The clearest effect of AI on work in the evidence here is a shrinking bottom rung: early-career roles are taking the hit, while more experienced workers in the same occupations are not.
- Entry-level work is where the harm shows. Several studies find less entry-level employment and fewer junior hires in AI-exposed jobs, while headcount for older workers in the same occupations grew.
- How a firm uses AI sets the direction. In one executive survey, firms using AI to automate routine tasks tended to cut entry-level hiring, while firms using it to move juniors into harder work tended to add it.
- Freelance markets show it first. Demand for automatable writing, coding and image work fell after ChatGPT's release, early evidence suggests buyers now weigh price more and credentials less, and AI-related freelance work grew.
- Exposure is not replacement. What a model could do is far from what organisations actually change, and forecasts in this field have a poor record.
- The real loss is how people learned the job. Routine junior tasks are going first, so the practical move is toward judgement, client contact and being able to say why an output is wrong.
A May 2026 survey of nearly 1,500 US executives found that firms using AI mainly to automate routine tasks tend to reduce entry-level hiring — while firms using it to move junior staff into more complex work tend to increase it. Same technology, opposite outcomes. The management decision determines the direction, not the tool.Strada Institute for the Future of Work, Entry-Level Hiring in the AI Era, May 2026, read at source 10 Sep 2026 — “nearly three times (2.7 times) as many senior talent leaders expect AI use to increase entry-level hiring in 2026 as to decrease it” — survey of 1,498 executives and senior talent leaders at US organisations, fielded March 2026, weighted by industry, size and geography.
A companion piece answers the question people actually search for — which jobs are safe from AI — and takes apart the ranked “AI-proof” lists, including one still built on a 2013 forecast that failed.
The evidence, both directions
The bottom rung is being removed
- Stanford researchers report a 16% decline in entry-level employment in AI-exposed occupations by 2025 — while headcount grew for older workers in the same occupations, and in less exposed roles.Reported by MIT Technology Review, A reality check on the AI jobs hysteria, 26 May 2026, on work by Brynjolfsson and colleagues, read at source 23 Sep 2026: “growing in 2025 to a 16% decline in entry-level jobs in AI-exposed occupations. In contrast, head count grew for older workers in the same occupations, as did the number of jobs in the less exposed occupations.”
- 38% of employers say they have moved basic data entry and processing off entry-level staff and onto AI; 31% have raised experience requirements for entry-level roles.ZipRecruiter Economic Research, More Jobs, Higher Bar: The 2026 AI Employer Report, 29 Jul 2026 — survey of over 1,000 US employers, read at source 23 Sep 2026: “38% of employers have shifted basic data processing away from entry-level workers and onto AI, and 31% have raised experience requirements for entry-level jobs as a result.”
- IMF analysis finds employment in AI-vulnerable occupations 3.6% lower after five years in regions with high demand for AI skills.Kristalina Georgieva, New Skills and AI Are Reshaping the Future of Work, IMF Blog, 14 Jan 2026, read at source 17 Sep 2026: “In fact, employment levels in AI-vulnerable occupations are lower in regions with high demand for AI skills—3.6 percent lower after five years than in regions with less demand for these skills.”
Adoption is not translating into cuts
- In the same executive survey, nearly three times as many senior talent leaders expect AI to increase rather than decrease entry-level hiring in 2026.Strada Institute for the Future of Work, Entry-Level Hiring in the AI Era, May 2026, read at source 17 Sep 2026: “Nearly three times (2.7 times) as many senior talent leaders expect AI use to increase entry-level hiring in 2026 as to decrease it, indicating a mixed and often positive near-term outlook.” An earlier version also gave a 46% figure for 2025 that could not be matched to the survey’s own wording, and was removed.
- Research by one frontier lab measuring actual usage rather than theoretical capability finds that although models could in principle assist with over 90% of tasks in some fields such as computer science and office administration, observed automation is far lower — held back by legal requirements, human verification, implementation cost and organisational inertia.
- One lab's own labour-market analysis found a ~14% drop in job-finding rates for exposed occupations post-2022 but described it as "just barely statistically significant", and noted the affected young workers may simply be staying put, switching fields or returning to study.Anthropic Research, Labor market impacts of AI, March 2026, read at source 10 Sep 2026 — the lab’s own framing is that AI “is far from reaching its theoretical capability”; its empirical core is Claude conversations and API traffic, so it reflects Anthropic’s own user base more directly than the economy — the lab's own caveat, quoted rather than paraphrased
- Some researchers argue the decline in these roles began before ChatGPT, and question whether labour markets could react as fast as the AI explanation requires.
When you read a headline about AI and jobs, check which workers it measures. The evidence of harm sits mostly with entry-level roles, and totals across all workers can hide it.
The gig economy: where it shows up first
Freelance marketplaces are the clearest place to watch this, because contracts are short, hiring is fast, and the data is public. What happens in employment over years happens here in months.
The demand side contracted. Research published in the INFORMS journal Organization Science found freelancers in AI-exposed services saw roughly a 2% monthly decline in contracts and about 5% lower earnings after generative AI became widely available; a separate study of a large freelancing platform found job posts for automation-prone writing and coding down around 21% within eight months of ChatGPT's release, and image-creation posts down about 17%. A separate firm-spending study found the share of company budgets going to online labour marketplaces fell from 0.66% to 0.14%, with more than half of businesses that used them in 2022 spending nothing by 2025.Hui, Reshef & Zhou (Organization Science), as summarised by WashU Olin Business School, 24 Aug 2023: “the number of monthly jobs for writing-related freelancers on Upwork declined by 2%, while monthly earnings declined by 5.2%”. Demirci, Hannane & Zhu, Who Is AI Replacing?, CESifo working paper, 2024: “a 21% decrease in the number of job posts for automation-prone jobs related to writing and coding” and “a 17% decrease in the number of job posts related to image creation”. Ramp Economics Lab, Payrolls to Prompts summary, 18 Feb 2026 (paper submitted 28 Jan 2026): “The share of total spend going to labor marketplaces fell from 0.66% in Q4 2021 to 0.14% in Q3 2025” and “More than half of the businesses using freelancers in 2022 have stopped entirely.” All read at source 23 Sep 2026. The spending study covers one payments provider's customers, not the whole economy, and says so. Until 23 Sep 2026 this paragraph credited the 21% and 17% figures to the Organization Science study; they come from the separate Demirci, Hannane & Zhu paper.
A study of 49,610 freelancers and 2.26 million completed contracts, tracked quarterly from early 2021 to early 2026, examined not just how much hiring fell but what buyers started weighting differently. In the most AI-exposed categories, contract volume fell about 7% relative to unexposed ones — and the signals clients had previously relied on lost their pull: the predictive importance of credentials, reputation and self-presentation dropped roughly 8%, while price gained. The shift appeared only in exposed categories, and it has been strengthening rather than levelling off — in the last four quarters measured, contract volume fell about 9.6% and the weight on those signals about 10.1%.
That is the mechanism behind a phrase people use loosely. Competence became a commodity — not because workers got worse, but because clients stopped believing they needed to pay for the difference.Siddiq & Zhang, Human Capital, AI, and Labor Commoditization, UCLA Anderson, submitted 20 Jun 2026, read at source 23 Sep 2026: “the combined importance of all human capital signals in the most exposed job categories falls by approximately 7.8%”, and for the final four quarters, “demand falls by 9.6%, the combined importance of human capital signals falls by 10.1%, and price importance rises by 1.8%” — a difference-in-differences study using text embeddings and Shapley values, covering 21 quarters around ChatGPT's release. A working paper, not yet peer-reviewed, and the strongest available evidence on this specific question rather than a settled finding.
And the same market split, sharply. While commodity work fell, demand for AI-related freelance work is reported to have grown, and clients are reported to prefer people who use AI to augment their work over those who either avoid it or let it do the thinking — platform reporting that was not read at source for this page. The pattern matches the wider evidence above: not less work, but a different distribution of who gets it — with the bottom repriced and the top revalued.
If you sell freelance work that AI can do, early evidence suggests buyers now weigh price more and credentials less. The demand that grew is for AI-related work and for people who use AI to augment their work rather than do their thinking.
What both sides agree on
Exposure is not replacement. Every serious study distinguishes what a model could theoretically do from what organisations actually change. The gap between the two is where all the argument lives, and it is large.
The effect is age-shaped, not sector-shaped. The consistent signal across otherwise conflicting studies is that early-career workers absorb it while workers over 25 in the same occupations do not. That is a different problem from "jobs disappearing", and arguably a worse one: it removes the way people used to acquire the experience that makes them employable later.
Nobody's forecast is reliable. This field's prediction record is poor — as documented elsewhere on this site, a widely repeated forecast that 90% of online content would be synthetic by 2026 — repeated for years in the name of a 2022 Europol report, whose revised version removed the statement — simply did not happen.Europol Innovation Lab, Facing reality?, 2022, publication page read 17 Sep 2026: “In the updated version, a statement from an inaccurate source on the expected future share of synthetically generated content was removed.” Treat every number about 2030 as a scenario, not a measurement.
The tasks that used to be handed to juniors — the routine ones — are the ones being automated first. Those tasks were how people learned the job, which is the real loss. The practical response is to get proximity to the parts that are not routine: judgement, client contact, deciding what the work should be. Being able to operate the tools is now assumed rather than impressive; being able to say why an output is wrong is what remains scarce.World Economic Forum, Future of Jobs Report 2025, as reported by FM magazine, 18 Feb 2025, read at source 23 Sep 2026: “Analytical thinking remains the top core skill needed in 2025, unchanged from the WEF’s previous findings.” The WEF’s own pages refused an automated read that day.
The honest summary is boring, which is why almost nobody publishes it: the evidence gathered here does not show mass unemployment, and it does show the ladder losing its bottom rung. Both can be true. Anyone telling you AI has had no effect on work is ignoring the entry-level data; anyone telling you it is a jobs apocalypse is ignoring the evidence above that adoption has not, so far, turned into broad cuts. The interesting question is not how many jobs exist — it is who gets to start.
Plan around who gets to start, not a count of jobs. If you hire, how you use AI helps decide whether junior roles shrink or grow; if you are early in a career, move toward the parts of the work that are not routine.