AI and Jobs: Assertive Headlines, Fanciful Data
Why the evidence still cannot tell us whether generative AI is responsible for today’s labor-market shifts.Photo by The New York Public Library on UnsplashA few months ago I happened to voice some criticism of Professor Floridi’s piece “Not Even Wrong: Forecasts of AI Impact on the Labour Market,…
Why the evidence still cannot tell us whether generative AI is responsible for today’s labor-market shifts.Photo by The New York Public Library on UnsplashA few months ago I happened to voice some criticism of Professor Floridi’s piece “Not Even Wrong: Forecasts of AI Impact on the Labour Market, from Frey & Osborne to ChatGPT 2013–2026.”Actually, the criticism wasn’t aimed at the piece itself, which frankly I hadn’t read, but at a summary that seemed to conclude in an overly nihilistic way: let’s admit we know nothing and be done with it.The disagreement was more apparent than real: I argued that the weakness of the available information couldn’t excuse us from trying, especially for policymakers, who in any case couldn’t avoid forming some opinion. Which was exactly the thesis of the article, even if perhaps not of the summary I’d stopped at, which simply called for more honesty, replicability, and rigor instead of fantasy in power.Since then things haven’t improved, and imagination is still firmly in charge.Now we’ve been enriched with a strand of work that would like to go beyond forecasting into analysis of actual outcomes, since at least three years have now passed since the advent of generative AI: we’re trying to see whether this advent has actually produced something in the labour market. Unfortunately, even here, I don’t see any more solidity in the ex-post analyses than what was (not) found in the forecasts.And the reason is, in a sense, very simple: we still don’t have sufficiently robust data on firms that have actually adopted generative AI. Maybe because firms have only really started adopting it “today,” metaphorically speaking, in particular since generative AI improved qualitatively at coding.And coding matters here, it’s worth stressing, not just because AI that can code can replace a programmer, but because AI that can code can do things, like calculations and data analysis, that it previously couldn’t do, which has made it genuinely more useful than before.Frankly, I believe that it’s only been a couple of years since enterprises have started to take into real consideration the hypothesis of using Gen AI for real tasks in the real world, and even less time since they started actually using it in practice.So, without adoption data, or rather, with adoption data still too recent for any measurable effect to have actually been produced, we inevitably fall back into forecasting. And the merry-go-round starts again: exposed occupations and how exposed, substitutable vs. complementary tasks, vacancy analysis as a proxy for future occupations, and so on. And above all, a useful scapegoat for whatever has actually happened in the labour market since 2022 onward.Let’s take, for example, the vacancy data. Which, together with the categories we mentioned regarding occupations more or less exposed to AI, have been, and still are, a workhorse for growing strands of researchers. And let’s have a look at some data from the UK labour market (you’ll be wondering why I’m talking about the UK, but the answer is simple: it’s the only European country whose language I understand that has deep, public, available data).Figure 1Figure 2You’ll immediately see that, unlike the past (Figure 1) — where there was a fairly consistent correlation between movements in vacancies and movements in employment, this correlation no longer seems to hold (Figure 2) from 2022 onward (let’s say from the height of the post-pandemic period on), which unfortunately is exactly the period that everyone considers relevant for analyzing the impact of AI on jobs, given that ChatGPT’s spread worldwide dates back to November 2022.We’re seeing an overall collapse in vacancies accompanied by an increase in both the number of jobs and the number of people employed. Granted, the post-2022 period is still short, and this correlation might reassert itself with a lag, etc., but for now it’s absent, whereas in the past it was quite precise and fairly timely, which explains why economists love it so much. And so one has to wonder: why rely so readily on proxies like vacancies and job postings if their reliability as leading indicators of future employment is showing cracks?Another point: since we lack robust data on adoption by firms, we fall back on the only categories we’ve been working with for years: exposed occupations, and try to understand what’s happened to employment within these categories. Too bad there’s a fair amount of variety in the rankings of exposed occupations, and above all an interesting variety of theoretical constructs distinguishing exposure-implying-substitutability from exposure-implying-complementarity.From the variability of these constructs follows a variability of results, which nonetheless seem to converge on one single result: young people (the very young) are being employed less than in the past, unlike more mature workers.And here comes a set of philosophical musings, including the ones I make myself, about the value of judgment that resists the commoditization of techne (of procedural competence), which supposedly explains the resilience of older workers and the decline of young people in sectors and occupations most exposed to AI (exposed according to various metrics).But this fails to account for the fact that (a) the rise in older people working is by now a decade-long phenomenon driven by the fact that no one wants to pay pensions, i.e., rents, to anyone who still has even a shred of energy left to work, a phenomenon entirely separate from generative AI, which is a much more recent thing.As for the decline in youth employment, it’s a real and relatively recent phenomenon (let’s say post-pandemic), affecting a highly disparate range of occupations, including some that would be very hard to link to generative AI, and it deserves to be investigated in its own right, not as a secondary byproduct of the question “what is AI doing to jobs.”For instance, at least two competing hypotheses come to mind: effects on youth education and socialization stemming from the pandemic and its lockdowns (including school closures), or from phenomena like social media, whose harmful degenerations we are only now finally starting to sanction, see the Meta case; and wages offered so low they fall below the “resistance threshold” (in Italy, say, job offers for pharmaceutical assistants, after a difficult five-year degree, offer 22,000 euros a year, if that seems normal to you), such that young people prefer not to work (because working is tiring, let’s not forget).And then a third hypothesis, which holds at least for wealthy countries (and certainly for the UK): the increase in the stock of young people due to migration inflows exceeding what the economy can actually absorb. In the UK, for example, inflows into the labour market are actually growing (mostly), the problem is they need to be measured against a pool that’s growing even faster.Maybe Daron Acemoglu is actually right when, alone among contemporary economists, he argues that laments over falling birth rates might not be such a huge problem for humanity (though perhaps only for capitalists who want their reserve industrial army; that addition is mine, not Acemoglu’s).Okay, I’m aware I’m extending to the whole world conclusions that, for now, only the UK data actually license — things might look different, say, in the US (I don’t know US statistics well enough to verify this, but I can look into it…), or in the rest of Europe. But what I’d really like is for everyone who uses a given measure or method to also explain why they consider that measure or method significant. And if it is, they should say clearly under what conditions and in which places it actually works.In any case, before we hand generative AI the blame for whatever is happening to vacancies, wages, or the young, it might be worth checking whether the crime scene actually has its fingerprints on it, or whether we’ve simply arrested the most convenient suspect.Sources:Floridi, L., Novelli, C., & Morley, J. (2026). Not even wrong 1: AI and the labour market, from Frey–Osborne to ChatGPT, 2012–2026 (Revised version)Daron Acemoglu, David Autor, Keelan Beirne, Andrew Scott, “Baby Busts and Growth Booms: Demographic Change and the Macroeconomy” (NBER Working Paper, luglio 2026)This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!AI and Jobs: Assertive Headlines, Fanciful Data was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI