Pulling Up the Ladder
AI is taking the entry-level work that used to train juniors in knowledge-work professions, and the response from those industries — though more substantial than I'd assumed — isn't yet anywhere near the scale of the displacement. A look at the data, and an honest take from someone sitting on the comfortable side of the age split.
When I wrote about the uncomfortable parts of building software with AI two weeks ago, one stat sat with me longer than the others. Stanford's analysis of ADP payroll data (the latest revision, November 2025) shows that employment for workers aged 22–25 in the most AI-exposed occupations has dropped about 16% since late 2022, whilst employment for older workers in the same occupations has stayed essentially flat [1]. For software developers specifically, the 22–25 cohort is down closer to 20% from its late-2022 peak [2]. The senior end of the curve is doing fine, but the bottom end is being quietly demolished.
I'm on the comfortable side of that split, since I'm nearly in my forties with almost two decades of governance, audit, and security experience behind me, and the way I work with AI assumes I already know how to read code that an assistant has just produced. I'm not the person who would have been competing for a junior dev role at a Calgary consultancy in 2026. But I would have been that person twenty years ago, and the version of me who got his first IT job in Johannesburg in the early 2000s would not have had a route in if the entry-level work I did then was being done by AI now. So I think it's worth being honest about what's happening, since the people best placed to talk about it are mostly people who don't have to live with the consequences.
What the data actually says
The Stanford paper is called "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence" (Brynjolfsson, Chandar, and Chen, 2025) [1]. It uses ADP payroll micro-data covering several million workers across tens of thousands of US firms, which is a more granular dataset than the standard BLS releases. The headline finding, after controlling for firm-level shocks, is that workers aged 22–25 in the most AI-exposed occupations have seen a 16% relative decline in employment since the late-2022 onset of widespread generative AI. The earlier draft of the paper (data through July 2025) had 13%; the November update extended the series to October 2025 and the number grew [3].
The framing I keep seeing in tech press is "AI is taking entry-level jobs whilst senior roles boom." That's not quite what the paper says. Older workers in AI-exposed occupations are stable, not growing. The growth is in less-exposed occupations across age bands. For software developers, the 22–25 cohort is down nearly 20%, the 26–30 cohort is down slightly, and older cohorts show no noticeable change in trend [2]. So it's not that AI is making senior developers more valuable in any direct sense. It's that the work that used to be done by juniors is now being done by AI working in tandem with seniors, and the result is that nobody is hiring the juniors.
There are reasonable competing explanations the authors address in their follow-up post [3]. Interest rates rose, the labour market for new grads is genuinely tougher across the board, and there are other macro factors at play. But the displacement they find is concentrated specifically in occupations where AI is more likely to automate than augment human work, which is the cut you'd expect if AI exposure were doing real explanatory work. It's not a slam-dunk causal claim, but it's the most defensible read of the available data, and the trend has been going in one direction for three years.
The ladder, and what was on it
When I was 22, the work I did was almost entirely the work AI is best at now. I wrote internal documentation that nobody would ever read carefully. I ticked-and-bashed audit samples until late into the evening. I sat in audit meetings and took notes that I'd later turn into reports that senior people would review, edit, and return with changes. I audited firewalls, and OSes, and applications by following a checklist somebody else had written. The work was not glamorous, and a lot of it was, frankly, boilerplate. But every piece of it taught me something specific about how systems behaved when they hit the real world, and the people I worked alongside corrected me when I got it wrong, which is how I learned what "wrong" looked like in the first place.
The ladder out of that early-career work had two functions. The first was filtering, since you found out fairly quickly whether a junior could think about the system as a whole or whether they could only execute the immediate task. The second, and the one that matters more for this argument, was training. You learned by doing the unglamorous work, by being corrected, by making the small mistakes that don't break anything serious and then watching what happens when you don't catch them in time. There was no other way to learn that. Reading books didn't teach you how to audit a set of call center tapes for a touchy insurance client. Doing an audit bootcamp didn't teach you what it feels like to write a report that management hates and fights you on. You learned by occupying the bottom rung long enough that the rungs above it started to make sense.
AI is now genuinely good at the bottom rung. AI can write the documentation, tick-and-bash better than I ever could, generate the reports, and (with the right set of deterministic rules) audit the firewall, OS or application (with caveats that I've written about elsewhere, since "good at" doesn't mean "trustworthy without governance"). But the ladder isn't just a set of tasks you can outsource. The tasks were the training mechanism, and removing them from the workflow doesn't just remove the cost, it removes the curriculum. AI is encouraging business leaders to cut the bottom rung off the ladder and isn't equally encouraging them to build another way up.
The CISO observation, because of course
I spent enough years in audit and security to develop a fairly cynical instinct about systems that can't reproduce themselves. The ones that depend on a small group of people who all came up through a now-closed path tend to be brittle in ways that aren't visible until those people retire, leave, or burn out. You see it in industries where the apprenticeship pipeline collapsed and nobody built a replacement (the trades in South Africa and much of Canada are a textbook case). You see it in regulated industries where a generation of compliance officers got their training by doing the boring on-the-ground work, and then the boring on-the-ground work got centralised or automated and the next generation never built the same instincts. The pattern is always the same: the system runs fine for ten years on the people who learned the hard way, and then it doesn't, and by the time you notice, you've lost the institutional capacity to teach the new people.
A workforce with no junior pipeline is exactly this kind of system. It works fine as long as the seniors keep working, and the AI assistants keep being available, and nothing fundamental changes about how software gets built, or systems get audited, or laws are interpreted and argued. None of those are safe assumptions over a ten- or twenty-year horizon. Seniors retire. AI vendors change their pricing models, their terms of service, and occasionally their entire technical direction. The work itself shifts in ways that make today's senior expertise less load-bearing than it currently feels. And when one of those things happens (probably all of them, eventually), the industries will discover that they've spent a decade not training anyone, and there is no fast fix for that.
What is and isn't being done
The thing that genuinely surprised me when I went looking is how thin the response has been from the companies most responsible for the shift. Anthropic, OpenAI, Google, Microsoft, and GitHub all have programmes aimed at early-career talent, but none of them are dedicated AI-native junior apprenticeships of the kind this moment seems to call for. Anthropic's Fellows Program is research-focused (May and July 2026 cohorts, applications open as of this writing), and it's aimed at applicants with quantitative technical backgrounds who want to work on alignment research, not at the 22-year-old who can't get a foot in the door [4]. OpenAI's Residency programme pays well ($18,333 a month, which is a real signal of intent) but targets adjacent-field researchers from maths, physics, and neuroscience, not new grads [5]. Google's Software Engineering Apprenticeship [6] exists (and credit to them for keeping it running), but it predates the AI shift and isn't structured around it. GitHub has launched an Agentic AI Developer certification through Microsoft Learn [7], which is a credentialing scheme rather than a hiring programme, and it puts the onus on the candidate to figure out how to get hired.
On the government side, the UK has done the most concrete thing so far. TechFirst, announced in June 2025 at London Tech Week, is a £187m programme broken into four sub-schemes: TechYouth (£24m for secondary-school AI literacy), TechGrad (£96.8m for undergraduate scholarships), TechExpert (£48.4m for PhD support), and TechLocal (£18m for regional tech employment) [8]. Whether it works is an open question (these things usually take five years before you can tell), but it's at least an attempt at a coordinated response, and £187m is real money. Canada has nothing equivalent at the federal level. TECHNATION's Career Ready Program offers wage subsidies through the existing Student Work Placement framework [9], and the Policy Options piece from October 2025 made the case for a dedicated youth AI-employment scheme [10], but as of now there's no Canadian counterpart to TechFirst and no signal that one is coming. For a country that talks a lot about being an AI powerhouse, that's a fairly conspicuous gap.
What I notice in all of this is the absence of anyone treating the junior-employment problem as the responsibility of the industries that created it. The frontier labs are building products that demonstrably displace early-career work; the response from those labs is to run research fellowships for people who already have advanced degrees. That's not a moral judgment so much as an observation about how incentives align, since nobody at Anthropic or OpenAI gets promoted for solving labour-market problems that take a decade to manifest. But it does mean that the people best placed to design AI-native apprenticeships are the ones least likely to do it.
And I haven't even begun to touch on the non-tech industries where the bottom of the ladder is being upended by AI. Junior lawyers used to do document review, contract markup, discovery, and case-law summarisation, all of which AI can now do faster and (with the right guardrails) more consistently. Junior accountants did variance analysis, journal entry preparation, and audit fieldwork that gets thinner every year as automation eats the bottom of the audit programme. Junior analysts in consulting, finance, and policy spent their first two years building decks, doing literature reviews, and summarising primary research, and that work is now thirty seconds of prompting. Junior journalists wrote the local-news filler and the press-release rewrites that funded the newsroom while they learned how to actually report a story.
Each of these professions has started to grapple with the question publicly, with uneven seriousness. Law has gone furthest. The American Bar Association's Task Force on Law and AI released its Year 2 report in December 2025 framing AI as legal-profession infrastructure rather than experiment [11], Thomson Reuters Institute has been writing about how AI is forcing law firms to rethink the training of junior lawyers [12], Canadian Lawyer Magazine has been running op-eds since 2025 about the apprenticeship model being hollowed out [13], and Goodwin Procter rolled out an explicit AI-era response: a firm-wide AI training initiative called Propel, plus a separate five-week pre-billable First Year Development Program designed precisely because AI does the work first-year associates used to learn on [14]. Accounting is close behind. ICAEW launched its Next Generation ACA in September 2025 (the biggest qualification redesign in three decades, with AI-era practice integrated through the work-experience component) [15], Deloitte publicly reshaped its three-year graduate training programme to match [16], and the Journal of Accountancy ran a March 2026 piece titled simply "How will accountants learn new skills when AI does the work?" [17] Consulting is grappling with it visibly too, though more through firm-by-firm restructures (Alvarez & Marsal flipping the pyramid into what's been called a "box" model, one of four emerging alternative shapes [18], McKinsey's North American chair publicly committing to 12% more graduate hires in North America in 2026 than 2025 [19]) than through industry-body reform. Journalism is the weakest of the four. Reuters Institute [20] and Poynter [21] are doing serious work, Columbia's CJS2030 AI Initiative has revamped curriculum [22], but most of the public conversation in journalism is about AI ethics, labour rights, or upskilling existing reporters — not how the next generation of cub reporters trains when AI does the cub work.
None of these reforms are yet at the scale of the displacement they're responding to. A five-week bootcamp at one firm and a redesigned UK qualification are not nothing, but they're not the answer either, and the conversations are happening at firm and industry-body level, mostly out of view of the 16-year-old trying to decide what to study. But the conversations exist, and they're worth crediting, because the closing argument of this piece is that the response needs to be much bigger and much faster — and "much bigger and much faster than X" lands harder when you've named what X actually is.
The Stanford data only studies the cohort it can measure (US payroll workers in AI-exposed occupations), but the pattern almost certainly generalises wherever entry-level work is mostly information-processing, and that covers most of knowledge work.
What a realistic path might look like
I'm wary of writing the section where the solo founder pretends to have the answer to a structural labour-market problem that spans half the economy, since I plainly don't. But I think there are a few honest things worth saying about what an actual path forward might look like.
The first is that apprenticeships across knowledge work need to be reframed around AI-pair work rather than AI-replaced work. The junior in 2026, whether that's a developer or a paralegal or a junior auditor or a financial analyst, isn't going to relearn the skills of a 2010 junior, since most of those skills are now table-stakes for the assistant. What they need is the harder, more durable skill: knowing when the assistant is wrong, knowing what to verify, knowing how to read its output for the things it doesn't say. That's a skill that can absolutely be taught, but it has to be taught alongside the work, by someone senior who knows what good looks like. It can't be a bootcamp or an online course. It has to be a structured time-bound apprenticeship of the kind the trades have run for centuries, and the relevant industries would need to actually fund it. Right now most of them are funding share buybacks or investments in AI infrastructure instead.
The second is that this is genuinely a place where solo founders and small firms can do disproportionate good, since we're the ones most directly using AI in ways that displace junior labour, across every knowledge-work sector. If you're a solo lawyer using AI for discovery, a small accounting practice using it for fieldwork, a one-person consultancy using it for analysis, or in my case a solo founder building software with AI assistance, taking on a junior to learn alongside you (paid properly, treated seriously, not unpaid-internship nonsense) is one of the few things at the individual scale that pushes back against the trend. I haven't done this yet myself, since I'm still figuring out whether my own workflow is stable enough to support it. But I think anyone working at this scale should be asking themselves the same question, and I'm asking it more seriously than I was six months ago.
The third is that the industry conversation, in every industry this affects, needs to stop pretending this is a temporary adjustment. Every cycle of automation in the last fifty years has been followed by reassurances that displaced workers would find new roles in the new economy, and whilst most of the time that's been broadly true at the aggregate level, it unacceptably discounts the individual lives that got crushed in the gap. This one might be different, since the rate of change is faster than previous cycles and the affected work is the work people used to learn by doing. We don't have a track record of multiple industries simultaneously replacing their own entry-level training mechanism and then successfully rebuilding it from scratch. There's no guarantee we'll manage it this time. Pretending otherwise is convenient if you're already past the bottom of the ladder, which - unsurprisingly - most of the people setting the narrative in each of these industries are.
The personal bit, again
I started this piece by saying I'm on the comfortable side of the age split, and I want to come back to that. The 16% number isn't an abstract data point for me, since I have nephews and friends' kids who will soon be roughly the age it's measuring, and the conversations I'll be having with them about what they should study and where they should look for work are noticeably harder than the equivalent conversations would have been five or ten years ago. A few of them won't be looking at tech at all. They'll be looking at law, accounting, consulting, journalism, policy, the professions that used to be the safe alternative to tech for a smart kid who didn't want to code. And I won't have great advice for any of them. The honest version is something like "the path I took isn't really available anymore in tech, and the equivalent paths in the professions you're looking at are quietly closing too, and the reforms the industries have started aren't yet at a scale that you, as a 16-year-old trying to decide what to study, can plan around, and I don't know what the new path looks like at the scale this actually needs."
That's not a comfortable thing to write, and I notice that the temptation to soften it (to add a hopeful close, a "but the human spirit prevails" gesture, a list of skills that will future-proof the next generation across every industry) is strong. I don't think any of those would be honest. The right close is the uncomfortable one, which is that I'm building products with AI in ways that I think are good and useful and probably correct, and other people in other industries are doing the same with their own AI use, and we're all also benefiting from a structural shift that's costing real opportunities to real people who would have been our peers in another decade. Both of those things are true at once and the least any of us can do is name them.
There's a chance the next decade sorts this out. Apprenticeships might come back, in tech and in the professions alongside it. Industries might step up, individually or collectively. Governments might do something coordinated rather than the patchwork we're seeing now. The data might turn out to be measuring a transition rather than a new equilibrium. I genuinely hope so, and I'd be the first person to write a follow-up piece in 2030 saying I was too gloomy about it. But hope isn't a plan, and right now the plan is mostly silence from the people who built the thing that's pulling the ladder up, and a response from the affected industries that's real but nowhere near the scale of the displacement. The talking has started in some places. The talking now needs to kickstart rapid action at a scale that matches what's actually happening — replacing the rungs we're quietly stripping away, and not just acknowledging that we're stripping them.
Sources
1. Stanford Digital Economy Lab — Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence — Brynjolfsson, Chandar, and Chen (2025); 16% relative decline in employment for workers aged 22–25 in AI-exposed occupations since late 2022, ADP payroll micro-data, November 2025 revision. Also reported by Fortune, CNBC, and TIME. 2. Bharat Chandar — A primer on "Canaries in the Coal Mine" — Author's plain-language summary of the paper, including the software-developer-specific cut: "Employment for 22 to 25 year old software developers in the ADP data is down nearly 20% between its peak in late 2022, around the time of ChatGPT's launch, to July 2025." Also notes older developer cohorts are flat in trend. 3. Stanford Digital Economy Lab — Canaries, Interest Rates, and Timing — Follow-up post extending the data series to October 2025; the 13% figure in the original paper grew to ~16% by October 2025. Explicitly addresses interest rates as a competing explanation; touches on other macro factors more briefly. 4. Anthropic Alignment — Fellows Program 2026 — May and July 2026 cohorts, research-focused, aimed at applicants with quantitative technical backgrounds (no PhD or prior alignment experience required, but a strong quantitative background expected). 5. OpenAI — Residency — Six-month programme at $18,333/month, targeting adjacent-field researchers from maths, physics, neuroscience; 2026 cohort applications closed. 6. Google Careers — Software Application Development Apprenticeship (March 2026 cohort) — Existing programme with March 2026 and October 2026 cohorts; not framed as AI-native. 7. Microsoft Learn — GitHub Certified: Agentic AI Developer — Credentialing scheme via Microsoft Learn, beta as of writing. 8. gov.uk — Free AI training for all (June 2025) — TechFirst announced 8 June 2025 at London Tech Week, £187m total across four sub-schemes: TechYouth (£24m), TechGrad (£96.8m), TechExpert (£48.4m), TechLocal (£18m). Sum: £187.2m. Also Learning News coverage. 9. TECHNATION — Career Ready Program — Federal Student Work Placement Program; 50% wage subsidy up to $5,000 per placement; Summer 2026 intake. 10. Policy Options / IRPP — AI disruption and young workers (October 2025) — Argues young Canadians need targeted help gaining AI-era experience, not more education; gap in federal response identified. 11. LawNext — ABA Task Force on AI: from experiment to infrastructure (December 2025) — Coverage of the ABA Task Force on Law and AI Year 2 Report, December 2025, framing AI as legal-profession infrastructure rather than experiment. 12. Thomson Reuters Institute — The AI Law Professor: When AI forces us to rethink how we train junior lawyers — Directly frames junior-lawyer training as the central question raised by AI adoption in law firms. 13. Canadian Lawyer Magazine — Paul Saunders urges training overhaul (April 2026) — Op-ed series on the apprenticeship model being hollowed out. Companion piece: AI risks hollowing out the law firm apprenticeship model (May 2026). 14. Goodwin Procter — Next-Generation Training Models (December 2025) — Announcement of Propel (firm-wide AI training) and the separate First Year Development Program (five-week pre-billable bootcamp). Profile of both programmes: Law.com — Inside Goodwin's ambitious AI training program (April 2026). 15. ICAEW — Next Generation ACA qualification to launch in 2025 — Announced October 2024, launched September 2025; described as the biggest change to the ACA in three decades, with AI-era practice integrated through the work-experience component. 16. ICAEW — How Deloitte is reshaping its audit training for the AI age (March 2026) — Deloitte's three-year graduate training programme explicitly redesigned because AI does the confirmation letters and data prep that trainees used to learn on. 17. Journal of Accountancy — How will accountants learn new skills when AI does the work? (March 2026) — AICPA publication directly addressing the training-pipeline question. See also AICPA's "Profession Ready" initiative on early-career skills. 18. Innovaiden — The Consulting Pyramid Is Broken — Names four emerging alternatives to the traditional consulting pyramid (diamond, obelisk, box, inverted pyramid). The "box" model is associated with Alvarez & Marsal's structural shift. Companion read: Consultancy.uk — AI may up-end the consulting pyramid. 19. Business Standard — McKinsey to hire 12% more junior employees in North America (September 2025) — Eric Kutcher (McKinsey North American chair) committing to 12% more graduate hires in North America in 2026 than 2025. Bloomberg's paywalled coverage (April 2026) provides broader context on McKinsey/BCG/Bain entry-level hiring shifts. Also Irish Times — Top consultancies freeze starting salaries as AI threatens pyramid model (December 2025). 20. Reuters Institute for the Study of Journalism — AI and the Future of News — Research programme; first dedicated AI-and-journalism conference held March 2025, second in March 2026. 21. Poynter — AI Innovation Lab — Launched October 2025; trained 1,000+ working journalists in AI tooling and ethics. Skills focus, not entry-level pipeline reform. 22. Columbia Journalism School — CJS2030 AI Initiative — Required AI-ethics course revamp; CS+Journalism dual masters track housed at the Tow Center for Digital Journalism.