13% drop in employment for young workers in AI-exposed roles. update on the new ai automation paper from stanford. yes, I was right. confirmation bias caution.

I don't know what to put here, ask claude for subtitle before publishing [DON'T FORGET AS IT WILL LOOK AMATEUR HOUR]

Incoming. Last week we got more data, for those of us that like to hang some data on our opinions.

• In Young People Can’t Get Jobs, Now What, I argued that the first rung of the white-collar ladder was collapsing because AI

• In Unbundling the Job, I explained why this matters: the job isn’t just wages, it’s a bundle — training, identity, security, mobility. If young people don’t get jobs, then…

• But Uh oh, what’s this? conflicting data. In AI and the Missing Junior Automation, I reconciled early conflicting data from the Economic Innovation Group (EIG) who found that young people were not actually losing their jobs.

• And now this…

---

1. what happened

Stanford economists (credentials!) Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen analysed payroll data from ADP, which tracks millions of US workers monthly. By combining age and job-level AI exposure, they’ve given us the clearest picture yet of AI’s impact on work. They say the following things:

• Among workers aged 22–25 in highly AI-exposed roles — software developers, customer support, junior auditors, analysts — employment has fallen 13% since late 2022, around the launch of ChatGPT.

• Older workers in the same roles saw little to no change.

• In non-AI-exposed roles — healthcare aides, trades, care work — employment for young workers rose.

They say something very important: this isn’t an economy-wide impact. In fact, even more interestingly, this isn’t a company-wide impact; it’s department specifc.

Now, confirmation bias alert: this is what we would have expected when we looked around us so be careful with this sort of data.

Derek Thompson summarises this well in his latest piece. After months of contradictory narratives — “maybe yes,” “definitely yes,” “almost certainly no,” “maybe yes again” — the ADP dataset resolves the tension: we weren’t looking at the right slice of the labour force. Aggregate job numbers are fine because new grads are a small share of the workforce. As I said in my last piece.

This lines up almost perfectly with what we’ve been tracking here since May:

• The first effects are cohort-specific. Juniors, interns, and graduates, not mid-career professionals, are hit first.

• Task substitution comes before job substitution. Firms keep headcount stable while quietly replacing entry-level tasks with AI.

• The pipelines are shrinking. Internships, analyst programs, and apprenticeships — the traditional “learning by doing” pathways — are evaporating.

2. why this matters

Well it matters because it’s data to support my thesis, which is nice. But beyond me being right, does it matter? Yea I think the paper makes some useful points, see a,b,c, below.

a) Task substitution is real

The paper distinguishes between automative and augmentative uses of AI.

• Automative roles — where tasks are codified and discrete, like formatting a document, answering a customer query, or writing boilerplate code — are being hollowed out first.

• Augmentative roles — requiring judgement, tacit knowledge, or physical interaction — are holding steady for now.

This validates the framework from Unbundling the Job: AI isn’t killing jobs wholesale yet, but it’s stripping out the “glue work” juniors once learned through repetition and exposure. The apprenticeship layer is dissolving.

b) Cohort masking hid the effect

The earlier EIG analysis suggested AI’s impact wasn’t showing up in aggregate unemployment data. That was true. But misleading. By grouping all AI-exposed workers together, the losses for young entrants were hidden under the stability of incumbent mid-career workers. Classic.

We now have the missing piece that connects our leading indicators:

• Revelio: 35% fewer entry-level postings since Jan 2023

• Adzuna: ~30% fewer UK internships since 2022

• NACE: net intern hiring down 3.1% in the US this year

• Surveys: 70% of employers say AI can already replace an intern

c) The lag is shrinking

The Stanford data is based on employment changes through mid-2025 — meaning the 13% drop already happened. Employers saw GPT-4’s capabilities in 2023, paused graduate schemes, and the effect hit payrolls in 2024–2025.

Now we’re heading into a new model cycle — GPT5, Gemini 3.5, Claude 4 — all improving in agentic multi-step, multi-hour workflows. If GPT-4 triggered the first 13% decline, the next wave of models will deepen the effect. I’m not gonna show you the chart again. You don’t need to see it. Picture it in your mind. A straight line up and to the right. Agents doing longer tasks. Each new model climbing higher and higher into our future. Deeper and deeper into the human job market.

Like, I’m not gonna call it from one set of confirmatory data. But also, like we all have eyes. I am no longer going to have the discussion if AI will effect junior white-collar work. Let’s discuss how fast, how far, and how deep.

Come on, now. I’m absolutely bouncing over here.

3. the big picture

some a’s,b’s and c’s for you.

a) The job bundle is breaking (see: Unbundling the Job)

The job has always been more than wages — it bundled training, progression, security, and identity into a single institutional package. Employers absorbed volatility on behalf of workers: you got a payslip, a pension, healthcare, and an apprenticeship path in exchange for loyalty.

AI is eroding the bottom of that system. When discrete tasks are automated, firms reduce entry-level headcount. When fewer juniors enter, fewer mid-levels get promoted. Over time, entire career ladders flatten. This isn’t a temporary “skills mismatch.” It’s a structural unbundling. If you are a policymaker reading this, and I see the email domains, so I know there are a few: putting money into “retraining” isn’t going to fix this. And my god, if you annouce some program to help young people “use AI”, I swear I will give up.

b) Labour markets will polarise (see: Dirty Work)

The Stanford data reinforces a pattern we’ve been writing about:

• AI-exposed, codifiable work = declining junior openings, slower career mobility

• AI-resistant, physical or tacit work = trades, care, field services gain demand

Again, we’re producing more graduates trained for knowledge work at the exact moment the on-ramps to those jobs are disappearing. Without intervention, we’ll force a generation into underemployment or into sectors they weren’t trained for, without the prestige, pay, or support infrastructure to make that transition attractive.

The answer: blue collar work. green transition. re-shoring. industry 5.0. call it what you want, but we need to get young people off screens, and into the trades asap as soon as possible.

What’s good for the goose, is good for the gander as they say.

c) We need a new lens: think tasks, not jobs

One reason this debate has been noisy is because we keep talking about jobs — binary “safe” vs “replaced” framings. That’s the wrong abstraction. A “job” is a bundle of tasks, and AI is eating through them unevenly.

People often say:

“An agent can’t do my job.”

That may be true. But it misses the point.

What percentage of my job’s tasks can an agent already do today?

What percentage will it be able to do in two, three, five years?

These are good questions.

The real inflection point isn’t when AI “replaces your job.” It’s when AI systems reliably complete 50%+ of the tasks inside the role, at lower cost, higher speed, and comparable quality. Well I don’t know exactly if that’s the inflection point, it might differ for every role? maybe the percentage is higher for jobs in regulated industries? Maybe the last remaining 10% of “client-facing work” is where the real value always was and now you do 50% more that?

I don’t know, but I do know, we would have a better debate if we talked in terms of “bundle of tasks that make up my job title” rather than “my job”.

This is where the conversation should go next: mapping task coverage curves rather than arguing whether AI can “replace jobs.” The Stanford paper gives us evidence of effects; the next step is measuring trajectories.

4. electric motors not steam power, tasks not jobs

I think this shift from jobs to tasks is akin to the shift from steam engine to electric motor in the industrial revolution. Electric motors were commercially available from the 1880s, but factories didn't see major productivity improvements until the 1910s and 1920s. 30-40 years of people going: where all that productivity we were promised?

The breakthrough came when engineers like Henry Ford realized they could redesign everything around electric power. Instead of tall, narrow buildings constrained by belt systems, they built sprawling single-story factories. Instead of workers clustered around power transmission points, they could distribute workstations optimally for material flow. The assembly line became possible precisely because electric motors freed manufacturing from the geometric constraints of steam power.

We're clearly in that transitional period now with AI and task automation. Most companies are using AI to speed up existing job functions - having ChatGPT help analysts write reports faster, or using AI to screen resumes more quickly. They're essentially putting electric motors on their old belt-driven systems.

It’s not obvious what a “task-first” organisation might look like. It will surely have fewer humans and more agents. But I am no Henry Ford. I don’t have a good sense of how that might look. It seems to rhyme with Amazon’s and the focus on external APIs and written communications. Figuring this out seems like a very valuable activity. If I am McKinsey, my next pitch is: “becoming a task-first organsation”. You’re welcome.

5. the bottom line

In Young People Can’t Get Jobs, I argued that the squeeze was coming. In AI and the Missing Junior Automation, I showed the early cracks. In Unbundling the Job, i explained why entry-level erosion matters far beyond first jobs.

Now, the Stanford paper confirms the signal:

• 13% drop in employment for young workers in AI-exposed roles

• Collapse concentrated in codifiable, automatable workflows

• Effects hidden in aggregate data but sharp at the cohort level

The ball I am running with at the moment is thinking in tasks-not-jobs.

---

In order to grow my audience and get more engagement, let me know what you think in the comments?

Actually, you know what, don’t. I don’t want your engagement. I want you money. Subscribe and pay me.

Bye.