Is AI making workers more vulnerable?

金融时报（科技），金融时报科技版块 - 来自金融时报的最新科技新闻与分析。关注

This article is an on-site version of our The AI Shift newsletter. Premium subscribers can sign up to get the newsletter delivered every Thursday. Standard subscribers can upgrade to Premium here, or explore all FT newsletters

Welcome back to The AI Shift, our weekly dive into how AI is changing jobs and work. Sarah is away this week so it’s a solo edition from me split over two sections. First up, in light of a flurry of new research I wanted to return to the question that kicked off this newsletter ten months ago: where are all the AI-driven job losses? I then have some thoughts on a fascinating recent study on how AI-use is warping education, which found the effects vary widely between individuals.

Is AI taking jobs yet? Q3 2026 edition

First up, a brief recap of where the AI job loss debate currently stands:

Official labour market data shows little to no evidence of a clear link between an occupation’s AI exposure and its employment growth at the economy-wide level
But analyses of very fine-grained administrative data do find weaker employment growth in the very most AI-exposed occupations such as software and customer services
The most consistent finding across different analyses and data sources is that where we do see any indication of displacement it’s concentrated among the most junior roles

Official labour market data shows little to no evidence of a clear link between an occupation’s AI exposure and its employment growth at the economy-wide level

But analyses of very fine-grained administrative data do find weaker employment growth in the very most AI-exposed occupations such as software and customer services

The most consistent finding across different analyses and data sources is that where we do see any indication of displacement it’s concentrated among the most junior roles

Against this backdrop, new research published a few days ago by a team of European researchers is a very valuable addition, applying the same analysis to both fine-grained proprietary data and official labour market statistics, and looking beyond employment levels in the search for an AI effect.

Expanding on an initial piece of work they carried out last year, José Azar, Mireia Giné and Javier Sanz-Espín have crunched data on employment levels, wages and job-to-job transitions, using both millions of job postings and professional profiles on sites like LinkedIn and (separately) the US Bureau of Labor Statistics’ flagship survey of earnings and employment in tens of thousands of US firms.

Like most others before it, the new study found no evidence that the most AI-exposed occupations have seen a decline in employment relative to less-exposed jobs and industries, but it did find a negative impact of AI exposure on wages and moves between jobs, suggesting that even if AI is not replacing workers yet, it may be starting to erode their bargaining power and broader status within the labour market.

Depending on the dataset used, the researchers found that wages for the most AI-exposed workers are down between five and 10 per cent relative to the least-exposed occupations, with the steepest declines among the most inexperienced workers in those jobs. And while there was no evidence of AI-exposed workers losing their jobs, they have become almost a third less likely to move between jobs compared to their unexposed counterparts. As the authors note, the two dynamics are likely related, given that moving between employers is a key source of wage growth.

How I’m thinking about this

For me this study acts as a potential missing link between the widespread reports that AI is leading employers to rethink whether they need as many people in certain roles and the dearth of hard evidence on actual job displacement. Weak wage growth and reduced worker mobility while employment levels hold just about steady looks a lot like what we might expect to see if AI was beginning to erode the value of certain skills and jobs, but had yet to show up in outright job losses given how much slower hiring and particularly firing tend to respond to incremental technological shifts.

The repeat of the finding that less experienced workers are the hardest hit also fits neatly with another recent study, which looked at automation risk through a new lens: whether a job mainly consists of execution or evaluation. The authors argue that AI is better at doing than evaluating, and find that execution-heavy roles have seen weaker employment growth in recent years while roles heavy on exercising judgment have performed better.

I think this is a very useful framework. Following instructions to execute narrow and distinct tasks as opposed to using accumulated expertise to evaluate outputs and make decisions maps more neatly on to junior versus senior roles within occupations than on to differences between entire occupations. Junior and senior consultants are more clearly on different sides of this divide than consultants and software developers are, for example.

AI and conscientiousness

Now for something completely different. If you’ll permit me a brief digression, something I spend a lot of time thinking about is the increasing importance of conscientiousness or self-discipline in the modern digital environment.

Highly conscientious people find effort less frustrating than others, making them less likely to procrastinate, and more resistant to distracting lures or easy but ultimately detrimental shortcuts.

The devices and technologies that are now embedded into daily life and work (think smartphones, social media, increasingly generative AI) can clearly be incredibly useful, valuable and generally beneficial if used consciously to help accomplish deeply intended goals. But they have in many cases been designed to tempt people into absent-minded and excessive or even compulsive use (‘doomscrolling’, short video consumption, upward social comparison) that often has negative impacts. In this world where powerful tools and boundless knowledge are only a fingertip away, but are competing with unprecedented and ever-present distraction, conscientiousness perhaps more than ever determines whether someone sinks or swims.

I mention this because a study from earlier this summer showed signs of exactly that divide playing out with AI use in education. The paper, led by Swedish economist David Stromberg, and covered (and exquisitely charted) last week by The Economist, found that after a group of Chinese high school students began using AI to help with their homework, marks in homework assignments improved sharply while completion time fell, but performance in subsequent closed-book offline exams cratered.

The headline finding is that using AI for schoolwork shortcuts the learning process and has damaging long-term consequences. But what I found especially interesting was that this was not a universal pattern. Where students used AI with their homework but still spent just as long on the work as students without AI (presumably they were using it more like a tutor than an answer book) they did just as well on the final exam as students who didn’t use AI at all. Whereas people who just used it to give them the answer saw the collapse in exam performance.

I do worry that this risks becoming a recurring pattern in many areas AI touches, including the world of work. While workers in some fields report that AI supercharges their output, and are evangelising about their tool set-up and the things they have produced with it, we’re also seeing reports that large numbers of workers who use AI in their job do so secretly and feel that the way they are using it is “cheating”. This sounds a lot like some are using the technology to accelerate their professional development and enhance their skills, while others are using it to shortcut the professional learning process and de-skill themselves.

Divergent use shapes divergent discourse

It strikes me that this might help explain why the prospect of AI-powered education is hugely exciting for some people but deeply unsettling for others. If you and everyone in your circles are highly motivated and enjoy challenging work (traits that I suspect are over-represented in Silicon Valley), then it would seem obvious that if students have access to AI they will use it as a Socratic tutor, pushing them hard to deepen their knowledge. If you know young people for whom schoolwork goes against the grain and the whole process is just a frustrating means to an end, the equally obvious outcome is widespread shortcutting and learning loss.

I also suspect this is another part of the AI story that differs depending on whether we’re talking about generating code or writing words. As we’ve written before, for many knowledge workers automating coding shrinks a routine part of the job that was, in a sense, acting as friction slowing down the amount of thinking and doing in the job; automating writing risks removing the thinking and doing steps altogether.

At a moment when schools, universities and workplaces are rushing to establish rules and practices on acceptable AI use in the face of rapidly changing behaviours, I think it’s important to keep in mind that getting this wrong risks ushering in a great skills divergence where some race ahead as others fall behind.

Before you go . . .

We will be speaking at the FT Weekend Festival on Saturday September 5 in London. Register here to save 10 per cent off your ticket with the code FTNewsletters.