Will AI Replace My Job? What the Data Actually Says (Not the Headlines)

TL;DR

Anthropic published a study with a new metric: “observed exposure,” combining LLM theoretical capability with real usage data
Most exposed jobs: computer programmers (75%), customer service reps, data entry operators
30% of workers have zero exposure: cooks, mechanics, electricians, plumbers
No evidence of increased unemployment in the most exposed occupations — yet
But there is tentative evidence of something concrete: hiring of young workers (22-25) in exposed jobs dropped 14% since ChatGPT launched
The most exposed jobs tend to be better paid, higher education, and held predominantly by women

There are two types of headlines about AI and employment.

First type: “AI will destroy millions of jobs.” Second type: “AI will create more jobs than it destroys.” Both have been circulating for years. Neither comes with data worth taking seriously.

This week Anthropic published something different: a rigorous academic study, with new methodology, based on real usage data. Not predictions. Observations.

The results are more nuanced than either side of the debate wants to admit.

The problem with previous studies

Most analyses of AI and employment share the same flaw: they measure theoretical exposure.

Take an LLM, ask it questions about work tasks, and if it answers well, mark that task as “exposed.” The problem is that a model that can do something and a model that is being used to do it are two very different things.

It’s like measuring the impact of cars on cities by calculating how many routes a car could theoretically drive, without checking if there are actually cars on the road.

The metric that changes the analysis: “observed exposure”

Anthropic introduced a new metric that combines two things:

Theoretical capability: what percentage of a job’s tasks an LLM can cover
Real usage: what Claude users are actually doing in their conversations

The result is “observed exposure” — not what AI could do, but what it is doing in a measurable way.

This distinction matters a lot. A job might have high theoretical automation potential but low real exposure if people in that role aren’t using AI significantly yet. And vice versa.

The concrete numbers

Most exposed jobs

Occupation	Observed exposure
Computer programmers	75%
Customer service representatives	High
Data entry operators	High
Financial analysts	High
Writers and editors	High

The 75% exposure figure for programmers doesn’t mean AI is doing 75% of their work. It means 75% of their tasks are the type that LLMs are already covering in real usage. There’s a difference.

The 30% with zero exposure

This data point gets less attention but is equally important: 30% of workers have essentially zero exposure.

Cooks. Mechanics. Electricians. Plumbers. Carpenters. Floor nurses. Warehouse workers. Drivers.

These jobs require physical presence, manual skills, or contextual human interaction that current LLMs can’t cover. It’s not that AI won’t eventually reach them — it’s that there’s a structural difference between what these jobs involve and what language models do.

What’s not happening yet

Here’s the part the alarmists don’t want to hear: there’s no evidence of increased unemployment in the most exposed occupations.

Programmers are still working. Financial analysts too. AI has increased their productivity in many cases, but it hasn’t replaced them en masse.

This isn’t surprising if you understand how labor markets work. Transitions are slow. Contracts, organizational structures, training, regulations — all of these act as buffers between “technology can do this” and “technology is doing this at scale.”

What is happening

Here’s the part the optimists don’t want to hear: there is tentative evidence of an effect on young worker hiring.

Since ChatGPT launched, hiring of 22-25-year-old workers in highly exposed occupations has dropped approximately 14%.

This isn’t unemployment. It’s something subtler: companies are hiring fewer junior profiles in the most automatable roles, probably because senior profiles with AI can cover some of that work.

It’s the difference between “AI replaces workers” and “AI changes who companies hire.” The second effect is real and already happening.

The profile of the most exposed jobs

There’s an interesting pattern in the occupations with the highest observed exposure:

Better pay than average
Higher education required
Predominantly held by women

That last point deserves attention. Not because it’s an inevitable outcome, but because if AI’s impact concentrates in well-paid jobs held by women, the gender implications are significant and largely underdiscussed.

This includes roles like writing and editing, skilled administrative work, data analysis, and parts of legal and accounting work.

What this means if you work in data or tech

If you work in analytics, programming, or any data role, your occupation is in the high-exposure group. That’s a fact.

What’s not a fact is what it means for you personally. There are two ways to be in a high-exposure job:

As someone who uses AI: Your productivity goes up, you produce more, your value as a professional increases. The exposure works in your favor.

As someone who doesn’t use it: You produce the same as before, but the people around you are producing more with AI. The exposure works against you, slowly.

The real risk isn’t AI taking your job tomorrow. It’s falling behind while the bar for what’s expected of a technical professional keeps rising.

This is the same pattern we see in how vibe coding is transforming software development: the tool amplifies those who use it well and leaves behind those who don’t adopt it.

Why both sides are wrong

The alarmists have a timing problem. They’ve been predicting mass employment collapse for years and it hasn’t arrived. Labor markets are complex, transitions are slow, and real AI adoption in work processes is well below its adoption in consumption.

The optimists have a granularity problem. “AI creates jobs” is true at a macro level over 20 years. But at a micro level, over 3-5 years, some profiles will see reduced demand. Juniors in highly automatable roles are already experiencing it.

The truth is in the data, and the data says: real but slow impact, uneven across sectors and profiles, with effects already visible in hiring even if not yet in unemployment numbers.

What the data still can’t tell us

Anthropic’s study is the most rigorous to date, but it has limitations the authors themselves acknowledge.

Claude usage data isn’t representative of the entire labor market. Claude is a tool, and its user base has biases — probably more technical, more educated, more English-speaking than the average worker.

We also can’t yet know what new jobs AI will create. Every significant technology revolution has created jobs that didn’t exist before. That will probably happen again. But what jobs, with what distribution, and requiring what training — that’s still speculation.

The bottom line

Will AI replace your job? The honest answer is: not massively yet, but it’s already changing who gets hired and what profile companies look for.

The clearest effect so far is the 14% drop in junior hiring in exposed roles. It’s not catastrophic. It’s also not irrelevant if you’re 23 and looking for your first job in tech.

What is clear: professionals who use AI well will be in a better position than those who don’t. Not because AI protects them from change, but because change will separate those who adapt from those who don’t.

That’s happened with every significant technology of the last hundred years. There’s no reason to think this time will be different.

Keep exploring

Are AI Tools Worth Paying For? — Which tools have real ROI for technical professionals
Vibe Coding: What It Is and Why 90% Do It Wrong — The concrete transformation of development work
AI Won’t Take Your Job (But This Will) — The analysis on automation and technical roles
Why 95% See No Results with AI — Why real adoption is well below the hype