Is generative artificial intelligence destroying jobs or transforming them? A study from Stanford University and King’s College London provides one of the first internationally comparable answers. The paper, titled “How Does AI Change Labor Demand? Evidence from 41 Countries” and written by Bharat Chandar of Stanford and Bouke Klein Teeselink of King’s College London, concludes that firms adopting AI hire slightly more overall, but shift the composition of their workforces in favor of senior staff.
King’s College publicized the results on October 6, 2026, after the Stanford Digital Economy Lab published the paper on September 21. In the days since, its figures have circulated widely in the public debate on AI and work. According to the British university’s summary, AI adoption at a firm is associated with an increase of around 3.3% in the number of employees, with senior employment growing by 6.7% and junior employment falling by 2.5%.
The scale of the data is remarkable. The authors analyzed a sample of 1.25 billion job postings and 154 million employment records across 41 countries. As Chandar explains on his blog, the data come from Revelio Labs, a workforce analytics company that collects public information from sites such as LinkedIn and from job boards. The authors themselves caution that their estimates should be understood as relevant to digitally attached multinational firms.
How do you measure whether a company has adopted AI? The researchers searched more than a billion postings for terms such as ChatGPT, Claude, Gemini, large language model, prompt engineering, copilot and retrieval-augmented generation. They then asked a language model to review the candidates to filter out stray matches, such as an airline’s copilots, the name Claude or “LLM” as a law degree, and ads that mention AI only to bar its use. The authors consider the result a lower bound on actual adoption.
The countries with the highest adoption, according to the analysis, include India, Israel and Singapore, in line with other recent measurements, and adoption is rising everywhere, fastest in IT and finance. To make fair comparisons, the study focuses on multinational affiliates in each country and matches those that adopt AI with similar ones that do not: same country and industry, and similar size, AI exposure and junior share of employees before ChatGPT’s launch in November 2022.
To separate causation from correlation, the authors add an instrumental variables approach. The idea is that an affiliate whose parent company sits in a city with high AI adoption is more likely to be affected by the technology than one whose parent sits in a city with lower adoption, for reasons unrelated to the local market. Chandar acknowledges the method is not foolproof and that the study includes many alternative tests, whose results are broadly consistent.
The aggregate result shows an increase in employment of around 3.3% at affiliates that adopt AI, a figure the authors describe as only marginally statistically significant. That growth comes from a 6.7% rise in senior employment, while junior employment shrinks by 2.5%. Overall, the junior share of employees falls by about two percentage points. The Stanford Digital Economy Lab summary stresses that this decline in the share is driven mainly by growth in senior employment.
The study also detects shifts in the type of work. Among seniors, employment moves toward occupations more exposed to AI, while the estimates for juniors point to a shift in the opposite direction, although that result is not statistically significant. The pattern is especially clear in computer and mathematical occupations, which gain weight in total employment but see the largest drop in the junior share.
The international dimension is one of the most relevant contributions. In the 31 countries with sufficiently precise estimates, the junior share of employment falls in 23, and the effect is statistically significant in seven countries as diverse as Brazil, Saudi Arabia and the United States. The authors find signs that effects on juniors are more negative in richer and more digitized countries, and less negative where labor markets are more rigid, although they present these correlations as suggestive.
Klein Teeselink summed up the finding in the King’s College release: it is difficult to predict which occupations new technologies will replace and which they will complement, and while the study does not settle the question, it indicates how workforces change after adopting generative AI, with a similar pattern in many labor markets. Chandar added that the results suggest AI is changing the mix of jobs within organizations rather than reducing overall employment.
However, Chandar himself qualifies an overly optimistic reading on his blog. The fact that AI-adopting firms grow relative to competitors does not mean total employment rises: if one company gains share because a rival closes, it can hire more even as aggregate employment falls. That is why he distinguishes between studies comparing adopting and non-adopting firms, which explain mechanisms, and those measuring each occupation’s exposure, which are more useful for knowing whether total employment is going up or down.
His personal conclusion, after reviewing the literature, is cautious: no widespread displacement so far, but likely sizeable negative impacts for the most exposed junior workers. That reading matches other work, including the “Canaries in the Coal Mine” study Chandar co-wrote with Erik Brynjolfsson and Ruyu Chen, and Klein Teeselink’s earlier research on the British labor market.
The implications for workers, companies and governments are clear. If AI expands the work of experienced staff and shrinks that of newcomers, the traditional path of on-the-job training could narrow. Companies could face a shortage of senior talent in the medium term if they hire fewer juniors today, and education systems will need to prepare new entrants for tasks that AI does not cover.
In short, the Stanford and King’s College study paints a nuanced picture: at multinationals that adopt generative AI there is no net job destruction, but rather a reshuffling that favors seniors and hurts juniors in most of the countries analyzed. The authors insist that their estimates describe digitally attached firms and that their methodology has limits, but the global pattern they describe is hard to ignore.
