{"slug": "the-ai-job-apocalypse-is-a-mirage", "title": "The AI Job Apocalypse Is a Mirage", "summary": "A new analysis of European employment data finds no systematic link between AI exposure and youth job losses, contradicting fears of an AI-driven job apocalypse. The study, based on exposure measures by Felten, Raj and Seamans, shows that between 2022 and 2025, the correlation between language model exposure and youth employment change is statistically indistinguishable from zero, with youth employment falling in only nine of thirteen most-exposed activities while total employment in that quintile grew by 5%. The findings follow an update to the 'Canaries in the Coal Mine?' paper by Brynjolfsson, Chandar and Chen, which confirmed a 19% shortfall for young workers in highly exposed occupations but ruled out overall displacement and could not attribute the decline to AI.", "body_md": "For three years the story has been sold as settled: generative AI is devouring entry-level work, the first rung of the career ladder is collapsing, and young people are the canaries. I, too, had become convinced that this had to be true. But at some point I asked myself: where, exactly, is the carcass?\n\nOnly a few days ago, Brynjolfsson, Chandar and Chen published an update to their landmark paper Canaries in the Coal Mine?, which had first sounded the alarm about young workers threatened by AI. While confirming a decline — young workers aged 22 – 25 in highly AI-exposed occupations stand roughly 19% below the level they would have reached if they had kept pace with same-age peers in less-exposed occupations – they nevertheless ruled out overall job displacement and, above all, were unable to show, even with their firm-level analysis, that this decline among young workers was caused by AI.\n\nAnd in Europe? What do the statistics say? After all, three years had passed since ChatGPT was released in late 2022, and since 2023 increasingly powerful versions were available. Some effect ought to have become visible by then, I thought.\n\nAt first, it seemed that one had. A rough analysis combining 2025 data on generative AI adoption with employment trends by economic activity showed that, from 2023 onwards, youth employment in sector J (information and communication, including technology and audiovisual activities ) and sector M (professional, scientific and technical services) had followed a different trajectory from that seen in the rest of the economy.\n\nBut this was too coarse a measure to be truly useful. Not to mention that I was using evidence on AI use collected in 2025 to explain earlier movements — a connection that could be made only with considerable strain.\n\nSo I dug deeper. Without access to firm-level data, I tried at least to move to more granular economic activities and, above all, to see whether a recognisable pattern linked the degree of AI penetration to employment outcomes.\n\nThat could not be done using adoption data. Official statistics do not measure AI adoption at a sufficiently granular level and, by the way, for mysterious reasons European statistics fail to cover all economic sectors and subsectors: public administration and finance are excluded…the big players, so to speak.\n\nEven more importantly, adoption recorded at company level does not necessarily correspond to AI’s actual penetration of business operations. I am sure, for example, that before 2026 many companies had no official processes affected by generative AI and would therefore have answered no to survey questions on the subject. In practice, however, they may already have been significantly affected through their employees’ own use of the technology.\n\nSo I turned to AI exposure and constructed a ranking of individual economic activities by exposure, using measures developed and validated for the US labour market by Felten, Raj and Seamans in 2023 and widely used in studies around the world.\n\nAnd what did I find? Nothing clearly related to AI, frankly.\n\nLook across sixty-three economic activities ranked by their exposure to language models. If the machines were already replacing young workers, employment losses should become larger as exposure increases. Above all, they should be greater in the most exposed subsectors and smaller or absent in the least exposed ones.\n\nThey are not.\n\nBetween 2022 and 2025, across the full distribution, the relationship between exposure to language models and the change in youth employment vanishes. The correlation is statistically indistinguishable from zero. No neat gradient. No systematic slaughter of junior jobs. The scatter plot is a flat line with noise.\n\nIt is true that youth employment fell in nine of the thirteen activities in the most exposed quintile, but it is also true that total employment in that quintile grew by 5%.\n\nAnd the pattern was not unique to the top. Employment also fell in eight of thirteen activities in the least-exposed quintile, compared with four declines in each of the three middle quintiles. Exposure therefore does not produce a simple gradient.\n\nEven if we focus only on the most exposed quintile, the results remain mixed. Publishing took a beating. A few professional and information activities lost ground. But in that same quintile, legal and accounting activities, financial auxiliaries, education and insurance hired more young people, not fewer.\n\nIf exposure mechanically destroyed junior employment, why did four of the thirteen most exposed activities still add young workers? It is true that the highest-exposure quintile remains unusual: six of its nine expanding activities still lost young workers. That concentration deserves attention, but it does not establish an AI mechanism.\n\nWhat emergest, instead, across the economy, is not AI displacement but ageing. The employment share of workers aged 50–74 increased in 56 of the 63 activities between 2022 and 2025, and in every one of the 61 activities with complete data between 2012 and 2019.\n\nBut this trend predates generative AI and would be very difficult to attribute to it.\n\nFinally, setting youth employment aside, greater exposure did not predict weaker overall employment growth. If anything, the relationship was weakly positive, but statistically indistinguishable from zero.\n\nWhere is the bloodbath the models were supposed to deliver?\n\nAs I said, even in Europe there are signs that young people have been less present in the labour market since 2022. One of the simplest explanations has been to blame AI. But alternative explanations are not exotic.\n\nFor example, between 2022 and 2025, the share of young Europeans who were not employed but were participating in education or training increased by almost one percentage point. The decline may also reflect a widening mismatch between the skills employers seek and those young people acquire through education. Firms increasingly reliant on remote work may have become more reluctant to train inexperienced workers. Or, more simply, some of the routine drafting and research tasks now associated with AI were already losing their status as distinct junior roles well before ChatGPT existed.\n\nThe technology may be accelerating this process in some places. It has not produced the broad displacement the commentary class keeps announcing. My suspicion is that skills mismatch matters, perhaps partly because COVID school closures and remote learning left lasting learning losses. But that is a question for another study.\n\nThe data do not prove that AI will be forever harmless. They do show that the simple story — that greater exposure is already producing greater unemployment, especially among young people, across the economy — has not appeared in Europe’s official statistics.\n\nIn some sectors, the first rung may be getting thinner, and AI may be one of the reasons. Or it may simply have arrived in time to take the blame for older, slower forces. How long will we keep searching for correlations that are simply being forced onto the data? Even if there were a problem with youth employment, that would be all the more reason to start investigating other causes. Without wasting any more time.\n\nSources:\n\n[Canaries in the Coal Mine? – Stanford Digital Economy Lab](https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/)\n\nFelten, Raj & Seamans — Language Modeling AIOE and AIIE Public Dataset[https://github.com/AIOE-Data/AIOE](https://github.com/AIOE-Data/AIOE)\n\n[https://doi.org/10.1093/qje/qjag027](https://doi.org/10.1093/qje/qjag027) on remote working\n\nhttps://ec.europa.eu/eurostat/databrowser/view/lfsq_egan22d/default/table?lang=en\n\nhttps://ec.europa.eu/eurostat/databrowser/view/lfsi_emp_a/default/table?lang=en\n\nhttps://ec.europa.eu/eurostat/databrowser/view/edat_lfse_19/default/table?lang=en\n\nhttps://ec.europa.eu/eurostat/databrowser/view/edat_lfse_20/default/table?lang=en\n\n[The AI Job Apocalypse Is a Mirage](https://pub.towardsai.net/the-ai-job-apocalypse-is-a-mirage-065ccca81d9d) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.", "url": "https://wpnews.pro/news/the-ai-job-apocalypse-is-a-mirage", "canonical_source": "https://pub.towardsai.net/the-ai-job-apocalypse-is-a-mirage-065ccca81d9d?source=rss----98111c9905da---4", "published_at": "2026-08-18 15:01:02+00:00", "updated_at": "2026-08-18 15:41:36.732671+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-policy"], "entities": ["Brynjolfsson", "Chandar", "Chen", "Felten", "Raj", "Seamans", "ChatGPT"], "alternates": {"html": "https://wpnews.pro/news/the-ai-job-apocalypse-is-a-mirage", "markdown": "https://wpnews.pro/news/the-ai-job-apocalypse-is-a-mirage.md", "text": "https://wpnews.pro/news/the-ai-job-apocalypse-is-a-mirage.txt", "jsonld": "https://wpnews.pro/news/the-ai-job-apocalypse-is-a-mirage.jsonld"}}