🧭 Grass Brief

First inning: everyone claims to know what AI is doing to jobs. The best research contradicts itself. Now ETH has looked at Switzerland.

August 17, 20268 min readDennis Gorican

Anyone who follows the coverage of AI and the labour market experiences a peculiar kind of whiplash. One week a headline announces that AI is eliminating knowledge work on a large scale. The next, another study finds no measurable effect at all. Then a third explains that the whole story was misread from the start.

For most readers this is background noise. For executives in a career transition it is something more serious. These headlines shape decisions: which direction to reposition in, which skills to acquire, which industries to avoid, how to explain one’s own field in an interview. People quietly rebuild their careers around this quarter’s confident claim.

That is why this week we are doing something slightly unusual for the Grass Brief. Instead of summarising a single study, we show the actual state of the evidence, contradictions included, and what the first serious Swiss analysis has found. Because the honest picture is more useful than the confident one.

The contradiction at the top of the field

The Hamilton Project at the Brookings Institution recently summarised the leading academic research on AI and employment, and the title alone says everything about the maturity of the field: research on AI and the labour market is still in the first inning.

Consider what the top studies actually find. A team led by Stanford economist Erik Brynjolfsson linked measures of AI exposure to detailed wage data from the payroll provider ADP and found clear declines in employment among career starters in AI-exposed occupations, while experienced and senior workers in the same occupations remained stable or grew. This is the study behind most “AI is eating entry-level jobs” headlines, and it matches the Swiss jobs.ch data we covered in our article on the senior paradox.

Then came Eckhardt and Goldschlag, using publicly available data from the Current Population Survey. Their finding: unemployment rose less in occupations with higher AI exposure. Not more. Less. The core result held across different exposure measures.

And finally Iscenko and Millet, who did find that job postings fell more in AI-exposed occupations, but: the decline began in 2022, before the public launch of ChatGPT. Their argument is that the timing fits the macroeconomic turn of rising interest rates better than the launch of large language models. In other words, part of what the world attributes to AI may simply be expensive money.

Three rigorous pieces of research. Three substantially different answers. That is not a scandal; this is what a young field of research looks like. But it should permanently change how you read the next confident headline.

What the ETH found when it looked at Switzerland

Into this contested landscape now comes the first serious Swiss evidence. KOF economists Jeremias Klaeui and Michael Siegenthaler of ETH Zurich have for the first time applied the AI exposure methods developed for the US market to Swiss labour market data, using two independent data sources: the administrative data of all jobseekers registered with the employment offices, and job posting data covering practically the entire Swiss online job market.

Their central finding: after the introduction of ChatGPT and other large language models in autumn 2022, the labour market in highly LLM-exposed occupations developed significantly less favourably than in occupations with low exposure. Since November 2022, the number of registered jobseekers in highly exposed occupations has risen on average 27 per cent more than in occupations with low exposure. Consistently, job postings in highly exposed occupations fell significantly more over the same period.

The most visibly affected groups: application programmers, software developers, systems analysts, journalists, and advertising and marketing professionals. Least exposed: cleaners, caretakers, seamstresses, occupations that rely on physical presence.

Two methodological details give the study its weight. First, the results hold across both independent data sources and two different exposure measures. Second, and more importantly: the trends in occupations with different exposure ran in parallel for years before 2022. The divergence only begins at the end of 2022, precisely with the arrival of the language models. As Siegenthaler puts it, the results suggest fairly strongly that generative AI is behind the recent rise in unemployment in some highly AI-exposed occupations.

One Swiss nuance deserves attention. In the US data, the effects are heavily concentrated among younger workers. In Switzerland too, jobseekers under 50 are somewhat more affected than older ones, but the difference between the age groups is much less pronounced. The Swiss pattern is broader and flatter than the American one.

What the ETH study does not say

This is where epistemic discipline matters, because this study will be cited for years, often wrongly.

The KOF analysis measures relative shifts between occupational groups, not the effect of AI on total employment. General developments that affect all occupations equally, including the economic cycle, are deliberately factored out. And the context is essential: these shifts are happening within growing overall employment. The authors themselves note that AI is plausibly making Swiss firms more productive, which should support overall employment, effects their study deliberately leaves out.

The honest summary of the best Swiss evidence is therefore: AI is visibly redistributing pressure between occupations, with routine-intensive knowledge work absorbing the most. But according to current evidence, it is not shrinking the Swiss labour market. Those are two different statements, and almost every alarming headline squeezes them into one.

What we can reasonably treat as established

Putting together the credible evidence, global and Swiss, a few conclusions hold across studies and can bear the weight of a career decision.

The pressure is real but unevenly distributed. Highly exposed, routine-intensive knowledge work, drafting, programming standard components, standardised analyses, formatted communication, faces measurably deteriorating demand, in Switzerland as elsewhere.

The seniority gradient keeps reappearing. Whether in Stanford’s ADP data or in the Swiss job posting data from our senior paradox article, the pattern repeats: pressure concentrates where work consists of producing output and eases where work consists of judging it. Demand for senior profiles in Swiss job ads rose by 26 per cent, while junior ads fell by around a third.

And the overall numbers remain resilient. Swiss employment is at a record level, and no serious study currently demonstrates aggregate job destruction by AI.

What nobody actually knows yet

Just as important is what nobody can tell you yet. Whether the occupational shifts are permanent or a temporary adjustment. How much of the decline in job postings belongs to AI and how much to interest rates and the economic cycle; the timing puzzle of Iscenko and Millet is unresolved. Whether the productivity gains create new roles fast enough to absorb those displaced, as every wave of technology has eventually done so far. And which specific occupations will be redefined rather than reduced.

Anyone who claims certainty on these questions is selling something.

How to decide when the evidence is contested

Which brings us to the practical question: how should an executive in transition act when the world’s best researchers contradict each other?

First: stop repositioning yourself along headlines. We see clients who want to rewrite their entire narrative every time a new study appears. A career strategy that changes with the news cycle is not a strategy. The evidence justifies exactly one structural conclusion: shifting your positioning towards judgement, responsibility and context, the level that every study shows is holding or growing, and away from pure output production. This conclusion has survived every contradictory finding of the past year. Build on what is robust across studies, not on what a single one claims.

Second: treat your own market conversations as data. The big studies describe averages across millions of job ads. Your transition takes place in a specific segment, at a specific level, in specific companies. Ten real conversations with the people who can actually influence your next step, former colleagues with hiring responsibility, board members, trusted headhunters, industry contacts, tell you more about the demand for your profile than any exposure index. When the macro evidence is contested, the micro evidence you gather yourself becomes the most reliable input you have.

Third: hold your conclusions loosely and your process firmly. The honest stance in a first-inning field of research is provisional belief: strong enough to act on, loose enough to update. What should not be provisional is the process: clear positioning, a warm network, visible judgement, financial and psychological stamina for a realistic timeline. These things work under every scenario the research currently allows.

The market will keep producing confident headlines. The researchers, to their credit, keep producing honest disagreement. Somewhere in between, good career decisions are made: informed by the evidence, anchored in your own market and built on the one finding that refuses to go away, that experience and judgement are what the machines make ever more valuable.

Sources: Klaeui & Siegenthaler (KOF, ETH Zurich), “KI und der Schweizer Arbeitsmarkt: Erste Evidenz zu Auswirkungen auf Arbeitslosigkeit und Stellenausschreibungen” (registered jobseekers in highly LLM-exposed occupations up 27 per cent more since November 2022 than in occupations with low exposure; two independent data sources; age effects less concentrated than in the US; relative shifts within growing overall employment); The Hamilton Project / Brookings, “Research on AI and the Labor Market Is Still in the First Inning”, 2026; Brynjolfsson, Chandar & Chen (Stanford / ADP payroll data), 2025; Eckhardt & Goldschlag (Current Population Survey), 2025; Iscenko & Millet (job postings and interest-rate timing), 2026; Stanford HAI, AI Index Report 2026; JobCloud, jobs.ch KI Report 2026.