Exposure Is Not Replacement: What the Evidence Actually Says About AI and Jobs
- Younhee Shin
- Jul 6
- 3 min read
A new report from the Institute for Machine Learning and the AI Institute @ St. John's University
"Will AI take the jobs?" is the question everyone asks. After spending several months reading the actual evidence — studies from OpenAI, the National Bureau of Economic Research, the IMF, the OECD, the Bureau of Labor Statistics, and the latest labor-market data — we think it's the wrong one.
The honest picture is more specific and useful than either panic or hype.
One word does most of the damage
About 80% of U.S. workers have at least 10% of their tasks exposed to large language models. That figure gets repeated as if it means 80% of jobs are disappearing. It doesn't.
Exposed means AI can touch the work. Whether it replaces, assists, cheapens, or upgrades that work is not decided by the technology. It's decided by institutions, business models, regulation, and the choices managers make. Exposure is a starting condition, not a verdict.
Once you hold that distinction steady, the evidence tells a coherent story.
What the evidence shows
AI is changing the structure of jobs before it changes the number of jobs. The clearest early effect isn't mass unemployment; it's a rearrangement of which tasks within a role carry value, which skills become the baseline, and what beginners are allowed to learn.
Augmentation is showing up before automation. A large NBER study of customer-support agents found that generative AI raised productivity by about 14% on average, with the largest gains going to the least experienced workers. AI can compress the learning curve. But the same productivity gain can be spent two ways: better service and higher wages, or fewer people doing the same work. The technology doesn't choose.
Layoffs are real — but AI is only part of the story. Layoff announcements citing AI have climbed sharply, and AI has now led all cited reasons for U.S. job cuts for four consecutive months. Yet those cuts sit inside a wider picture: post-pandemic over-hiring, higher interest rates, investor pressure for profitability, and heavy spending on AI infrastructure. A company citing AI is not proof that a model replaced each worker, and many of these cuts are happening at profitable firms expanding their AI budgets, not collapsing ones.
Technical work is being redefined, not eliminated. The BLS projects strong growth for software developers, QA analysts, and testers through 2034. Routine coding, basic testing, and documentation face the most pressure, while systems thinking, cybersecurity, data engineering, and human judgment gain value.
The risk almost nobody is pricing in
The danger isn't a jobpocalypse. It's quieter than that.
Careers are built on beginner work — drafting, testing, researching, documenting, and routine analysis. Those tasks aren't just low-value labor to be optimized away; they're how people develop judgment. If AI absorbs the bottom rung of the ladder, firms may hire fewer beginners or expect new graduates to arrive already expert. Nobody becomes an expert by prompting a model. They become experts through practice, and that practice is exactly what's most exposed.
That's an apprenticeship problem, not just a jobs problem, and it's the one that most rewards attention now.
The real question
So the better question isn't whether AI will take the jobs. It's the one this report is built around:
What kind of labor market are we choosing to build around AI?
The answer isn't set by the technology. It's set by employers who redesign work honestly, universities that strengthen the path from learning to work, and workers who learn not just to use these tools but to question, evaluate, and govern them.
Read the report
The full report and a two-page policy brief are available free to read and share under a Creative Commons license (CC BY 4.0). No sign-up, no gate — it's meant to be used, cited, and built on.
If you found this useful and want to go deeper on building the AI judgment this moment demands, that's the subject of my book, Generative AI for Business Professionals: From Understanding to Action — but the report above stands entirely on its own.