Fewer junior hires, not mass layoffs: what AI job data shows so far

Two sets of numbers published within a month of each other describe the same labour market in very different ways. On 12 August 2026, researchers at the Stanford Digital Economy Lab reported that employment of 22- to 25-year-olds in the occupations most exposed to AI stands about 19% below where it would be had it kept pace with same-age workers in less-exposed jobs. The update, by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, uses ADP payroll records through June 2026. It finds no comparable gap for experienced workers and no sign of widespread, economy-wide displacement.

On 3 September, the outplacement firm Challenger, Gray & Christmas reported that US-based employers announced 52,881 job cuts in August. Employers cited AI for 3,462 of them, which put it fourth among stated reasons and ended a run of five months, starting in March, during which AI had topped the list of reasons every month. Over the year to date AI remains the leading reason, cited in 116,175 announced cuts, or roughly 22% of the total.

Neither release shows that AI is causing job losses. The Stanford authors describe their results as descriptive patterns rather than causal estimates, and Challenger's categories record the reasons employers choose to give. Read together, though, they suggest that where an AI-related shift is visible at all, it is appearing first as fewer entry-level hires rather than as mass layoffs. For organisations that automate junior work, that is where the longer-term cost could sit.

How the two measures work

The Stanford work draws on anonymised payroll data from ADP, which provides payroll services to firms employing more than 26 million US workers. The analysis uses a subset: a balanced panel of firms observed every month since January 2021, covering between 3.5 and 5 million full-time employees a month. Each worker's job title is mapped to an occupation, and each occupation is scored for exposure to AI.

Two exposure measures do most of the work. One rates how susceptible an occupation's tasks are to automation by large language models, drawing on a 2024 study by Eloundou and colleagues. The other is the Anthropic Economic Index, which maps queries to Anthropic's Claude model onto work tasks and classifies usage as either automating a task or complementing the worker. Occupations are sorted into five exposure bands, and employment is tracked by age group from November 2022.

The headline figure rests on a simple comparison. Employment of 22- to 25-year-olds in the two most-exposed bands fell by about 11% between November 2022 and June 2026, while the same age group in the three least-exposed bands grew by about 10%. The 19% figure expresses how far the first group now sits below where it would be had it grown in step with the second.

Challenger's figures are different in kind. The firm counts job cuts announced by US-based employers and files each under the reason the announcement gives. Reductions announced as a range or spread over several years are recorded at the announced figure. AI has been tracked as a distinct reason since 2023, and a separate category captures cuts tied to new technology where AI is alluded to but not named.

What the Stanford update found

Beyond the headline, three findings matter for employers. The first is that the gap has grown: on the same descriptive measure it was 15% at the July 2025 data vintage. Earlier versions of the study led with a regression estimate that adjusted for firm-level shocks, 13% on July 2025 data and 16% on September 2025 data, so comparisons with those numbers mix two different measures.

The second is that the adjustment runs mainly through hiring. The authors find that the gap reflects fewer young workers being taken on rather than more of them leaving, and that it shows up in employment rather than in base pay. That matters because slower hiring does not show up in layoff trackers.

The third is that the pattern depends on how AI is used. Declines are concentrated in occupations where usage mainly automates tasks. Where usage mainly complements workers, employment is flat or rising, particularly among experienced staff. The authors also report suggestive evidence that young workers are losing ground in occupations built on codified knowledge, while occupations that rely on tacit, experience-based knowledge show faster employment growth for experienced workers.

The team offers a possible explanation rather than a conclusion: AI may be taking over the routine, verifiable tasks that once justified entry-level headcount, while giving experienced staff more leverage. The divergence persists when technology firms and computer occupations are excluded, when exposure to interest-rate rises and remote work is controlled for, and when alternative exposure measures are used.

What the layoff announcements show

August's 52,881 announced cuts were up 58% on July's 33,429 but down 38% on August 2025, and the lowest August total since 2022. Through August, employers have announced 529,914 cuts, 41% fewer than in the first eight months of 2025.

Restructuring led August's stated reasons with 16,173 cuts, followed by market and economic conditions with 15,260 and closings with 6,743. AI's 3,462 was its lowest monthly figure since December 2025. A month earlier the picture looked very different: in July AI accounted for 10,970 announced cuts, a third of that month's total, and Challenger had counted 184,538 AI-attributed cuts since it began tracking the reason in 2023.

Technology companies remain the heaviest cutters. The sector has announced 155,126 cuts so far in 2026, up 52% on the same period last year and 29% of all announced cuts, although its August figure of 6,103 was its lowest monthly total of the year. Announced hiring plans point the other way: employers have announced plans to hire 119,825 workers so far this year, up 37%, with technology leading all industries at 19,751.

On 6 August Andy Challenger, the firm's chief revenue officer, said AI is 'shifting the labor market, it is not dismantling it.' In the September release he questioned how quickly announced roles are actually being filled, and whether employers will find people with the skills they need. One month's fall in AI-attributed cuts is too thin a signal to call a turning point.

What is disputed: attribution, causation and timing

The AI category in Challenger's data measures what companies say, not what drives their decisions. Andy Challenger has himself observed that naming AI in a layoff announcement can please investors while alienating current and prospective staff, and that companies have grown far less shy about doing so. He also expects firms to become more careful in their announcements as regulation takes shape, which would make the category harder to read. A fall in AI-attributed cuts could reflect changing language as much as changing practice.

Economists at the Stanford Institute for Economic Policy Research doubt that AI is having a large effect on overall employment. In a July 2026 policy brief, Neale Mahoney, Erika McEntarfer and Karsen Wahal judge that effect likely to be small. They note that unemployment among the most AI-exposed workers has risen by 0.77 percentage points since 2022, slightly less than the 0.85-point rise among the least exposed, a pattern they read as general weakening in the labour market rather than job losses driven by AI. They add that some layoffs linked to AI appear to be about freeing cash for AI investment or unwinding pandemic-era over-hiring.

On young workers the brief is cautious rather than dismissive. It reports new-graduate unemployment of 5.6% in early 2026, 1.6 points higher than three years earlier, and accepts that AI may be part of the reason. But it sets out competing explanations, including interest-rate rises that began in March 2022 and the shift to remote work, which can slow on-the-job learning and make younger hires less valuable. It also notes that experiments have often found generative AI tools help less experienced workers most, which sits awkwardly with fewer of them being hired.

The Stanford paper is candid about its own weak points. Some divergence between more- and less-exposed occupations predates ChatGPT, particularly around the pandemic. Estimates weaken when occupational education levels are controlled for, which could signal an alternative explanation or the very channel through which AI acts. The gap is also more pronounced in the ADP sample, which over-represents higher-exposure occupations and larger firms, than in national surveys, and estimates that account for overall changes in firm hiring are more sensitive to specification choices.

Against that, the authors point out that the gap kept widening long after interest rates peaked, and that it tracks automation-type AI usage with a clear age gradient, which the rival explanations do not predict. They present the figures as early indicators to be monitored, and the lab publishes a dashboard that it says is updated monthly, so the picture is still developing.

Why a thinner entry pipeline is an operational risk

If the pattern holds, the risk for employers lies less in current headcount than in future expertise. The Stanford data suggest experienced workers are faring comparatively well where AI complements them and in occupations that rely on tacit, experience-based knowledge. Much of that kind of knowledge is acquired in junior roles, through repetition, supervised mistakes and gradually widening responsibility.

Automating the routine tasks that used to justify those roles can look efficient within a single budget cycle. Often the earliest casualty is not an existing employee but an opening that quietly stops being posted. Organisations that stop hiring at the bottom may later find they have too few people ready to supervise automated systems, handle the exceptions and train the next cohort.

The first cost of automation may not be a layoff at all, but the junior job that is never advertised.

In Europe, regulation adds a further reason for care. The EU's Digital Omnibus on AI, published in the Official Journal on 24 July 2026 and in force since 27 July, moved the application date for the AI Act's Annex III high-risk systems to 2 December 2027. Employment is one of the areas Annex III covers, although not every AI system used in those areas is classed as high-risk. The amended Act also expects providers and deployers to take measures that support AI literacy among their staff.

Practical implications for organisations automating work

The evidence is early, contested and largely American. It still supports some practical choices for teams running AI agents, robotics or automation in real operations.

• Automate tasks, not the learning path. When agents or robots take over routine work, rebuild junior roles around reviewing outputs, handling exceptions and supervised escalation, so new staff still develop judgement.
• Track entry-level hiring and internal progression alongside productivity. The Stanford data suggest the adjustment shows up in who gets hired, which layoff counts and headcount snapshots miss.
• Favour deployments that complement staff where the work allows. The link between complementary AI use and stable or rising employment is correlational, but such programmes are also easier to explain to employees, customers and regulators.
• Be precise about attribution. Citing AI for cuts driven by restructuring or cost pressure may impress investors while eroding staff trust, and public claims are likely to draw more scrutiny as rules develop.
• Keep apprenticeship-style routes open. Pair new hires with experienced operators of agents and robots so that tacit, experience-based know-how is passed on rather than lost.
• If you operate in the EU, inventory the AI used in decisions about workers now, assess which systems could fall under Annex III before 2 December 2027, and document AI literacy measures for staff.