The future of AI at work is a hiring story, not a firing story

Ask about the future of AI and work, and you’ll get two confident answers. One says the jobs are going. The other says nothing has happened, because unemployment is fine. Both are arguing from numbers, and the numbers they pick decide the answer before the argument starts. Here’s the claim this piece defends: the measurable effect of AI on employment so far sits at the point of hiring, not the point of firing, and the headline series most forecasts lean on can’t see it either way.

That’s a falsifiable position, so the rest of this sets out the evidence for it, the strongest case against it, and the specific readings that would break it. Everything below comes from a source we opened: administrative payroll records, a Federal Reserve note reconciling three adoption surveys, a staggered software rollout across thousands of support agents, a randomised developer trial, and one lab’s own survey of its users.

The payroll data points at the door, not the desk

The most granular evidence in this piece comes from ADP payroll records, which cover millions of US workers at high frequency. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen used them in a Stanford Digital Economy Lab working paper published in August 2026, running through June 2026. Its first finding is the one both camps skip: no evidence of widespread, economy-wide job displacement.

Its second finding is the one that matters. Employment of workers aged 22 to 25 in AI-exposed occupations now sits 19% below where it would be if it had kept pace with their less-exposed peers. Experienced workers show no comparable gap. In levels, employment of 22 to 25 year olds in the two most exposed quintiles fell about 11% between November 2022 and June 2026, while the same age group in the three least-exposed quintiles grew about 10%.

That gap has widened rather than closed. The paper reports the same kept-pace measure at 15% using July 2025 data, against 19% as of June 2026. The mechanism is the part that gets misreported: the adjustment runs through reduced hiring of young workers, not through increased separations. Nobody’s being marched out. The entry-level door is opening less often, which is a slower and much quieter process than a layoff round.

A layoff has a date, a headcount and a press release. A job that never gets posted has none of those, which is why this showed up in payroll records before it showed up in the news.

Rundowns AI

The authors are unusually careful about what this is, and quoting them properly matters more than quoting the 19%. They call these early, descriptive indicators rather than causal estimates, and they say the data alone can’t establish how much of the divergence generative AI caused. They also list what weakens the reading: the estimates attenuate when you control for occupational education levels, some divergence between exposed and less-exposed occupations predates ChatGPT, and the pattern is stronger in the ADP panel than in national survey benchmarks. We covered the earlier vintage of this finding when the entry-level series first split from the aggregate, and the direction hasn’t reversed since.

Why the aggregate series can’t settle this

Which brings up the obvious objection. If something this large is happening, why doesn’t it appear in national statistics? It mostly doesn’t, and that’s consistent rather than contradictory.

The Budget Lab at Yale has been tracking exactly that question, in an analysis first published on 16 July 2026 and updated on 19 August 2026. Its conclusions are blunt. The occupational mix “is not yet changing in ways that clearly align with the introduction of AI into the workforce”. Measures of AI usage “show no connection to changes in employment or unemployment”. A synthetic differences-in-differences analysis of AI exposure “does not yet clearly indicate an AI-related labor market footprint”.

Read those two studies side by side and they aren’t in conflict, because they’re measuring different things. Workers aged 22 to 25 are a small slice of total employment, and hiring flows are a small slice of the stock of jobs. A shortfall concentrated in new entrants to exposed occupations can be real, growing, and still too small to move an occupational mix index built on the whole workforce. That’s what it means for a series to lack the resolution for a question.

So the honest summary is that the aggregate data is evidence of absence only at the aggregate scale. It rules out the strongest displacement claims. It can’t rule out the narrow one, and treating it as though it can is the most common mistake in this argument.

Adoption depends entirely on which number you quote

The same resolution problem wrecks the adoption debate, and here it’s been documented directly. Jeffrey S. Allen of the Federal Reserve Board published a FEDS Note on 3 April 2026 comparing three surveys that all claim to measure US AI adoption and disagree by a factor of four.

Survey	Reference period	What it counts	Result
Census BTOS	Year-end 2025	Firms using AI, firm-weighted	18%
RPS	November 2025	Individuals using generative AI for work	41%
SBU	November 2025	Workforce at firms that have adopted AI	78%
SBU	November 2025	Workforce at firms adopting LLMs specifically	54%

None of those numbers is wrong. They answer different questions. The employment-weighted figure is high because the largest employers adopt first and carry the most staff, while the firm-weighted figure is low because roughly 95% of US firms have fewer than 50 employees. Allen’s own framing is the useful one: which estimate researchers and policymakers look to “depends on the question they are seeking to answer”.

The Census Bureau’s own write-up of the Business Trends and Outlook Survey, published 26 May 2026 from collection running 14 December 2025 to 3 May 2026, puts national firm-level use at 19.8%. The split by size is stark: about 37% among firms with at least 250 employees, 32% at 100 to 249, and under 20% at firms with four or fewer. By sector, Information reached 39.7% and Finance and Insurance 33.9%, against 14% in Retail Trade. Census states plainly that use rose among firms with at least 20 employees but “didn’t change significantly among firms with fewer than 20 employees”.

That matters because it tells you where to look for labour effects and where not to bother. Concentrated adoption produces concentrated effects, which is the same shape the payroll data found. If you want the workflow-level version of this, we’ve written about which workflows actually survive automation once a firm gets past the pilot.

At the task level, the gains land on the inexperienced

Underneath the labour market sits the question of whether the tools work, and there the field evidence is better than most people assume. Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered rollout of a generative AI assistant across 5,179 customer support agents. Access raised issues resolved per hour by 14% on average. Novice and low-skilled workers improved by about 34%, while the most experienced and highest-skilled saw minimal impact.

That distribution is doing a lot of work in this argument. A tool that compresses the gap between a new hire and a veteran reduces the premium an employer pays for experience, and it also reduces the reason to hire the new person at all if the veteran plus the tool covers the volume. Both readings are consistent with the same experiment, which is why the payroll data is needed to break the tie.

The counterweight comes from METR, which ran a randomised trial on experienced open-source developers working in repositories they already knew. It found that AI caused tasks to take 19% longer, with a confidence interval running from +2% to +39%. METR has since revised how it runs the experiment, launching a second study in August 2025 with 57 developers across 143 repositories and more than 800 tasks. On whether the old result still holds, the organisation says it believes it’s “likely that developers are more sped up from AI tools now” in early 2026 than its early-2025 estimate suggested, while warning that selection effects undermine its ability to measure the current impact reliably.

So the task-level picture is genuinely mixed, and anyone telling you otherwise is quoting one study. Large gains for novices on structured work, contested gains for experts on complex work, and a measurement problem that the researchers themselves flag. We looked at the wider version of that evidence problem when weighing what the trials show against what executives report.

The strongest version of the displacement case

Here’s the argument against everything above, in its best form. Every dataset in this piece is backward-looking, and capability isn’t holding still. METR’s time-horizon measurements found that the length of task a frontier model can complete has roughly doubled every seven months across six years of models. Its own conditional statement is that if the trend from the past six years continues for another two to four years, “generalist autonomous agents will be capable of performing a wide range of week-long tasks”.

The capability benchmarks point the same way. OpenAI’s GDPval paper, submitted in October 2025, built tasks from the representative work of industry professionals averaging 14 years of experience, covering 44 occupations across the nine sectors contributing most to US GDP. Its conclusion is that frontier performance is improving roughly linearly over time and that the best models are “approaching industry experts in deliverable quality”.

Then there’s what the heaviest users expect. Anthropic’s Economic Index report published in June 2026 linked survey responses to actual usage for about 9,700 people, over conversations sampled between 10 April and 10 June 2026. More than 35% predicted AI would be able to do most of their work within a year. One in ten rated losing their own job as likely or very likely, and 38% of those attributed the forecast to AI. Over a third put the odds of a junior colleague losing their job in the next year above 60%, which lines up uncomfortably well with what the ADP records already show.

The people who delegate to Claude the most are the most optimistic about their future labor market outcomes, and feel their skills are growing in value.

Anthropic Economic Index report: Cadences, June 2026

That last finding cuts against the fear, though the sample can’t carry much weight. Anthropic says so itself: Computer and Mathematical occupations make up roughly 30% of respondents against 4% of US employment, and management 23% against a 7% share. It’s a survey of people who already chose to use the product, which is close to the least representative group available.

Even so, the steelman holds up. Adoption is concentrated in large firms and climbing, capability trends are steep, and the people closest to the tools expect their jobs to change substantially inside a year. If you believe those three things, “no aggregate effect yet” reads as a timestamp rather than a verdict.

The official forecast still assumes ordinary decades

Set against all of that is the projection the US government actually publishes. The Bureau of Labor Statistics released its 2024 to 2034 employment projections on 28 August 2025. It expects the economy to add 5.2 million jobs, reaching 175.2 million, a rise of 3.1%. That’s a marked slowdown from the 13.0% growth over 2014 to 2024.

Where AI appears in that release, it mostly adds jobs. Demand for AI-based systems is cited as a driver behind professional, scientific and technical services at +7.5% and the information sector at +6.5%. Healthcare and social assistance leads everything at +8.4%, driven by an ageing population and chronic conditions rather than by technology. Retail trade is projected to lose the most jobs of any sector at -1.2%, attributed to automation, consolidation and e-commerce, which is the pre-generative story continuing. AI does turn up on the losing side too: BLS expects automated systems including AI to contribute to declining employment among office and administrative support workers.

What makes that projection worth reading is the technical note attached to it, because BLS says the quiet part itself. Its method assumes technological progress in line with historical experience. In its own words, “in a future state where technology advances much more rapidly than it has historically, it is unlikely that historical relationships would hold, and therefore BLS projection methods are unlikely to yield reasonable results”. So the official baseline isn’t a rebuttal of the displacement case. It’s a scenario that has stated in advance which assumption breaks it.

Source	What it can see	What it can’t
ADP payroll panel, through June 2026	Monthly hiring by age and occupational exposure	Causation, and firms outside the ADP client base
Yale Budget Lab, updated August 2026	Whether the national occupational mix has bent	Effects confined to one age band or to hiring flows
Census BTOS, to May 2026	Share of firms using AI, by size and sector	How intensively, or to what effect
Field studies (NBER, METR)	Task-level speed and quality inside one firm or one repository set	Whether firms convert that into fewer roles
BLS projections, 2024 to 2034	The official ten-year baseline	Anything faster than its own revision cycle

The point of that table isn’t that any one row is authoritative. It’s that the disagreement between them is structural, not a matter of one side being sloppy. Each series was designed for a question, and the future of AI at work happens to fall between the questions.

What would change our minds

Four readings would break the claim, and all four are already being published on a schedule, which is the useful part.

The first is separations. If the Stanford team’s monthly updates start showing the gap driven by rising separations in exposed occupations rather than by reduced hiring, then the door metaphor fails and this is a firing story after all. The second is age. An equivalent gap opening for experienced workers would mean the effect isn’t about entry-level tasks, and the mechanism argued here would be wrong.

The third runs the other way. If the young-worker gap stops widening while firm adoption keeps climbing past 20%, the case that AI is driving the divergence gets much weaker, and the interest-rate and education explanations the paper wrestles with get stronger. The fourth is the aggregate one: if the Budget Lab’s occupational mix analysis starts finding a footprint, the narrow reading here is superseded by a broad one, and it should be.

Until one of those lands, the defensible position is narrower than either camp wants. Something real is happening to the first rung, it’s happening through hiring rather than dismissal, it’s concentrated in large firms in a handful of sectors, and nobody has established that AI caused it. Anyone arguing the future of artificial intelligence from the unemployment rate is arguing from a number that would look identical either way. If you’re deciding what to learn on the strength of this, the job postings are a better guide than the forecasts, because they’re the series that moves first.