AI Coding Data Shows Widespread Use but No Industrywide Code Share

AI tools are common in software development, but recent findings measure different kinds of use and output, leaving no comparable industrywide figure for how much code AI produces.

At Google, AI-generated code approved by engineers accounted for 75% of new code on April 22, up from 50% the previous fall. That figure covers new code at one company. It cannot be applied to software development as a whole.

A survey conducted in October 2025 and published in January 2026 asked professional developers to estimate the share of code they committed that AI had generated or significantly assisted. Respondents put the 2025 share at 42% and projected 55% for 2026 and 65% for 2027. All 1,149 participants had used AI at work in the preceding year, and 979 answered the code-share question. The figures are estimates from that group, not measurements of code across the industry.

A report released June 10 presented a different measure: 67% of 200 U.S. technology decision-makers placed AI-generated or significantly refactored code at 51% to 75% of their organizations' weekly output. Participants were managers or more senior staff at larger companies already using AI in software engineering.

Those two surveys ask different people about different kinds of AI involvement. Generating or significantly assisting code a developer commits is not the same measure as generating or significantly refactoring an organization's weekly output. Neither estimate can be combined with Google's figure to calculate an industry average.

Adoption figures answer a separate question. A 2025 global survey of nearly 5,000 technology professionals found that 90% of software development professionals used AI at work. Respondents reported a median of two hours of AI use on a typical workday. Frequency and time spent with a tool do not reveal how much of the resulting code it produced.

Another 2025 survey collected 49,009 responses from 177 countries. It found that 84% of respondents used or planned to use AI tools, while 51% of professional developers used them daily. On the accuracy of AI output, 46% expressed distrust and 33% expressed trust. The full response count describes the survey; it is not necessarily the denominator for each AI question.

Repository data tracks changes to code but cannot by itself identify their cause. Across an analysis of 623 million code changes from 2023 through 2026, duplicated code blocks per million changed lines rose 81%, from 40.3 in 2023 to 73.0 in 2026 year to date. The share of moved code fell from 21% of changed lines in 2022 to 3.8% in 2026 year to date. Those patterns do not establish that AI caused either shift.

A security benchmark examined a narrower outcome. AI-generated code achieved an average 56% pass rate on a test set of 80 function-level coding tasks covering four programming languages and four vulnerability types. The latest model group contained 11 models. Performance on those tasks does not establish the rate of vulnerabilities in production software.

Results on developer speed also depend on the task and setting. In a randomized trial conducted from February through June 2025, 16 experienced open-source developers were assigned 246 tasks. Completion took 19% longer when AI tools were allowed. The trial tested those developers on familiar open-source projects with tools available at the time; it did not measure every kind of software work.

In February 2026, the researchers behind the randomized trial described tentative follow-up estimates pointing toward faster work with AI. Participant and task selection effects, along with unreliable time tracking when developers ran multiple agents, left only weak evidence for the size of any improvement.