Large language models (LLMs) \u2014 today\u2019s premier form of artificial intelligence \u2014 still have their limitations. They can\u2019t reliably fulfill most white-collar workers\u2019 every function. But AI progress is compounding on itself. As LLMs automate the process of building better LLMs, they will kick off a feedback loop of exponential self-improvement<\/a>.<\/p>\n Thus, by the end of next year \u2014 if not this one \u2014 AI will render much of America\u2019s professional class obsolete and push unemployment to 20 percent<\/a>. Within a decade, the technology could wipe out virtually all forms of knowledge work<\/a>.<\/p>\n Or so many of AI\u2019s champions and detractors believe.<\/p>\n In recent weeks, the drumbeat of catastrophic labor-market forecasts has grown louder, with tech CEOs<\/a>, financial analysts<\/a>, and journalists<\/a> penning viral predictions of an impending unemployment crisis.<\/p>\n In my view, the threat of AI-induced unemployment is worth taking seriously. And I\u2019ve sketched out the case<\/a> for alarm in past essays<\/a>.<\/p>\n If the AI doomers\u2019 concerns are warranted, however, their certainty is misplaced. Artificial intelligence could trigger mass white-collar layoffs in the near future. But there are plausible arguments against that scenario.<\/p>\n To inject some balance into the AI discourse \u2014 and\/or, reassure myself that my hard-won verbal skills aren\u2019t about to be less economically valuable than my flimsy biceps \u2014 I\u2019ve sought out reasons for optimism about the white-collar labor market. Here are the four that I found most compelling:<\/p>\n 1) You can see the AI age everywhere except in the jobs data<\/p>\n The first reason to doubt the doomer scenario for AI and unemployment is that it keeps not happening.<\/p>\n Or, more precisely: Despite the astounding capacities of today\u2019s LLMs, there still aren\u2019t many signs of large-scale, AI-induced job loss.<\/p>\n It takes time for firms to adopt new technologies, of course. But generative AI has been remarkably powerful for a while now. As of late 2024, it could already automate many coding tasks, generate research reports, write ad copy, review legal documents, and make terrible music<\/a> at a near-human level<\/a>.<\/p>\n Yet America\u2019s unemployment rate has barely budged over the past two years, hovering near 4 percent.<\/p>\n Even in the industries most suited to AI-driven automation, employment shifts have been modest. Job postings for software developers<\/a> have actually increased over the past year. Employment<\/a> in market research<\/a>, meanwhile, went up after ChatGPT hit the market. Even customer service representatives \u2014 arguably, the workers most threatened by chatbots \u2014 have not suffered massive job losses: Although employment in the field fell 10 percent from 2023 to 2024,<\/a> it has held steady since then and remains close to its pre-pandemic level.<\/p>\n What\u2019s more, there are few indications<\/a> that mass, white-collar layoffs are on the horizon. In a December survey by the accounting firm KPMG<\/a>, 92 percent of CEOs said they were planning to grow their head counts, even as 69 percent were dedicating a large share of their budgets to AI deployment.<\/p>\n Similarly, a January survey from EY-Parthenon<\/a> found that 69 percent of CEOs expected that AI would lead them to either maintain or expand their payrolls.<\/p>\n One could dismiss this as sunny bluster. But there is evidence that these executives\u2019 ostensible intuition \u2014 that AI adoption and downsizing don\u2019t necessarily go together \u2014 holds true in practice. In a study of 12,000 European businesses published in February, firms that adopted AI<\/a> saw a 4 percent increase in labor productivity \u2014 yet did not reduce their staffing in response.<\/p>\n Granted, if you scour the jobs data for portents of an AI-driven unemployment crisis, you can come up with a few. For one, between November 2022 and January 2026, America\u2019s core white-collar industries<\/a> \u2014 finance, insurance, information, and professional and business services \u2014 cut their staffing by 1.9 percent. This is unusual; outside of recessions, these sectors have historically added jobs at a steady rate.<\/p>\n For another, a Stanford Digital Economy Lab study<\/a> suggests that young workers in heavily AI-exposed fields have seen declining job prospects, relative to those in other sectors, since ChatGPT debuted.<\/p>\n Forecasts of an impending white-collar \u201cbloodbath<\/a>\u201d tend to put a lot of weight on these data points. And yet, both developments likely have less to do with AI adoption than with monetary policy.<\/p>\n As two economists at Google recently observed, America\u2019s most AI-exposed industries<\/a> began to slash hiring six months before ChatGPT hit the market in November 2022. And white-collar job postings fell most precipitously in 2023, when corporate deployment of LLMs had barely begun; in the fourth quarter of that year, fewer than 10 percent<\/a> of large businesses said they were even planning to use AI in the next six months.<\/p>\n This timeline is hard to square with the theory that AI drove the slowdown in white-collar hiring. By contrast, the timing neatly aligns with the Federal Reserve\u2019s tightening cycle.<\/p>\n In March 2022, the central bank began hiking interest rates at a historically aggressive pace. A little over one month later, job postings began to fall in white-collar fields. When the Fed paused its hikes in 2024, that decline bottomed out; when the central bank began cutting rates in 2025, job openings started rebounding.<\/p>\n Courtesy of the Economic Innovation Group<\/p>\n Critically, interest rate hikes disproportionately impact AI-exposed industries. The sectors most susceptible to artificial intelligence \u2014 tech, finance, and professional services \u2014 are also among the most sensitive to tightening financial conditions. And when companies come under strain, they often pause entry-level hiring<\/a>.<\/p>\n A pullback in employment caused by the Fed could therefore look a lot like one triggered by LLMs.<\/p>\n None of this is to deny that artificial intelligence has reduced employment in some occupations (for example, AI is almost certainly implicated in the recent decline of computer programming jobs<\/a>). The point is just that the overall labor market impacts have been remarkably modest, given the scale of AI\u2019s current capacities.<\/p>\n 2) White-collar workers don\u2019t need to outperform AI to remain economically valuable<\/p>\n The absence of a one-to-one correlation between increases in AI\u2019s capabilities \u2014 and declines in white-collar employment \u2014 isn\u2019t entirely surprising.<\/p>\n To remain economically valuable, a human worker does not need to outperform a machine at their job\u2019s core tasks; they merely need to usefully complement that machine\u2019s operations.<\/p>\n Consider translators. LLMs can convert text from one language to another at a speed and cost that no human could ever match. For many tasks, if corporations, authors, and publishers were forced to choose between having access to AI \u2014 or the world\u2019s most gifted linguist \u2014 they would choose the bot.<\/p>\n And yet, a human translator working with an LLM still produces better text than the machine does by itself. While the latter blitzes through a first draft, the former can correct excessively literal translations of idiomatic expressions, tailor tone to the intended audience, and catch subtle errors that invite confusion or legal risk.<\/p>\n So long as human translators retain this utility, AI progress won\u2019t necessarily reduce demand for their services. In fact, the technology could conceivably increase such demand.<\/p>\n That claim might seem unintuitive. After all, it surely takes fewer people to translate any given quantity of text in the age of generative AI than it did in years prior.<\/p>\n Yet humanity\u2019s appetite for translated text is not fixed. If you drastically increase the efficiency of translation \u2014 and thus, reduce its cost \u2014 then people will purchase more of it.<\/p>\n And indeed, since the introduction of ChatGPT in 2022, demand for translation has surged<\/a>. Perhaps for this reason, even as machines have come to match or exceed the skills of human translators across several dimensions, employment in the industry has grown in the European Union<\/a> and stayed<\/a> roughly<\/a> level in the US.<\/p>\n And you can tell a similar story about myriad other fields.<\/p>\n AI can read medical images faster \u2014 and, for some types<\/a> of cancer<\/a>, more accurately \u2014 than any human. Still, a radiologist working with an AI yields better diagnoses<\/a> than the machine working alone. And as LLMs have made radiology more efficient, demand for imaging has spiked<\/a> \u2014 and with it, radiology employment<\/a>.<\/p>\n 3) People want some things done by people<\/p>\n In some domains, white-collar workers may retain an advantage over AI simply because they are human.<\/p>\n As the economist Adam Ozimek notes<\/a>, many contemporary occupations could have been automated out of existence long ago, were technology the only concern. We\u2019ve had player pianos and recorded music since the late 19th century. Yet many hotels and bars still pay human beings to tickle the keys for their customers.<\/p>\n \u201cPeople are often willing to pay a premium for the \u2018human touch.\u2019\u201d <\/p>\n For decades, it\u2019s been easy to book your own travel online, relying on aggregators like Expedia and reviews on Yelp. Yet 67,500 Americans still make a living as travel agents<\/a>. Workout videos make it possible for anyone to perform yoga at home, yet many hire personal instructors<\/a>. Mechanical reproductions of famous paintings can be had at a low cost, yet people shell out millions<\/a> for visibly indistinguishable versions that were produced by a specific human hand.<\/p>\n You could have asked ChatGPT to give you four reasons why AI won\u2019t cause mass unemployment, and it would have instantly spit out a listicle. Instead, you\u2019re reading an artisanally crafted explainer that Vox Media Inc. paid me to produce.<\/p>\n In other words, people are often willing to pay a premium for the \u201chuman touch.\u201d<\/p>\n This won\u2019t preempt an AI-induced employment crisis, all by itself. Consumers don\u2019t typically care how their smartphone apps were coded or insurance claims were processed or tax returns were prepared. But a market for explicitly human-produced goods and services is likely to persist in many realms \u2014 including sales, medicine, legal services, and entertainment.<\/p>\n Heck, there might even be durable demand for journalism that\u2019s conspicuously free of AI\u2019s bizarre syntactical tics. That\u2019s not just cope \u2014 it\u2019s a serious possibility.<\/p>\n 4) AI progress won\u2019t necessarily be exponential<\/p>\n All these arguments may count for little, if AI\u2019s capacities are truly growing at an exponential rate.<\/p>\n After all, exponential processes tend to creep up on you. When 32 cases of a supervirus become 64, almost no one notices. If that bug keeps doubling every couple days, however, the world will wake up a month later to 2.1 million infections. In that scenario, a glance at the pathogen\u2019s impact on day three would have told you little about its consequences four weeks later.<\/p>\n In a world where AI progress is exponential, similar principles apply. Look around three years after ChatGPT\u2019s debut and you might see little job loss. But if artificial intelligence is recursively self-improving \u2014 such that every advance accelerates the next \u2014 then today\u2019s AI is only a pale imitation of 2030\u2019s. The former may be to the latter as a hot-air balloon is to a space shuttle.<\/p>\n If so, then examining AI\u2019s impact on jobs over the past four years wouldn\u2019t shed much light on its effects over the next four. Likewise, the fact that white-collar workers can usefully complement AI today would scarcely guarantee their utility in the future.<\/p>\n But it\u2019s not clear that AI has actually been improving at an exponential rate, much less that it will keep doing so, for years on end.<\/p>\n Without question, LLMs\u2019 capabilities have been growing rapidly. But claims that this progress has been exponential tend to rest on a single, widely cited benchmark.<\/p>\n The AI research institute METR has long been the authority on the speed of AI progress. To gauge that pace, it tracks the duration of tasks that LLMs can complete with at least 50 percent accuracy. In this context, duration is measured by how long it would take a skilled human worker to complete the same assignment.<\/p>\n METR\u2019s charts<\/a> of how this has changed over time are ubiquitous in discussions of AI. And the trends are eye-popping.<\/p>\n Faced with these vertiginous slopes, many jump straight to wondering whether they will enjoy life as a \u201cmachine God\u2019s\u201d pet<\/a> \u2014 forgetting to first ask themselves, \u201cWait, how does METR know that?\u201d<\/p>\n Which is unfortunate, since the short answer is it doesn\u2019t.<\/p>\n METR isn\u2019t spying on every white-collar laborer in America, implanting bugs and honeypots in their break rooms, so as to determine how long it takes each worker to perform their jobs\u2019 tasks.<\/p>\n Rather, to generate its estimates, the institute presents human software engineers with a bucket of coding assignments, measures how long they take to complete their tasks, and then sees whether AI models can perform the same feats. Through this process, METR estimates that the latest version of Claude can autonomously perform tasks that would take a skilled worker up to 14.5 hours to execute.<\/p>\n And yet, as NYU\u2019s Nathan Witkin argues, there are massive problems with METR\u2019s methodology<\/a>, defects that severely limit what its findings can actually tell us about AI\u2019s capabilities. To name just a few:<\/p>\n METR\u2019s tasks are unrealistically basic. <\/strong>In METR\u2019s own analysis<\/a>, the bulk of their sample tasks differ from real-world engineering problems in systematic ways. Specifically, the former occur in static environments, require no coordination with other people (or agents), and include few resource constraints. METR also largely excluded tasks in which a single mistake could derail the entire project, so as to \u201creduce the expected cost of collecting human baselines.\u201d<\/p>\n When the institute charted AI\u2019s progress on its \u201cmessiest\u201d tasks \u2014 which is to say, its most realistic ones \u2014 this was the result:<\/p>\n Courtesy of METR<\/p>\n Viewed like this, AI progress does not look terribly exponential.<\/p>\n METR\u2019s human baselines are unreliable. <\/strong>The sample of engineers who established METR\u2019s baseline for human performance was neither large nor representative. Rather, as of 2025, its testing included only 140 people, recruited primarily from METR staffers\u2019 professional networks.<\/p>\n More critically, on the more complex tasks, these recruits were often operating outside of their areas of expertise. In real life, these assignments would typically be handled by specialists, who would surely complete them more rapidly than random engineers with little domain knowledge.<\/p>\n Making matters worse, METR paid its baseliners on a per-hour basis, giving them an incentive to drag out their tasks.<\/p>\n AI could have simply memorized the answers to many of its assigned tasks. <\/strong>About one-third of the tested tasks had publicly available solutions. For these assignments, the models may have just been recalling answers they had encountered on the internet, in which case their success wouldn\u2019t necessarily reflect growth in their general capabilities. (If a high-school student gains access to a calculus test in advance, and memorizes the answer, their performance on that problem wouldn\u2019t tell us much about their general math skills.)<\/p>\n None of this is meant to disparage METR\u2019s intentions, or to suggest that its data has zero utility. The pace of AI progress is not an easy thing to measure. And the organization is making an admirable effort.<\/p>\n Still, the fact that its charts are AI boosters\u2019 (and doomers\u2019) go-to evidence for exponential progress \u2014 despite the extreme limitations of its figures \u2014 calls the existence of that progress into question.<\/p>\n\n
![\"Chart](\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2026\/03\/Screenshot-2026-02-22-at-10.39.18\u202fAM.png\"\/)
<\/a><\/p>\n![\"Chart](\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2026\/03\/Screenshot-2026-02-22-at-1.40.44\u202fPM.png\"\/)
<\/a>![\"Chart](\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2026\/03\/Screenshot-2026-02-22-at-1.39.06\u202fPM.png\"\/)
<\/a><\/p>\n![\"Chart](\"https:\/\/www.europesays.com\/ie\/wp-content\/uploads\/2026\/03\/Screenshot-2026-02-22-at-3.11.50\u202fPM.png\"\/)
<\/a><\/p>\n