Generative artificial intelligence (AI) has significantly reduced the time developers spend writing code, but a survey found it is also increasing review and maintenance burdens as low-quality code flows in.

A Linux Foundation report cited by IT media outlet TechRadar on Wednesday said respondents estimated coding time fell by an average 52 percent after adopting AI. Maintainers, however, assessed that 25 percent of incoming pull requests (PRs), or code-change requests, amounted to low-quality “AI slop.”

The survey also found that time saved was used for other development work. Some 66 percent of respondents said they use time freed up by AI to address technical debt, or accumulated code and design issues. Those who said they use it to develop new features came to 63 percent, while 58 percent said they spend it on learning and skills development.

Individual time savings did not necessarily translate into efficiency gains across the development process. The faster code is generated with AI, the more maintainers have to validate for quality and security. The reduced burden at the coding stage shifts to the review stage.

AI’s impact on code review also varied by task difficulty. It helped with simple tasks, but 21 percent of respondents said complex work takes longer. False positives and duplicate submissions in reports written by AI were also cited as factors that consume reviewers’ time.

Pre-submission verification and disclosure of AI use were also insufficient. Some 61 percent of respondents said they submit code only after closely reviewing AI-generated code and testing full functionality. Procedures for informing the next person that AI was used were also not properly followed, adding to the burden on maintainers responsible for final checks.

The burden was more pronounced in large projects. Some 59 percent of maintainers said code contributions increased after generative AI spread, and the share rose to 69 percent in large projects. The proportion of PRs entering large projects estimated to be low-quality “AI slop” was also higher than the overall average, at 31 percent.

Governance systems, however, did not keep up with the pace of AI adoption. Only 35 percent of projects had policies that officially allow the use of generative AI. Some 32 percent said they had no related policy and no plan to establish one.

The Linux Foundation stressed that rather than excluding AI from coding, governance should be strengthened. It said clear contribution standards for AI use should be set and responsibility for code testing and disclosure of AI use should be placed on contributors. It added that tools should also be improved so maintainers can manage incoming PRs efficiently.