In 2023 the technology industry was sold a forecast that many programmers took as a death notice: AI would replace software developers within a few years, delivering tireless digital colleagues that never asked questions and never made mistakes. Three years on, the industry looks different, but not in the way the forecast promised. AI has changed how software is built; it has not removed the people who build it.
The layoffs were real
The job cuts that accompanied the forecast were not imaginary. According to the tracker Layoffs.fyi, 152,922 employees were laid off across 551 technology companies in 2024, following even larger waves in 2022 and 2023. In early 2025 several large firms eliminated thousands more positions and described the moves as restructuring around AI.
At the same time, executives claimed that a growing share of new code at major companies was being written by automated systems. Those claims were hard to check from the outside, and the broader pattern of corporate AI spending told a more mixed story: many organizations reported impressive demonstrations but struggled to show measurable returns from production deployments, and AI systems proved weak at holding the context of large, complex codebases over time.
Speed in demos, friction in practice
The most striking evidence on productivity came from a randomized controlled trial published by the research nonprofit METR in July 2025. Experienced open-source developers working on their own repositories took 19 percent longer to complete tasks when they were allowed to use early-2025 AI coding tools, even though the same developers believed the tools had made them about 20 percent faster. The authors cautioned that the result applied to experienced contributors on mature projects and might not generalize, but it punctured the assumption that AI assistance automatically speeds up skilled engineers.
Code quality researchers have raised a related concern: code produced by prompting in natural language tends to be repetitive and structurally simple, duplicated rather than refactored. Some engineers have taken to calling the result a “slop layer,” software that runs but whose internal logic is hard to understand or fix when it breaks. Accumulated across large codebases, that becomes a maintenance burden, and security reviewers have reported basic vulnerabilities in generated code. The practical effect is that senior engineers spend more of their time supervising and correcting automated output rather than less.
The junior pipeline problem
The deepest effect has been on people. When companies concluded that AI would handle entry-level work, hiring of junior developers fell sharply. That removes the first rung of the ladder that produces the next generation of senior engineers. New graduates who do find jobs are often expected to work directly in complex systems without the gradual apprenticeship that used to come first.
The balance of power in the labor market shifted as well. Salary growth stalled and in some segments reversed, and the AI narrative was used as leverage in negotiations: if machines are doing so much of the work, the argument went, human skill should cost less. In practice, organizations continued to depend on exactly the people who could catch and repair AI’s mistakes.
A tool, not a replacement
None of this means the technology failed. It means its role was misunderstood. AI coding tools are good at repetitive work, at drafting boilerplate and at surfacing patterns in large amounts of data. Where teams treated them as the sole solution, problems multiplied; where they paired them with human judgment, experience and accountability, results improved. That frees engineers to spend more time on design, architecture and the decisions that machines handle poorly.
A growing number of companies are now investing in retraining, upskilling and what they describe as human-centered AI strategies, and the industry conversation has moved toward responsible deployment, transparent systems and keeping people at the center of decisions. The lesson of the past three years is that the future of software is neither fully automated nor purely human. The developers who do well will be the ones who learn how the tools work, understand their limits and build the balance between the two into their everyday practice.
