widespread displacement and high unemployment. If there is absorption of the displaced, it will be into worse jobs that might not even be around for long.⁴ On that view, AI is markedly abnormal: while past technologies have complemented distinctly human contributions, such as intelligence or creativity, AI is sufficiently broad and general that we should expect it to exceed and supplant these human contributions to the economy, leaving human workers largely obsolete.

Much ink has been spilled on the question of which view is correct. To little avail: in the face of a novel technology, we’re still far away from any actionable conclusions. One of many ways to think about the underlying dynamics is computer scientist Ethan Mollick’s notion of the “jagged frontier”: AI systems are getting better, but it turns out they’re drastically better at some kinds of skills (such as writing computer code) than others (wooing a client).⁵ As it stands, the jagged frontier allows for synergistic cooperation between humans and AI systems: our limited supply of AI tokens is best used where AI is best—the spikes—leaving an economically valuable task profile for humans among the gaps. As long as AI capabilities retain this shape, human participation in the labor market remains comparatively advantageous, and we’re spared the worst of the scenarios: humans would neither be relegated to only relational occupations like health care or artisanal manufacturing, nor would they be displaced outright. And so, every sector with a jagged frontier is a sector that favors human employment and provides refuge for the displaced.

The future shape of this jagged frontier, then, is what underlies the augmentation-automation race: if the frontier remains jagged, then niches along the jagged frontier remain. It would remain economically efficient to employ humans in the tasks they’re good at, and the future of the firm would be AIs and humans working together. Will that remain the case? Perhaps. The argument in favor is the one from comparative advantage: it costs scarce research skill and computational power to run AI systems and make them good at novel tasks, and the most efficient uses will be ones that keep the models doing what they’re good at. The argument against is: sooner or later, the most effective next step in AI development will be to fill in the niches in the jagged frontier where comparatively advantageous human labor sits, either because it will suddenly be very easy to fill these niches as systems grow more capable, or because employers and developers are willing to pay a premium to get rid of burdensome employees and the costs they raise. In that case, we face what economist Erik Brynjolfsson calls the “Turing Trap”: we’d make a labor market of machines just like us, which means an economy of machines with little use for us.⁶

Two Paths, Not One Truth

Instead of litigating the economics of the case and declaring one or the other theory correct, I believe policymakers should draw a very different conclusion: they should assume both hypotheses are technically sound. Then, they should reconcile them by understanding them as possible futures, contingent on policy choices. If we make augmentation along the jagged frontier synergistic and efficient, comparative advantage holds, and the years ahead look as the first view suggests. If integrating human and AI work is burdensome, friction-rich, and hard, firms will grow frustrated and eventually find it cheaper to bypass humans entirely, and the second view is the best prediction of our medium-term future.

Parts of Europe have done democracies the world over the favor of running this experiment before the advent of AI. Employment protections there make firing a worker many times more expensive than in the United States, and a long line of work in the creative destruction tradition, most prominently by economist Philippe Aghion, finds that these protections dampen innovation, firm entry, and reallocation across sectors.⁷ The mechanism, in short, is first and most simply that firms hire fewer workers at the margin when firing is costly. But beyond that, rigid labor rules lock in the status quo structure of the corporation itself, making exactly the kind of radical reorganization that general-purpose technologies demand almost prohibitively expensive. As a result, analysts have remarked that European firms seemingly refine established industries very well but reorganize around new technical realities very badly. In a slow-moving labor market, that approach looks enviable: workers are taken care of, welfare costs are high but tenable because the unemployed reenter work fairly quickly. As long as things remain static—large employers stay large, major sectors remain major sectors—the system is stable.

Once the pace of technological progress and economic transformation picks up, however, the dynamic radically changes. Because job security rides on the sustainability of major employers and the economic usefulness of specific, long-trained task profiles, the protection crumbles once either falls away. Technological progress has a way of causing that: rapidly growing new firms become a larger part of the labor market, and task profiles are made redundant by the advent of technology. That means that high employment can only be maintained through new firm entry, quick reallocation of workers to new task profiles, some firms being able to hire quickly and speculatively, and others being able to stay in business by firing half the workforce instead of going under and ultimately losing all of it. In the long run, however protective policy may be, new firms will enter the market. But if they face an open question of whether to hire humans or deploy AI systems, they will be more likely to deploy AI systems if hiring humans is costly, bureaucratic, and competitive in a highly restricted labor market. In that case, the fastest growing parts of the economy might or might not still show up in headline GDP, but they will not show up in the labor market. That is the lesson from the European lag: if the labor market does not leave room for rapidly rising firms to draw in labor, the labor market loses out on the beneficial employment effects of growth and innovation.⁸

The trends in AI technology exacerbate this effect because AI is itself a fast-moving technology that also accelerates technological progress elsewhere. Already today, AI is among the fastest-adopted technologies in history, and it has found its way into most firms and shaken up the way workers do their work. If policymakers respond to the short-term threat from AI by piling on friction, they will slow down their labor markets in exactly the ways that have already overexposed Europe to technical disruption in recent years.

We can derive a policy design principle from the tension between the protectionist instinct and the threat of ultimate disruption: we must build the conditions for human-AI teams to stay on the productivity frontier as more autonomous capability rises. These conditions could be disruptive in the short term in their own right: adopting aggressively creates efficiency gains that can and will cause local layoffs and restructurings. But still, if you take the implications of the race seriously, the only way out of the AI job displacement threat is through intermediary disruption.

This policy principle will make pro-worker voices recoil at first glance, and for good reason: the vague promise of long-term resilience seems like an insufficient reason to allow short-term disruption, especially when the following recommendations align with a tech policy agenda that frequently does not seem to have workers’ welfare at heart. But the underlying reason for this “strange bedfellows” dynamic is that, executed well, augmentation is not simply a waystation on a speedy path toward full automation that many pro-labor voices fear the AI developers pursue. If pro-worker voices succeed in retaining augmentation as the economically optimal paradigm, they do not need to fight the adversarial fight against AI as a technology right now, do not need to go up against a stock-market-critical data center buildout or hundreds of millions in political donations. A world where American workers are supercharged by American AI systems because we’ve simultaneously encouraged uptake and discouraged wholesale replacement still comes with a dynamic market rich in demand for tokens and bespoke AI systems, much unlike the friction-rich world that alternative pro-worker advice might counsel. From that position, the eventual policy space to grapple with the prospect of obsolescence through automation will be much wider.

But creating this mutually beneficial market is downstream of our choices today because the technology’s trajectory is path-dependent: a workforce, and a political class, that clearly demands augmentation-forward technology shapes what developers design and deploy. Demand pulls the technology toward the shape that keeps augmentation viable. And an economy that has bet on augmentation, through industrial policy or otherwise, acquires its own incentive to defend that balance. It can do so through its own investments, and above all through its scarcest assets: the proprietary data and process knowledge that automation-seeking developers cannot generate on their own. The firms that hold these assets can deploy them to shape the technology in augmentation’s favor—or sit on them until selling out to automation becomes too attractive. An affirmative bet on augmentation as the design principle of the American AI labor market creates the preconditions for augmentation to remain the dominant paradigm for far longer.

Extending the Centaur Era

Beyond all empirical substance, the pressing constraint is political. Whatever policymakers conclude about the augmentation-automation race, the political economy of AI labor impacts pulls them in one clear direction: toward heavy-handed responses and crackdowns on AI use, even and especially where that use is augmentative.

The underlying issue is this: long-run trends will continue to pull toward automation. As AI systems become more efficient and effective, as data becomes less relevant and computational power becomes more abundant, it will be easier to build machines that can do anything humans can do. A rapid advent of the automation that trend would bring is unfavorable to workers: right now, there’s little political salience and interest in dealing with the heavy automation futures some AI executives predict; workers aren’t organized and voters aren’t yet engaged, and the policy solutions they’d want to call for still do not exist. Perhaps we can avoid that future altogether, if policy continues to be shaped according to the interests of workers who seek to retain an economic role.

But right now, most AI use is effectively augmentative: we’re in a “Centaur era” of human-AI collaboration along mixed task profiles that make up complex jobs. The data, so far, shows very little direct job displacement.⁹ Instead, it shows the content of task profiles shifting: software engineers spending less time writing computer code by hand, lawyers spending less time researching and more time interfacing with clients and courts.¹⁰ Very rarely are wholesale AI agents deployed to outright fill the roles of workers as they exist right now; the few high-profile instances of AI-driven layoffs have often been linked to longstanding layoff plans unrelated to technological progress. Right now, when firms use AI—when AI becomes more abundant, easier to use, more present in the workplace—our economy is running a million small experiments on how to make augmentation work. Any hastening now adds to this experimentation more than to the process of full-on automation, with the latter much more bottlenecked by progress in model capabilities governed by progress in San Francisco than by deployment in the rest of the country. The systems we have today are of a shape that integrates into the jagged frontier, not attempts to smooth it over. The better we get at exploring the confines and advantages of this augmented paradigm, the more competitive this paradigm might be once Silicon Valley manages to build integrated systems that can compete at parity with today’s workers.

But even if the sweeping debate over a new social contract needs to be held eventually, it shouldn’t be held today: workers would likely lose it. Instead, right now, we should seek to retain a broad-based distribution of economic power, income, and jobs—because with them comes democratic leverage. Without an economic contribution to make, citizens lose their immediate value to, and recourse against, both their employers and their governments.¹¹ The more essential they remain to the economy, the more likely it is they enter negotiations from a strong position; the easier it is to replace unpleasant workers with obedient machines, the harder it will be to sue for an equitable arrangement. To improve on the long-term policy treatment of humans’ potential economic obsolescence, we should favor augmentation over automation in the short term and extend the Centaur era.

Yet the politics pull toward adding frictions even in this current phase of fundamentally worker-friendly adoption trends. Two asymmetries in particular produce that pull.

The first is an asymmetry of visibility. The costs of immediate AI adoption are often localized, concentrated, and legible. A round of layoffs at one major employer is attributed to AI by the CEO, reported by major media, and boosted by every technoskeptical politician.¹² The benefits are comparatively diffuse and boring: adoption yields a few basis points of productivity growth, a marginal edge in international competitiveness and invisible national security gains. This general dynamic, of mobilizing concentrated costs versus diffuse benefits, is well known, but it’s worse in the case of AI, which will be blamed for many future localized disruptions entirely independent of whether they actually have anything to do with AI. For the next decade, every corporate restructuring, every hiring freeze, every soft graduate labor market could plausibly be read as an AI story; some of these readings will be right, many won’t be, and the political effect will be the same either way.¹³

The second is an asymmetry of traction. The moderate economist’s view counsels patience, measurement, and modest intervention. But that position is mostly rhetorically inert: you don’t win primaries on claiming the evidence remains inconclusive, and there is no good debate stage defense against the charge of selling out the workers to the coastal tech feudalists. Very generally, the pessimistic view supplies some of the most explosive political material: it trades in terms of a “permanent underclass” or the “end of work,” creating a generation locked not only out of the professions their parents took for granted but out of the main vehicle for climbing the ladder and fulfilling the American Dream. Politicians on both flanks can and will build platforms and careers on that rhetoric. What’s the skeptic’s incentive to push back on this instead of dodging the issue and fighting another fight, especially when the public’s own priors run in the same direction? On this, polling has been remarkably stable: people are deeply worried about automation, for themselves and for the country.¹⁴ In response, they seek measures to protect their ability to work for a living.¹⁵ An electorate in that mood, supplied by policymakers of this disposition, will reach for easy interventionist wins.

The result is a political market that gravitates toward friction. All easy policy solutions that reduce automation in the here and now ultimately cut against this early-stage augmentation required to win the race: while wholesale agents that can replace entire firms do not exist yet, plenty of current AI adoption is augmentative. Whether it’s restrictions on AI deployment, mandated humans in the loop, procedural layers around automation: all this satisfies demand for visible protection while, in effect, blocking immediate augmentation.

Losses from Friction

The policies most suited to meet this political moment are the least suited to winning the augmentation-automation race. Nearly every intervention you could dream up today to myopically reduce AI disruption of the workforce hamstrings near-term adoption. Adding burdensome requirements—such as last decade’s well-trodden idea of keeping humans in the loop,¹⁶ or the now popular “token taxes,”¹⁷ or slowing down data center buildouts, or restricting training data—makes the American workforce less AI-powered, ultimately leaving it more vulnerable to disruptions from outside. On the rare occasion where inducing friction has worked to retain jobs at some economic cost, such as in the case of the longshoremen’s union preventing port automation,¹⁸ this has worked because the relevant industries had been geographically captive. If ports are not automated, there’s still no better way to ship goods to America. The same is decidedly untrue for the jobs AI might prove most disruptive to: services jobs that can be carried out from anywhere over the internet and that are therefore offshorable to whoever uses AI best.¹⁹

The disruption that would result from a near-term, friction-inducing agenda comes through two channels. The first is domestic. Most protective measures protect jobs within legacy business models: they make it costly for existing firms to slot AI agents into human seats. But those frictions mean very little to new entrants that never seek to employ as many humans in the first place. Vertically integrated firms—the two-person consultancy with a swarm of agents, AI-only law firms—start outside the economic infrastructure that frictions are supposed to defend. Consider the analogy to neobanks over the last decade: as opposed to staff-intensive banks with tellers, branches and sprawling bureaucracies, financial technology start-ups took every shortcut available and now provide effective digital banking solutions at a fraction of the payroll cost. We might see the same compressing effects from new, AI-native firms entering established sectors in most domains of white-collar work. Reasonable people might disagree as to whether such firms are a desirable element of a new economy, but they should at least agree that such firms must compete on fair market terms. If they simply win because incumbents were shackled by rules asymmetrically hitting AI adoption in places with workers, that is policy failure: we would have handed entire industries to whoever moved fast on the outside instead of allowing the incumbents to, yes, cut down on workforce somewhat, but defend the ultimate concept of a firm with employees.

The international version of this effect is well-known: through the China shock that eroded manufacturing capacity and hollowed out the economic infrastructure, we have come to know the risks of exposing a slower-moving workforce to a competitor with a decisive cost advantage.²⁰ The deficit in industrial competitiveness in the face of Chinese trade practices was the underlying driver of this shock. No matter how good the post-shock social policy could have been, it could never have made up for all of what was lost: even the most generous wage insurance would have been unable to restore the value of a profession and a career; even the most expansive subsidy would have been unable to retain manufacturing at scale. A comprehensive solution would have needed to tackle the emerging competitiveness gap head on.

Today, the only visible way to do that is to hasten adoption, reduce frictions, ensure that American workers have a fighting chance through using the best AI systems in the world. Else, if workforces in India or the Philippines embrace maximal AI adoption—and their governments have every incentive to let them—their firms will undercut American ones across every internationally contestable service domain: augmented workers first, wholesale automation later. A second white-collar China shock would, in that sense, be partly self-inflicted, not as the result of a rival’s cheaper labor but as a result of rules of our own design that kept American workers from wielding the best tools available to them. The default greatly favors America in this context: it is where the best AI models are built and where most of the infrastructure needed to deploy them stands. Preferential access to this technology could be a direly needed competitive advantage for American workers as competition for services jobs becomes global.

Winning the Race

What should policymakers make of all this? The policy imperative is almost straightforward. If the real threat is decisive disruption down the line, the near-term agenda has to enable near-term disruption. The paradigmatic counterexample to this line of reasoning is perhaps the token tax, based on the popular notion that there should be usage-based taxes on AI, in effect placing a greater burden on those who adopt more rapidly or grow more quickly due to AI use. The problem with the token tax is not taxation as such but its symmetry: it burdens the firm that augments its workers and the firm that replaces them alike. Instead, policymakers should ensure that taxes favor widespread usage of AI in the workplace, where that usage does not trade off directly against employment—ensuring American workers are the best-augmented in the world. They should also enable data use so that accumulated American expertise actually flows into the systems American workers will rely on. Each of these will cause more local pain than doing nothing, or than maximizing frictions, but each of them puts forward a serious attempt to prevail in the race that matters.

Much necessary policy comes down to symmetry: nothing in the tax code or the regulatory structure should make it cheaper to replace a worker than to augment one. Today, the scales frequently tip entirely the wrong way: payroll taxes mean high overheads while capital expenditure on AI is incentivized.²¹ And it would be a bitter irony if humans turned out to be competitive with AI systems on the merits, but arcane tax incentives from a pre-AI era put them out of business anyway. Shifting part of the burden from payroll taxes onto AI-specific capital investment is the cleanest correction we could imagine: its effect would be roughly neutral for the firm that employs many workers and boosts them with AI, while being costly only for the decision to bypass humans wholesale. Unlike token taxes, this instrument does not punish adoption but outright replacement.

The same logic extends to data. Liberalizing data use is a cheap ask that usually primarily comes as a gift to the developers. An augmentation-forward version is narrower. The scarcest input for augmentation-shaped AI is not the open web but the data and process knowledge locked inside American firms: everything from physical worker movements to claims-department processes. Policy should make it easy and rewarding to turn that asset into augmentation: there should be clear rules that let firms train systems on their own operational data without legal ambiguity and data consortia through which midsized firms can pool process knowledge to develop bespoke augmentation tools, rather than each selling out to an automation-seeking buyer as a last resort. In every marketplace, there will be firms that skew toward a human-worker-heavy, augmentation-forward approach, and firms that will attempt to find shortcuts through heavier automation instead; whether that’s small agile start-ups, Big Tech firms attempting to vertically integrate, or legacy firms dead set on replacing their workforce. On the margins, data policy should favor the firms making an honest attempt at augmentation.

On the flip side, policymakers might consider intervening into attempts to collectivize ownership of worker data outside of their control, understanding that however well-compensated this transfer of ownership might be, it risks surrendering the one lasting asset workers have to enable augmentation that outperforms automation.²² Done this way, data policy is not a favor to the labs but a way to keep the race’s key dual-use asset on the side of workers.

All of this may sound to many readers like a defense of an accelerationism that will definitely benefit the makers of AI technology—and perhaps help workers later on. We should not embark on this attempt on faith alone; at every step of the way, we must put measures in place to track whether it’s working. We must verify if our attempts to encourage augmentation are making American workers, not just American firms, pull ahead of the new entrants and international competitors that threaten them. It’s become an almost tired cliché, but no less true, to note that doing any of this well requires greater state capacity. The federal government currently sits at the wrong end of a stark information asymmetry: frontier labs watch usage patterns across sectors and token spend across domains in real time, while the Bureau of Labor Statistics is stuck on decades-old survey-based approaches.²³ The developers, on their own, are not trustworthy narrators here—entirely too much of their revenue depends on telling a very specific story of white-collar uptake of their agentic systems. Policymakers should encourage and mandate these developers to share their economic usage data with statistical agencies: an augmentation strategy cannot be run, and its casualties can neither be identified nor compensated, without sufficient data.

The political challenge contained within that strategy is much more difficult to resolve. The agenda above asks elected officials to permit and even welcome visible, attributable, near-term job losses in order to avert a diffuse and deferred catastrophe. That is a very hard political sell. Just making it remotely pitchable to a principal requires a clear message and some way to deal with the immediate economic fallout before it creates constituencies.

The message has to put an explanation of this augmentation-automation race front and center. American workers are well-positioned to beat out the rest of the world by harnessing the world’s best AI: protecting them from this tool is to hamstring them. Protection in the near term, by contrast, means losing slowly to competitors less squeamish about the transition. Told properly, that can be framed as a story of foresight and strength rather than surrender. But as always, it must be told and developed and tested early. The more counterintuitive the political argument, the more lead time it requires to become popular with the electorate. By the next presidential primaries, the easy answers will be in position: scrambling to oppose them on the augmentation framing will be too late, so policymakers interested in protecting workers in the long term need to build their agenda now.

All of this requires collateral, with the stakes being the losers of the transition, chiefly workers displaced by the rapid adoption promoted by the agenda. They must be compensated not just well or acceptably but expansively. They should be provided broad wage insurance, workforce development funded well past the point of comfort, and outright transfers where the reach of other instruments is in doubt. Policymakers must spare no expense here, and they will be able to be generous because if the play works, the total number of those compensated will be much lower than in the automation-forward future. If we fail to get the compensation right, those affected will harden into an anti-augmentation constituency—and that constituency will demand, and in a democracy will get, the frictions that lose the augmentation-automation race. Paying now is a political imperative: boosterism alone is a wholly insufficient response to understanding the circumstances of this race.

None of this is a politically comfortable platform. The agenda promises early turbulence and greatly delayed vindication only later, betting on the strength of a mechanism most voters have never heard of. But the alternative is the late-2029 scenario sketched at the start of this essay, not through malice or stupidity, but through a series of reasonable protective choices undertaken by well-meaning pro-worker politicians; the shock hits all at once, with little room to respond. The augmentation-automation race is still winnable. But winning it requires admitting its existence and explaining its importance.

This article originally appeared in American Affairs Volume X, Number 3 (Fall 2026): 150–63.

---

Notes