Anthropic's AI scenarios meet Swiss labour market data

15 percent economic growth per year and at the same time unemployment as in a severe recession: Anthropic's new scenario model delivers a pointed combination.

The Anthropic paper states explicitly that the three pathways are not forecasts and are not assigned probabilities of occurring. The 15 percent figure in the Euronews headline reproduces the extreme model path correctly; without that context, however, it looks like an expectation. The question is rather: what would have to happen for artificial intelligence to fundamentally change growth, employment and income by 2030?

Three calculations, no crystal ball

The technical report «Working Paper 2026-02», dated September 2026, models three US paths to 2030. In the restrained scenario, gross domestic product in 2030 is only 1.6 percent above a development without AI. In the middle scenario the gap is 8.3 percent. Only the extreme path leads to a historic acceleration: in 2030 the annual growth rate reaches 15 percent; GDP then stands 32.4 percent above the comparison path.

That requires extraordinary conditions. AI would have to be more productive than humans at almost all cognitive tasks, handle a large share of them autonomously, be deployed rapidly in companies, and create hardly any new knowledge work for people. Anthropic links this scenario to self-improving systems. That is a thought experiment at the edge of the possible, not an expected value.

That is precisely why the distributional calculation is interesting. In the extreme US scenario, almost one in five people in a cognitive occupation would be unemployed. The share of labour income in the modelled total economy would fall from 60 to 45.2 percent. A significantly larger economy would therefore not automatically be an economy in which employees win.

In Switzerland the shift is already measurable

The US figures cannot be transferred to Switzerland because of differences in industry structure, vocational training and social protection. Nevertheless, there are first signals that fit the model's mechanism. A KOF study at ETH Zurich compares occupations with high and low exposure to language models. Since the end of 2022, the number of registered job seekers in highly exposed occupations rose by up to 27 percent more, in relative terms, than in occupations with low exposure. At the same time, online job advertisements declined more sharply there. Those affected included software development, journalism, advertising and marketing. The summary does not give an absolute number of people for the 27 percent relative difference; it must not be read as an increase in all job seekers.

That is not proof of general AI unemployment. The study measures differences between occupational groups before and after the release of ChatGPT; cyclical or sector-specific factors cannot be fully ruled out. In March 2026, the KOF estimated for the period since the end of 2022 that technological change could be connected with 7,000 to 10,000 lost jobs and could explain at most a fifth of the recent rise in unemployment. This is a model-based order of magnitude, not an official count.