the strike has barely moved

We have been hearing for years that artificial intelligence is going to destroy millions of qualified jobs. Dario Amodei himself, CEO of Anthropic, said last year that AI could affect half of administrative jobs entry level in the coming years. Mustafa Suleyman, head of AI at Microsoft, was more aggressive in your estimatesensuring that most professional work would be replaced within twelve to eighteen months.

Now the same Anthropic publish a study which, without denying that the risk exists, forces these predictions to be greatly qualified.

What the study measures. The research, signed by economists Maxim Massenkoff and Peter McCrory, introduces a new metric called “observed exposure.” The idea is that instead of asking what tasks AI could do in theory, the authors analyze what it is actually doing now in professional settings, using usage data from Claude in work contexts.

The gap between theoretical capacity and actual use. Taking the computer science and mathematics sector as an example, language models would be capable, in theory, of executing 94% of the tasks associated with these professions. In practice, Claude covers 33%, according to the study. In office automation and administrative positions, the theoretical capacity is close to 90%; actual use is far below.

The authors themselves illustrate their metric with an example: authorizing the refilling of medical prescriptions to pharmacies is a task that a language model could easily automate, but the study’s researchers have not observed that Claude was currently doing it.

And the barriers to AI not automating these types of tasks include legal restrictions, the need for human verification, barriers with software integration, and more. That is to say, the researchers show that all of these tasks could already be done theoretically by AI, but they are not yet being done due to these restrictions that the human being himself imposes.

Who are the most exposed. According to the studythe jobs with the highest observed exposure are computer programmers (74.5%), customer service positions (70.1%) and those who operate by entering data (67.1%). At the opposite extreme, 30% of workers have zero exposure: cooks, mechanics, lifeguards, or waiters. They are jobs that require physical presence and that, according to the study, no language model can replicate. For this we would still have to give robotics a lot of time.

The demographic profile of the most exposed group also breaks with the usual imagination. According to the study, these workers are 16% more likely to be women, earn on average 47% more, and have significantly higher levels of education. Anthropic reveals in the study that it is not the warehouse worker who is in the spotlight, but the financial analyst, lawyer or software developer.

Unemployment. This is the most striking fact of the investigation, because since the arrival of ChatGPT at the end of 2022 until today, the study says that there is no statistical evidence of a systematic increase in unemployment among workers most exposed to AI. The effect, according to the authors themselves, is “indistinguishable from zero.” The Bureau of Labor Statistics yes it projects that the most exposed jobs will grow less between now and 2034. We will have to wait a few years to study how the metrics progress.

The youngest, the most affected. The researchers do detect a worrying sign among workers aged 22 to 25: the rate of entry into jobs in high-exposure sectors has fallen by approximately 14% in the post-ChatGPT era compared to 2022. The authors attribute this phenomenon more to a slowdown in hiring than to layoffs. But they warn that the signal is “barely statistically significant” and that the causes could be several: from young people who simply stay longer in their current jobs, to those who opt for other sectors or going back to school.

What limitations does the study have? From Forbes, some analysts have pointed out that the research measures the use of Claude, not the use of AI in the economy as a whole. Companies also use ChatGPT, Microsoft Copilot, Gemini or own models, and those interactions do not appear in the data. The authors are aware of this and acknowledge it in the text. The conclusion that “AI is far from reaching its theoretical capacity” depends in part on the limits of what they can measure, and not just the actual limits of adoption.

So should we relax? The authors themselves advise against it. The proposed analysis is said to be designed precisely for scenarios in which the impact arrives gradually and is difficult to detect until it is too late. They point out that the effects of AI on employment could be more like those of the internet or trade with China than those of COVID: slow, diffuse, complicated to isolate from other economic factors.

They also warn that if the gap between theoretical capacity and actual use closes, as they expect to happen as models improve and adoption spreads, the most vulnerable groups will be precisely those who today have better salaries and more training.

Cover image | Unsplash (charlesdeluvio, Emiliano Vittoriosi)

In Xataka | NVIDIA has lost hope in China, which is why it has started manufacturing its own next-generation GPUs for AI