The mid-career squeeze

Fifteen years in, the fear is usually some version of the same sentence: I am expensive, I am replaceable, and I am too far along to start again. The evidence says that fear is pointed at the wrong thing. A mid-career squeeze does exist. It simply is not the one people are bracing for.

The displacement is happening somewhere else#

Start with what the payroll data shows, because it is the least speculative thing here. Brynjolfsson, Chandar and Chen, using ADP microdata covering millions of US workers, compared employment by age and by occupational AI exposure since ChatGPT’s release. Two findings matter.

• No widespread economy-wide displacement. The authors rule it out explicitly on current evidence.
• Employment among 22 to 25 year olds in highly exposed occupations sits about 19 per cent below where it would be had it tracked similarly aged workers in less exposed occupations. The divergence runs through reduced hiring rather than increased separations, and it concentrates where AI substitutes for human tasks rather than where it complements them.

Read that second point carefully if you are forty. The measurable damage is at the entry gate, not in the middle, and it shows up as doors not opening rather than people being pushed out.

History points the same way. Dauth and colleagues, using German administrative worker and plant data from 1994 to 2014, found that incumbent workers facing robot exposure largely kept their jobs and moved into new, higher-quality tasks inside their original plants. The cost fell on young entrants, who moved away from vocational manufacturing training altogether. Different technology, with the authors careful to note that generative AI need not behave like industrial robotics. But the pattern is worth knowing before you panic: in the closest large-scale analogue we have, being already inside was a considerable advantage.

The squeeze is on the premium, not the post#

Here is the part that gets less attention and deserves more. What a mid-career professional actually sells is a gap: the distance between what they can do and what somebody five years in can do. That gap is what the salary is for. And the evidence on where AI’s gains land is consistent across very different kinds of work.

• Customer support. Brynjolfsson, Li and Raymond studied a staggered rollout across 5,172 support agents. Resolutions per hour rose about 15 per cent on average, but the least experienced gained around 30 per cent, rising to 36 per cent for the lowest skill quintile, while the most skilled saw no significant gain at all.
• Driving a taxi. Kanazawa and colleagues followed a Japanese fleet through an AI demand-prediction rollout. Productivity gains accrued almost entirely to low-skilled drivers, narrowing the gap between best and worst by 14 per cent.

One is knowledge work in an office, the other is manual frontline work in a car. Both compress the distribution from the bottom. Nobody was made worse off in absolute terms, so this rarely registers as a threat, and the experienced worker’s relative advantage narrowed anyway. You do not have to get worse to be worth less.

Three reasons this is hard to see from the inside#

If it were obvious, people would have adjusted. Three findings explain why it is not.

You cannot tell whether the tool is helping you. METR ran a randomised controlled trial with sixteen experienced open-source developers across 246 real tasks on mature repositories they knew well, randomising whether AI tools were permitted. The developers were measured as 19 per cent slower when allowed to use AI. They had forecast a 24 per cent speed-up. Afterwards, having actually done the work and experienced the slowdown, they still estimated AI had made them roughly 20 per cent faster. Sixteen people is a small study, and the authors say so plainly: nothing here generalises to all developers or all software work. What it does establish is that the self-report can be wrong in the opposite direction to the truth, by a wide margin, in the population that feels most confident.

Seniority does not predict who benefits. Yu and colleagues randomised AI assistance across 140 radiologists on roughly 5,190 observations. The effect ranged from strongly positive to strongly negative between individuals, and experience, subspecialty and prior familiarity with AI all failed to predict which. Lower performers did not reliably gain either. Whatever determines this, it is not years served.

Experienced professionals do deskill, and quickly. Budzyn and colleagues looked at 1,443 unassisted colonoscopies performed by 19 endoscopists averaging 27.6 years of experience, before and after AI was introduced at four Polish centres. Adenoma detection in the unassisted procedures fell from 28.4 to 22.4 per cent, six percentage points, within months. Observational rather than randomised, covering one procedure in one country. It is also the closest thing we have to a direct measurement of an expert getting worse at the thing they are expert in.

Am I too senior to retrain and too junior to be safe?#

The question assumes the danger is being let go, and on the payroll evidence that is the least likely outcome for someone mid-career in the near term. So the framing is wrong rather than the worry being silly.

The real exposure is slower and less dramatic: your differential erodes while your title does not, and the first visible sign is not redundancy but a quiet change in what the market will pay for what you do. Retraining aimed at keeping the job is aimed at a risk that is largely not materialising. Work aimed at protecting the gap is aimed at the one that is.

How to analyse your own job, in one pass#

Autor and Thompson give the sharpest available lens. Across four decades of task data covering 303 US occupations, automation that removed the less expert tasks in a job raised wages and reduced employment, while automation that removed the expert tasks lowered wages and increased employment. Their data ends in 2018, so treat this as a question to ask rather than a forecast.

Applied to yourself, it is one question with an uncomfortable answer: of the tasks AI is taking from your week, are they the ones that made you expensive, or the ones that were merely time-consuming? If the tool is removing the routine perimeter and leaving you the judgement, your role is appreciating. If it is doing the diagnosis, the drafting or the call and leaving you to check its work, it is commoditising, and no amount of enthusiasm about the tool changes which of those is happening.

What actually protects a mid-career position#

• Keep some work unaided, deliberately, as measurement. Not as a principle or a protest. If everything you produce is assisted, you have no way of knowing whether the gap you are paid for still exists, and the METR result says you will guess wrong about it.
• Move towards the tasks that are hard to specify. The tools are strongest where the task can be stated clearly. Ambiguous problems, contested priorities, situations where the hard part is working out what is actually being asked, are where a decade of context still compounds.
• Take the accountability nobody else wants. Somebody has to stand behind a decision that a model contributed to. That responsibility is currently badly distributed and rarely priced. No tool absorbs it.
• Get in front of juniors. Not altruism. The Canaries finding says hiring at 22 to 25 is where the damage is concentrated, which means the people who can build judgement in others become scarcer and more valuable at the same moment that the pipeline thins.

Notice what is not on that list. Learning the tools is worth doing without being a moat, for the reason set out in why learn to prompt is weak career advice: the skill is neither scarce nor durable. PwC’s barometer, analysing close to a billion job advertisements, reports a 56 per cent wage premium for AI skills. Read it for what it is. Advertisements are stated employer demand rather than realised pay, and PwC sells AI services, so it is a signal of what firms are asking for rather than evidence of what they end up paying.

What if my employer makes me use it?#

Mandated adoption is common and mostly reasonable. Two things are worth doing rather than resisting outright.

First, ask what is being measured. If the reporting is adoption or hours saved, nobody is watching the thing that affects you. Asking in writing how the organisation intends to know whether people can still work unaided is a fair question. Second, negotiate for the unaided reps rather than against the tool. That is a request an employer can grant, it costs little, and no other version of this conversation avoids sounding like refusal.

On refusing outright, very little direct evidence exists about what happens to people who decline. The METR result complicates the assumption that refusers are simply forgoing productivity, since the experienced developers in that trial were faster without the tools. But sixteen people on mature codebases is not a basis for a career strategy, and the social and organisational costs of visible refusal are real and unmeasured. It is one of the clearer gaps in this literature.

What about older workers specifically?#

This one deserves a straight answer rather than a confident one. The strong age-stratified evidence that exists is about the young: the Canaries analysis measures 22 to 25 year olds, and Dauth measures entrants. Nothing comparable has been published on workers in their fifties and sixties in AI-exposed occupations.

What can be said is narrower. Yu found experience did not predict who benefits from AI assistance, which removes one common assumption in both directions: older workers are not automatically disadvantaged by unfamiliarity, and not automatically protected by expertise. Everything else circulating on this subject is inference, presented with more confidence than the evidence carries.

Where this sits in my own argument#

I have written at length about what happens to people entering a profession, in the missing rungs and synthetic seniority. This page is the same mechanism one career stage later, and my position on it is narrower than the panic and less comfortable than the reassurance.

What a mid-career professional sells is a differential. The seven capabilities I set out in SuperSkills are an attempt to name what stays scarce when the differential compresses, and the honest test of any of them is the one on this page: can you still do it with the tool switched off, and when did anybody last check.

What I have observed in organisations#

Many of the teams I used to work with had EAs who held the administrative spine of how the team ran: the meetings, the all-hands, assembling the material, taking the notes, assigning the actions. Much of that has gone. The work has not gone, and other people are doing it now, usually alongside their own jobs.

The clearer version is in marketing. I have seen teams who hand-built everything in customer acquisition, from the Google Ads through to the creative. That craft was the differential and it was hard-won. It is now substantially available to somebody with a subscription and no comparable experience.

Neither group was made redundant, which is the point. The job stayed and the scarcity left.

What this page does not claim#

It does not claim mid-career workers are safe. It claims the measured displacement is currently concentrated at entry, which is a statement about the last three years rather than the next ten, and hiring patterns can move.

It does not claim AI makes experienced people worse. Brynjolfsson found no significant gain for the most skilled, which is not the same as harm. Budzyn found decline in one procedure. Yu found the effect goes both ways and nothing predicts which. Summarised fairly: the effect on experts is real, individual and not yet predictable, and anyone telling you otherwise is ahead of the evidence.

And it does not offer a five-year plan. The mechanism here is compression of a differential, which is slow, hard to observe and specific to your domain. What the evidence supports is a way of checking, not a destination.