What future for technical jobs with AI?

Among technical occupations, software development is the one where AI is used the most, at least at Anthropic: in August 2025, 36% of conversations with Claude concerned computing and mathematical tasks, the 1st category, far ahead of education (about 12%). It is also the best-measured occupation: surveys of tens of thousands of developers, randomised trials in companies, series of job postings.

This analysis is limited to technical occupations. The exposure of all occupations is covered in Will AI replace occupations?, and the entry of young people into employment in What happens to entry-level workers when AI does their work?.

In 2025, 84% of the 33,662 respondents to the annual Stack Overflow survey used AI to develop or planned to do so, and 51% of professional developers used it every day. The survey asks, task by task, whether AI is already used for it, and for which tasks the respondent does not plan to use it.

We read in this a split that follows responsibility: AI is used where an error shows quickly and costs little, such as a piece of code that does not compile or an answer to check, and kept away from where an error goes into production, such as the review that validates the code, deployment and monitoring.

Trust remains low. 46% of respondents distrust the accuracy of AI tools, against 33% who trust them. The 1st frustration, cited by 66% of them, is an answer that is "almost right, but not quite" (our translation); 45% find it takes longer to debug code written by an AI.

Randomised trials give opposite results, and the difference comes down to the developer and the code.

The largest gain concerns a short task, written from scratch, with no existing code to know. The only trial that measures a loss concerns developers who know a very large codebase thoroughly and have contributed to it for several years. Before the trial, they expected AI to give a gain of 24%; afterwards, they estimated they had gained 20%, whereas they had taken 19% more time.

In 2026, METR could not repeat the measurement reliably: developers refused to work without AI, and the tasks where they expected the most gain dropped out of the trial. The confidence intervals contain zero. Other studies find the same gap between perceived and measured time; they are read in Does AI really make us more productive at work?.

Developer postings swelled during the pandemic, peaked in spring 2022, then fell: according to Indeed, the fall in postings in occupations exposed to AI began before the release of ChatGPT, at the end of 2022. Developer postings fell below their 2020 level in spring 2023, and were close to 33% below it in October 2024. They have been rising since 2025: 71% of the increase from May 2025 to May 2026 comes from experienced-level positions, and 37% from postings that mention AI in their title. Indeed notes that, since May 2022, the occupations most exposed to AI have fallen the most, a correlation whose cause it does not establish. None of the studies we have found separates the share due to AI from the other causes of this cycle.

The US Bureau of Labor Statistics separates 2 occupations in its projections from 2025 to 2035: employment of developers would grow by 10%, driven notably by the development of AI, and that of programmers, who translate into code the plans designed by developers, would fall by 7%, notably because companies would have to automate repetitive programming tasks with AI.

The International Labour Organization places software developers in its gradient 3, the 2nd exposure class out of 6, with a mean score of 0.53. Web and multimedia developers are in gradient 4, the most exposed (0.60), which mainly groups office jobs. For France, we have not found a public series of developer job postings that would allow the same comparison.

For mechanical, civil or process engineers, and for maintenance, laboratory or design office technicians, we have found neither a randomised trial nor a hiring series that isolates the effect of generative AI. Exposure studies rate the tasks of these occupations on what a language model can do with text and code.

The International Labour Organization itself points out the limit. In the survey used to rate the tasks, respondents answered "I don't know" 2 times more often for technical and manual tasks, which, according to the authors, suggests that they felt less able to judge them. Part of these occupations takes place on a building site, in a factory or in front of a machine, where a language model does not act.

What comes closest is the documentation side of these occupations: writing a calculation note, searching a standard, writing a test report, programming a controller. These tasks resemble the developer's, and the measurements above can be partly transposed to them, with caution: none measures them directly.