← Siddhant Goswami

What AI has done to jobs so far

In 2026, AI became the most common reason US companies gave for cutting jobs. Unemployment barely moved. Both are measured. Both are true. How?

The answer is not in the layoff announcements and not in the unemployment rate. It is in who is no longer being hired. The ladder did not break. The bottom rung was pulled up, and the people who would have stood on it do not show up anywhere a headline looks.

The short answer:

  • There is no economy-wide displacement wave yet. Every independent tracker that counts the whole economy says so, and in July 2026 Anthropic’s own head of economics agreed.
  • The damage is real, and it is on the first rung. Young workers in the most AI-exposed jobs are 19% behind where they would have been. Companies are not firing juniors. They are not hiring them. It has not yet shown up in graduate unemployment.
  • The rung moved up rather than vanished. Entry-level jobs exposed to AI now ask for judgement and ownership, and those are the ones still growing.
  • The big job-creation numbers are almost all forecasts. The measured ones are far smaller.

This piece covers 30 November 2022, the day ChatGPT launched, to 29 October 2026. Everything in it is dated, and everything is sorted into one of two piles: counted, or predicted. The first pile is evidence. The second is opinion, however official the letterhead.

Two people will walk through the numbers with us. Both graduated in computer science from a good college. One finished in May 2022. One finished in May 2025. Same degree, same ability. The three years between them are the subject of this piece.

Three questions to run on any AI jobs claim

Before any number, ask these. They take ten seconds and they sort almost everything you will read on this topic.

  1. Who counted it? A payroll record, a job posting, a layoff filing: counted. A survey of what employers expect, a model built from exposure scores, a forecast for 2030: predicted.
  2. Could it be proved wrong by 2027? If nothing could prove it wrong, it is not evidence yet. The World Economic Forum’s 78 million net new jobs by 2030 cannot be checked until 2030.
  3. Who benefits from saying it? A data-centre lobby counting construction jobs. An AI lab studying itself. A consultancy selling AI transformation. A chief executive telling investors why headcount fell. An AI school telling you to learn AI, which includes me. None of these are liars. All of them have a direction they would prefer the number to point.

The rest of this piece applies these questions to every source it uses, including the ones that support its conclusion. Where a source fails one, it says so beside the number.

The confession

Challenger, Gray & Christmas records the reason companies give for every announced US job cut. AI went from a footnote to the top line in eighteen months.

Period Announced US cuts citing AI Share of all announced cuts
2023 to 2024 about 17,000 about 1%
2025 54,836 about 5%
2026, January to September 120,136 about 21%, the top reason this year

In March 2026, AI led all stated reasons for the first time. In May it was cited for 38,579 cuts, 40% of the month. Since then the monthly share has fallen. In September, AI was the fifth most cited reason, with 3,961 cuts, about 9% of the month. Total announced cuts for January to September are down 39% on 2025, the lowest since 2022. AI still leads for the year, but the peak so far was in spring.

Run the questions. Who counted it? Challenger counts announcements, which is real. But the reason is self-reported by the company. Who benefits? Andy Challenger put it plainly: “Naming AI in a layoff announcement can win over investors while pushing current and prospective employees away. That’s why the messaging has swung from hedging to aggressively citing it.” Sam Altman has said the same: some companies are blaming AI for layoffs they would have made anyway. Fortune calls the pattern AI-washing, and Martha Gimbel of Yale’s Budget Lab suggests a motive: naming AI is easier than admitting that tariffs, immigration policy or a misread economy hurt the business.

The incentive does not run only one way. Companies also have reasons to under-claim AI, to avoid frightening their own staff. Challenger files another 20,219 cuts in 2025 under “technological update”, which may be AI by another name. The true figure is genuinely uncertain.

So the confession is real, but it is a confession to a motive, not proof of a crime. The named cases cut both ways. CrowdStrike cut 5% of its staff in May 2025, telling investors that AI “flattens our hiring curve”. Workday cut about 1,750 roles in February 2025 to shift spending towards AI. Salesforce took its support staff from about 9,000 to 5,000, and Marc Benioff’s explanation was “I need less heads”. Snap cut about 1,000 people, 16% of staff, in April 2026 to save costs, saying AI lets its teams do less repetitive work. Amazon’s Andy Jassy wrote in June 2025 that AI would “reduce our total corporate workforce” over the next few years. Amazon has since cut about 14,000 corporate roles in October 2025 and 16,000 in January 2026, though the January memo did not name AI.

There is a third route, and it needs no AI to do anyone’s work. On 26 October 2026, Microsoft offered voluntary buyouts to about 60,000 employees, more than a quarter of its 223,000 staff, while planning about $175 billion of capital spending this year, mostly on AI data centres and chips. Reuters reported the aim as containing costs and helping to fund that build. Three days later it reported more than $36.6 billion of profit for the quarter. Microsoft has not said AI is doing these people’s work. The money is moving from payroll to buildings that, as the data-centre counts at the end of this piece show, employ very few people. File it as capital reallocation, not displacement. It is an offer, not a cut, and nobody will know how many accept until December. Buyouts also mostly appeal to experienced staff, so this is a separate effect from the shrinking first rung, not more evidence for it.

The alibi

Every independent tracker that counts the whole economy finds the same thing: no displacement wave.

  • Yale’s Budget Lab, updated monthly, found in October 2026 that measures of AI exposure and AI use show no connection to changes in employment or unemployment. The cooling in hiring shows up in exposed and unexposed jobs alike.
  • California’s AI-Unemployment Tracker, built on actual unemployment claims, finds no statewide surge from AI-exposed occupations through May 2026. It can detect known AI layoffs when they happen. They are not yet big enough to move the state total.
  • Bloomberg Economics found in June 2026 that vacancies in AI-exposed UK roles were already falling before ChatGPT and have risen since summer 2024, and that UK tax data shows private-sector headcount in several vulnerable sectors up since ChatGPT launched.
  • Goldman Sachs and Morgan Stanley, scoring jobs by exposure, put AI’s net effect on US unemployment at about 0.1 percentage point. Goldman estimates that substitution removed about 25,000 US jobs a month over the year to April 2026 while augmentation added about 9,000, a net loss of about 16,000 a month, falling mostly on less experienced workers. Small, real, and a model rather than a count.

The loudest forecaster has also softened. In May 2025, Dario Amodei, Anthropic’s chief executive, told Axios that AI could wipe out half of all entry-level white-collar jobs and push unemployment to 10 to 20% within one to five years: “We, as the producers of this technology, have a duty and an obligation to be honest about what is coming.” Others said similar things that spring. Shopify’s Tobi Lütke told staff that “teams must demonstrate why they cannot get what they want done using AI” before asking for more people. Fiverr’s Micha Kaufman went further: “AI is coming for your jobs. Heck, it’s coming for my job too.”

In May 2026, Amodei invoked the Jevons paradox: when something gets cheaper, people tend to use more of it. He did not withdraw the warning. In July, Peter McCrory, Anthropic’s head of economics, wrote that he did not expect unemployment to be noticeably higher a year from now, at least not because of AI.

Study what people do, not what they say. Here is what the companies did next.

Company What they said What they did
Klarna February 2024: its AI assistant does the work of 700 agents. December 2024: AI “can already do all of the jobs that we, as humans, do”. Headcount fell from 5,527 to 3,422 between 2022 and 2024, mostly by not replacing leavers. By May 2025 it was hiring human agents again; the chief executive said the push for cost had cost quality.
IBM May 2023: about 7,800 back-office roles could be replaced by AI over five years. May 2025: AI agents took over the work of hundreds of HR staff, but “our total employment has actually gone up”.
Salesforce February 2025: “we’re not going to hire any new engineers this year”. April 2026: “We’re hiring 1,000 new grads & interns right now to ride the AI exponential.”
Duolingo April 2025: “AI-first”, and “we’ll gradually stop using contractors to do work that AI can handle”. Four weeks later: “I do not see AI as replacing what our employees do.” (Fortune)
Shopify April 2025: prove AI cannot do it before asking for headcount. The policy stands.
Amazon June 2025: AI will shrink the corporate workforce. About 30,000 corporate cuts since October 2025, then two smaller rounds in October 2026, about 1,250 roles in retail and devices, with AI not named.

The pattern is wider than a few reversals. Gartner predicts that by 2027, half the companies that cut customer-service staff because of AI will rehire people for similar work under different job titles. Of the 321 customer-service leaders it surveyed in October 2025, only 20% had actually reduced staff because of AI. Where companies tried to fully automate relational, high-stakes work, quality fell and people came back.

So the suspect has an alibi for the economy as a whole. That is not the end of the case.

Where it hurts: the first rung

Back to our two graduates. The 2022 graduate walked into a market where young workers in AI-exposed jobs were still being hired at the same pace as everyone else. The 2025 graduate did not.

In August 2025, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford used ADP payroll records to ask who was losing work in AI-exposed jobs. Their August 2026 revision, with data to June 2026, finds employment of 22 to 25 year olds in the most exposed occupations is now 19% below where it would be had it kept pace with less exposed peers, up from 15% a year earlier. Read that carefully: it is a gap against a comparison group, not a 19% loss. In absolute terms, between November 2022 and June 2026, young workers in the two most exposed fifths of occupations fell by about 11%. Young workers in the three least exposed fifths grew by about 10%, and so did workers aged 35 to 49 in the most exposed jobs. The losses sit where AI substitutes for work rather than complementing it, and they show up in jobs, not wages. The authors find no evidence of economy-wide displacement and call their findings early, descriptive signals, not proof of cause.

The mechanism is hiring, not firing. Seyed Mahdi Hosseini and Guy Lichtinger at Harvard traced 65 million résumés across 280,000 firms. Six quarters after adopting generative AI, firms had about 9% fewer junior employees than firms that had not. Senior employment was largely unchanged. Nobody fired the 2025 graduate. Nobody made the offer.

The clearest single case is the company that makes the model, which also fails question three. On 16 October 2026, four Anthropic economists, McCrory among them, published how employment is changing at Anthropic. Since early 2025, juniors have fallen from 21% to 14% of its technical staff, and from 27% to 9% of technical hires. Junior postings have dropped sharply since December 2025. In April 2026, about a third of machine-learning managers and a fifth of software managers said they were hiring less because of AI. Meanwhile, commits per engineer reached five times their level of six months earlier, and by August its staff’s agents were logging six hours of work for every human hour. The company is still growing and still hiring engineers. It is hiring fewer juniors among them.

The authors give two cautions: the junior slowdown is descriptive, not causal, and much of what Anthropic does cannot be replicated across the economy. Both are right. It is still the economist who expects no noticeable rise in unemployment co-signing a measured collapse in junior hiring at his own firm.

Now the confounds, because the method demands it

The Stanford gap opens in late 2022. So does everything else that could explain it. The Federal Reserve began raising rates in March 2022. Tech companies cut about 430,000 jobs across 2022 and 2023 after over-hiring through the pandemic, by Layoffs.fyi’s count. A US tax change that took effect in 2022 ended the immediate deduction of software R&D wages, making engineers more expensive to employ. Yale finds the hiring slowdown in exposed and unexposed jobs alike, which fits those causes better than it fits AI.

Some of the numbers this debate leans on predate the mechanism. At the largest technology companies, new-graduate hiring fell 25% from 2023 to 2024, before agentic coding tools existed. Overall IIT placement fell from 90% to 80% between the 2021-22 and 2023-24 academic years, according to a parliamentary committee. Indeed’s index of US software postings stood at 78 in early October 2026 against 100 in February 2020, but most of that fall came in 2022 and 2023. These show a weak junior market. They do not show AI.

What survives the confounds: the gap is specific to AI-exposed occupations and specific to the young; it kept widening into 2026 after the Federal Reserve began cutting in September 2024 and after the July 2025 tax law restored immediate expensing for domestic R&D; and firms that adopted AI show it more than firms that did not. That is three fingerprints pointing at one suspect. It is not yet a conviction.

Why unemployment does not show it

The New York Fed puts unemployment for US graduates aged 22 to 27 at 5.6% in the second quarter of 2026, against 2.9% for all graduates. Recent graduates always run above the all-graduate rate, so that gap on its own is normal. The level is what matters, and it is well above the roughly 4% of 2019. 42% of recent graduates are in jobs that do not need a degree. On 2024 data, computer science graduates had 7.0% unemployment and computer engineering graduates 7.8%, among the highest of any major.

Robert Fairlie and Jane Wu studied US survey data for the summer of 2026 and put unemployment for bachelor’s graduates aged 22 to 25 at 7.3%, inside the range of the previous four summers (6.3% to 7.8%). It has not risen against older graduates or young people without degrees, and they find no significant link to how exposed a graduate’s occupation is to AI. They call it an early snapshot.

These do not contradict Stanford. Stanford counts who holds the exposed job. Fairlie and Wu count who is still looking. Our 2025 graduate, who never got the software offer and took a job that does not need the degree, appears in the first and not the second. The first rung moves before the unemployment rate does. That is the resolution of the puzzle this piece opened with.

India

India looks different, and its data is thinner. TCS cut about 12,000 roles, 2% of its workforce, in July 2025, but its chief executive said it was not because of AI. Nasscom estimates the sector still added about 1.35 lakh jobs in FY26, taking it to 59.5 lakh, and Infosys hired more than 20,000 freshers. The IIT placement fall above predates the tools. There is no Indian equivalent of the Stanford payroll study, so the first-rung finding is, for now, an American one.

The rung did not vanish. It moved up.

If the first rung were simply gone, the advice would be to leave the field. It is not gone. It has been redefined, and the data on that is the most useful finding in this whole debate.

PwC’s 2026 AI Jobs Barometer read over a billion job ads, including 2.4 million US entry-level postings. Entry-level roles most exposed to AI are now seven times more likely to require skills traditionally asked of seniors: leadership, creativity, dealing with people face to face. Those seniorised entry-level roles grew 35% since 2019. Entry-level roles that were not seniorised shrank 10%. PwC sells AI transformation, so question three applies, but this is a count of ads, not a forecast.

Anthropic’s own data says the same from the hiring side. Mentions of candidates’ AI use in its interview scorecards have roughly tripled for technical roles since 2024. The comments that grew most were about autonomy and judgement. Discussion of domain expertise shrank.

Postings agree. Since Claude Code launched in February 2025, US software postings on Indeed have risen about 15% while postings overall fell 7%. But between May 2025 and May 2026, 71% of that growth came from senior roles. AI skills appeared in 2.5% of US job postings in 2025, up 55% on the year before, and postings asking for agentic AI went from 0.06% to 0.23% in a year, nearly 90,000 ads. PwC found jobs requiring AI skills grew 69%, about eight times the growth of the job market as a whole.

Augmenting more than automating, for now

What AI is actually used for tells you where the pressure lands. Anthropic publishes what people use Claude for. Computer and mathematical work is about a third of conversations on Claude.ai and nearly half of business use through the API. In September 2025, automation, where people hand over a whole task, overtook augmentation for the first time. By November 2025 augmentation was back ahead, 52% to 45%. Use is lowest at both ends of the pay scale. In Anthropic’s June 2026 report, early-career workers were the most worried about losing their jobs. Microsoft researchers mapped 200,000 Copilot conversations onto occupations and found interpreters, translators, historians and writers had the most overlap with what AI does well. They warn that reading this as a list of jobs about to disappear “would be a mistake”.

Here is the irony our 2025 graduate lives inside. The controlled experiments mostly show gains, and the largest go to the least experienced:

A tool that makes a junior perform like someone with years of experience could have made juniors more valuable. So far the posting data suggests firms are instead asking juniors to arrive already experienced.

Models are getting better at whole pieces of work. OpenAI’s GDPval asks experts to judge AI against professionals on 1,320 real tasks from 44 occupations. In September 2025, the best model, Claude Opus 4.1, matched or beat the expert 47.6% of the time. By April 2026, OpenAI reported 84.9% for GPT-5.5, a number it reports about its own model. But these are tasks, not jobs. A job is dozens of tasks plus the judgement of which one matters today.

Inside companies, the value is still hard to find. MIT’s NANDA project reported that 95% of organisations were getting zero return from generative AI pilots. Its sample was small and the work is not peer-reviewed, so read it as a signal, not a verdict. IBM’s 2025 survey of 2,000 chief executives found only 25% of AI initiatives had delivered the return expected, and only 16% had scaled across the company. The capability is high on tasks. The capture is still low on jobs and companies.

The rung that was removed was the one where you did routine work for two years and learned judgement along the way. The rung that replaced it asks for judgement on day one. The advice below follows from that.

What to do

A disclosure first. I co-founded 100xEngineers, an applied AI school, and advice that ends in “learn to work with AI” is advice I am paid to give. Run question three on this section too.

The evidence points one way, though it is three correlations and not proof. Plan to enter at the second rung. Expect nobody to let you learn judgement on the job for two years. Bring it, in a form a stranger can check.

If you are graduating now

The 2025 graduate’s problem was not the degree. It was that the degree certified routine ability, and routine ability is what firms are buying less of at entry. What they are buying instead, by their own job ads and their own interview notes, is judgement and ownership.

The only proof of judgement that survives an interview is work a stranger can check without you in the room: a repository that runs, a system with real users, even two, a write-up that says what broke and why. That is the only definition of done I trust, and it is increasingly the one employers use. Before you show any piece of work, ask: could someone verify this without talking to me? If not, it is not done.

Do the work yourself before handing it to the model, because you can only judge what you can do. The 2022 graduate learned that through two years of routine work. You have to do it on purpose, on your own time, because fewer employers will pay you to.

If you are deciding what to study

The US and India are moving in opposite directions. In the US, bachelor’s computing enrolment fell 3.1% in 2025, and 62% of computing departments reported declining numbers. The National Student Clearinghouse counts a fall of 8.1% in computer science undergraduates at four-year institutions. In India, B.Tech admissions reached an eight-year high of 12.53 lakh in 2024 to 2025, led by computer science, AI and data science. If the American pattern reaches India, more graduates will be arriving at a smaller routine rung.

Study computing anyway, with one condition. Do not study it to write routine code, because that is the skill the seniorisation data says is no longer bought at entry. Pair it with a field where a human has to sign off on the output: finance, law, health, infrastructure, physical operations. This is where I move from evidence to judgement, so treat it as such. In the most AI-exposed roles PwC finds skills changing 66% faster than elsewhere, and a domain where someone is accountable for the result is where the review work lands.

If you already work

You are the 2022 graduate, or older. You are not the one being squeezed. For now, AI is augmenting more people than it replaces, and companies most exposed to AI have grown headcount faster than those least exposed, by PwC’s count. PwC puts the wage premium for AI skills in job ads at 62%, up from 57% a year earlier. That is a gap between ads, not a raise you get for learning a tool, so read it as a signal of demand and not a promise.

The work most at risk is routine information processing: claims handling, template writing, basic analysis, first-line support. The work least at risk combines physical presence, high stakes, regulation and relationships. The move is up one level on every task you own: from doing it, to directing it, checking it and owning the result. The bottleneck is no longer the output. It is the review. At Anthropic, reading everything their agents produce would take even non-technical staff more than half the working week, and the share of sessions spent on building, reviewing and monitoring agents rose from 10% to 22% in six months. Be the person who can do the review, because that is the job the first rung now starts at.

What would change my mind

A conclusion with no number attached to its failure is an opinion. These are the numbers, and when I will check them.

Claim Fails if Check on
The first rung is structural, not cyclical Stanford’s gap for 22 to 25 year olds in exposed jobs falls below 17% in the next revision, or recent-graduate unemployment falls back towards 4% while the all-graduate rate stays flat The next Stanford revision; New York Fed quarterly updates
AI-cited layoffs are mostly a story Challenger’s monthly AI share returns above 30% for two consecutive months and US unemployment rises above 4.6% in the same quarter Monthly, through March 2027
The rung moved up rather than vanished PwC’s next barometer shows seniorised entry-level roles no longer growing, or AI-exposed entry-level postings falling more than 10% in a year PwC’s 2027 barometer, likely June 2027
AI skill is a durable advantage The AI-skills premium in job ads falls below 50% PwC’s 2027 barometer
Microsoft is capital reallocation, not displacement Microsoft says AI is doing the departing staff’s work, or follows the buyouts with forced cuts Microsoft’s disclosures, from December 2026

One test has already tripped. When this piece first went up on 7 September, it said AI-cited cuts breaking decisively above 10% of all cuts would shift the picture towards genuine displacement. They broke it in March and stand at about 21% for the year. I missed it because I was working from old numbers. It did shift the picture, but less than it sounds, and the new test is stricter for that reason: a stated reason crossing 10% turned out to be cheap, so the new test pairs the stated reason with the unemployment rate, which is not.

The macro risk is real and unresolved. US employers added only 29,000 jobs in September, July was revised to a loss, and unemployment rose to 4.2%. Nothing in that report ties the slowdown to AI. But in October, Goldman’s Joseph Briggs argued that “automation-related labor disruption is limited when the overall economy is healthy but accelerates during cyclical slowdowns”, and put eventual displacement at up to 9% of workers in rich economies, plus 3% through slower hiring. If he is right, a recession is when the AI story stops being only a story. That is why the second test watches both numbers together. Microsoft is the counterexample: it is shrinking while its profits rise, so at least some of this is not waiting for a recession.

Ask question one of every big creation number and it fails. The most quoted comes from the World Economic Forum. Its Future of Jobs Report 2025 projects 170 million jobs created and 92 million displaced by 2030, a net gain of 78 million. It is a survey of over 1,000 employers describing what they expect. That is useful. It is not a measurement.

The data-centre numbers are weaker still. A figure of 4.7 million US construction jobs circulates widely, credited to the American Edge Project, a policy group funded mainly by Meta. I could not find it in anything the group has published, and one local presentation gives it as a range of 4 to 7 million, so it may be garbled. A projection from Kevin McCarthy’s ALFA Institute says $100 billion of data-centre investment would support about 500,000 jobs, mostly in manufacturing.

When anyone counts, the numbers shrink. Virginia’s data-centre operators reported 1,610 net new jobs in fiscal 2025, against a sales-tax exemption worth about $1.9 billion: roughly $1.2 million of forgone tax per job, on the operators’ own figures. Dany Bahar and Greg Wright, writing for Brookings, first reported larger effects in May 2026. In August they revised their findings: a county’s first large data centre brings about 100 to 200 data-processing jobs, no effect on wages, and little evidence of wider local gains. Chip plants employ more people than server halls: TSMC’s Arizona plan promises about 6,000 direct jobs.

Nobody has a credible measured count of jobs created by AI. The measured signals, the AI-skill postings above, are real and small. Everything at the scale of millions is a projection, and the data-centre projections come largely from people who build data centres.

What these numbers cannot tell you

  • One company is not an economy. Anthropic’s figures come from an AI company growing fast, studying itself. Its authors say so, and call the junior trend descriptive.
  • Stated reasons are not causes. Challenger counts what companies announce, in the US only.
  • Models are not counts. Goldman’s 16,000 a month is an estimate built from exposure indices, and its 9% is a projection over a decade or more.
  • The confounds are not separated. No study here cleanly isolates AI from rates, the 2022 tech correction, the tax change, tariffs or immigration. The first-rung finding is three correlations pointing one way, not a causal estimate.
  • Single sources. The 4.7 million data-centre jobs figure traces to a Meta-funded group and could not be found in its reports. Klarna’s reversal rests largely on one Bloomberg interview. The Bloomberg UK analysis is paywalled and I read it through summaries.
  • India’s data is thinner. Much of it reaches the public through press coverage rather than primary AICTE or Nasscom documents.
  • The New York Fed’s figures by major lag by a year or more and carry wide margins of error.

Changelog

Date Change
29 October 2026 Restructured around the puzzle and the two graduates. Added the three questions, the confounds, the seniorisation data, Yale’s and Bloomberg’s tracker results, and dated tests with thresholds. Moved the job-creation numbers to a sidebar. Earlier the same day: graduate unemployment, September jobs, Goldman, Microsoft buyouts, Amazon’s October cuts.
19 October 2026 Added Anthropic’s employment report. Corrected METR, GDPval and the September Challenger figures; added caveats.
16 October 2026 Rewritten with figures to October 2026.
7 September 2026 First published. Listed AI-cited cuts above 10% as a test; it had already tripped in March.

Updated 29 October 2026.