The most convenient story about AI and jobs has a clean opening scene: ChatGPT arrives in November 2022, and the labour market starts to wobble. A January working paper on arXiv suggests the wobble started first. That is awkward for a narrative that already had its protagonist cast.
The paper does not claim generative AI is harmless. It claims something narrower and, for anyone who likes tidy causation, more irritating.
What the researchers actually did
The authors combined three kinds of records. One was US unemployment insurance data. Another was career histories drawn from online profiles. The third was university course information. Together these let them compare how things changed across occupations and across graduating cohorts, rather than staring at one aggregate unemployment line and guessing.
That design matters because the question is not whether the labour market got harder for some people. It is when, for whom, and in which jobs.
The timeline problem
The central finding is about dates. In occupations with high exposure to large language models, the risk of unemployment began climbing from a low point in early 2022. ChatGPT launched in November of that year, so the upturn predates it by almost twelve months.
The graduate data tells a similar story:
- The 2021 cohort. Graduates from the class of 2021 onward entered these high-exposure occupations less often than the cohorts before them.
- The timing of the gap. That gap opened before the end of 2022, again ahead of the chatbot that everyone now remembers as the starting gun.
If ChatGPT were the switch, the lights would not have flickered the winter before it was installed.
What the authors are not saying
The researchers do not argue that generative AI left the labour market untouched. Their point is that the launch cannot serve as a clean causal boundary, because the negative trends had already begun.
Their reading of the evidence points elsewhere, to economy-wide and sector-specific shifts that were already putting pressure on technology-heavy occupations. In other words, these jobs may have been in trouble for reasons that had little to do with a chatbot, and the chatbot then arrived at a scene already underway.
There is a tidy irony here. AI gets blamed for a slump that, on this evidence, had started without it. Technology is occasionally given credit for events it merely attended.
A reminder about what 'exposure' means
The analysis also carries a methodological warning that tends to get lost in headlines. Exposure to AI measures how far the tasks in an occupation could be affected by what language models can do. It is not a count of jobs actually lost to AI.
Those are very different things. An occupation can be highly exposed in principle and still lose workers for unrelated reasons, which is more or less what the timeline suggests happened.
The flip side deserves equal honesty. The data is observational, built from unemployment records and career histories, and it does not rule out later effects from LLM adoption in particular industries or regions. A paper showing that the story started early is not a paper showing that the story ended there. Deployment takes time, and an effect that arrives after the data stops will not appear in the data.
The education finding
The paper's more encouraging result concerns graduates. Those whose programmes had higher AI exposure did better after ChatGPT appeared. They earned higher pay in their first job and spent less time looking for it.
The authors are careful about how far to push this. It is an association, not a promise to any individual worker or group. Still, they see it as support for explicitly connecting skills, curricula and career outcomes, rather than hoping the three line up by themselves.
That is a modest prescription, and a sensible one. It also happens to be the kind of recommendation that survives whichever way the causation argument ends up.
What to take from it
Three things are worth keeping.
- Dates are evidence. When a decline starts before the proposed cause exists, the proposed cause needs company.
- Exposure is a forecast of possibility. It is not a tally of casualties.
- Absence of proof cuts both ways. The paper weakens the claim that ChatGPT triggered the downturn, but it leaves open later, more localised effects.
The practical upshot is less dramatic than either camp would like. Blaming AI for everything is lazy, and dismissing it entirely would be equally so. The harder job is measuring what is happening, occupation by occupation and cohort by cohort, which is slower than writing a headline and considerably less fun.