The Software Industry and Artificial Intelligence: Nearly Universal Use, Evidence That Ranges from Large Gains to a Slowdown, and the Question of Who Gets Hired

Purpose. The software industry both builds artificial intelligence and was among the first to use it on its own work, through coding assistants. This article asks where AI stands in software in 2026, what controlled studies say about developer productivity, and where the industry may be by about 2035. Design and methods. A review combining the share of the software research literature that involves AI (OpenAlex, 2010–2025), Eurostat's survey of AI use in EU enterprises in programming, consultancy and information services (2021–2025), three randomised studies of coding assistants with different results, a user study of security, and a study of early-career employment. Claims taken from the literature were checked against the sources cited. The scenarios are judgements built on this evidence. Findings. The share of software works that mention AI rose from 7.4% in 2018 to 34.6% in 2025, 4.7 times as much, and language-model works from 198 to 4,052, 46.3% of AI-related works, the highest share of the 20 industries. In the EU, 65.7% of enterprises in the sector used AI in 2025, the highest of the twelve sectors with Eurostat data and more than three times the 19.9% for all enterprises, and 42.6% used three or more technologies. Controlled studies disagree about productivity, and the disagreement is informative. In a laboratory experiment, developers with an AI pair programmer completed a programming task 55.8% faster; in three field experiments with 4,867 developers at three large companies, the number of completed tasks rose by 26.08%, with larger gains for less experienced developers; but in a randomised trial with 16 experienced open-source developers working on their own mature projects, allowing AI tools increased completion time by 19%, while the developers believed that it had cut their time by 20%. A user study found that participants with an AI assistant wrote significantly less secure code and were more confident in it. US payroll data show employment of workers aged 22–25 in AI-exposed occupations 19% below the level that matching less-exposed peers would imply, mainly through reduced hiring; this is descriptive. We expect AI coding tools to be near-universal by 2028, work to shift toward specifying, reviewing and testing in 2029–2032, and the entry-level pipeline to be the open question. Originality. The article places three randomised studies with conflicting results side by side, explains what each can and cannot show, and links them to evidence on code security and on who is being hired, in the industry that uses generative AI most.