Most software written by AI

How far are we from most software being written by AI?

AI tools already write a visible share of new code at some large companies. Google's prepared remarks for its Q3 2024 earnings call, dated 29 October 2024, said that more than a quarter of its new code was generated by AI and then reviewed and accepted by engineers. A news report says Google's chief executive put the figure at 75 percent in April 2026. Microsoft's chief executive said in April 2025 that roughly 20 to 30 percent of the code in some projects was AI-written. These are company statements, and the sources reviewed give no method for how code was counted, so they are not independent measurements. Benchmarks, controlled trials and security studies do not yet show that AI-written code is as reliable as human-written code, so most software being AI-written is still a long way off.

AI illustration · Illustrative scene, not a photograph of the specific device or experiment.
WHERE WE AREScaling up

Nearly here. It works and is being scaled up.

AI coding agents are sold as commercial products and used inside large companies, and some of those companies report that AI writes a large share of their new code, so the stage is scaling. It is not widespread. The share figures come from company statements that do not define how code was counted. The main independent measures are a 2025 controlled trial that found experienced developers took longer with AI tools, a 2026 follow-up whose authors call their data very weak evidence, and a benchmark whose validity is contested. No source gives a share of all software written by AI.

The tech tree

Everything recorded on Vrulum that this future needs, oldest foundations at the top and the goal at the bottom. Lines show what each step requires.

What it takes: Most software written by AI Unlocked Being researched Locked
  1. 2023AI coding agentsCommercial Unlocked.
  1. 2023AI agents resolve real software issues on a standard benchmarkIn progress Being researched.
  2. 2024AI writes a minority share of new code at large companiesDone Unlocked.
  3. 2025Controlled studies show faster completion by experienced developersIn progress Being researched.
  4. 2026Reported share of AI-written code at large companies keeps risingIn progress Being researched.
  5. Date not recordedSecurity and maintenance of AI-written code shown to match human-written codeNot started Locked.
  6. Date not recordedMost software written by AINot started Locked.
  1. GoalMost software written by AIGoalRequires: AI coding agents, AI writes a minority share of new code at large companies, Reported share of AI-written code at large companies keeps rising, AI agents resolve real software issues on a standard benchmark, Controlled studies show faster completion by experienced developers, Security and maintenance of AI-written code shown to match human-written code, Most software written by AI. Locked.

The road so far

1 of 6 steps done, 3 in progress.

  1. Done · 29 Oct 2024AI writes a minority share of new code at large companies

    Alphabet's prepared remarks for its Q3 2024 earnings call, dated 29 October 2024, say that more than a quarter of all new code at Google is generated by AI and then reviewed and accepted by engineers. In April 2025, at Meta's LlamaCon event, Satya Nadella said maybe 20 to 30 percent of the code in some Microsoft projects was AI-written, and he described newer Python code more favourably than C++. Neither statement gives a method for counting code. [1][3]

  2. In progress · 24 Apr 2026Reported share of AI-written code at large companies keeps rising

    Semafor reported on 24 April 2026 that Pichai said 75 percent of Google's new code is AI-generated. The article gives 25 percent for 2024 and 50 percent for the previous autumn. It does not say whether the statement was made on an earnings call, in a blog post or in another setting, and Alphabet's own transcript was not confirmed as the source. The figure is a company claim and is not independently checked. [2]

  3. In progress · Oct 2023AI agents resolve real software issues on a standard benchmark

    On the original SWE-bench test set of 2,294 problems drawn from real GitHub issues, the best model reported at publication, Claude 2, resolved 1.96 percent of issues. I could not open the official leaderboard to cite a current top score. Liang and colleagues argue that some benchmark results may partly reflect memorisation rather than reasoning. [4][5]

  4. In progress · Jul 2025Controlled studies show faster completion by experienced developers

    In METR's 2025 trial of 16 experienced open-source developers on 246 tasks, allowing AI tools increased completion time by 19 percent (confidence interval +2 to +39 percent). Developers had predicted a 24 percent speedup beforehand and estimated a 20 percent speedup afterwards. METR's February 2026 follow-up, with 57 developers, reported speedups of 18 percent for returning developers and 4 percent for new developers, and both confidence intervals include zero. METR says its raw results show some evidence for speedup, calls its data only very weak evidence for the size of the 2025 increase, and says the central estimate is likely a lower bound because of selection effects. [6][7]

  5. Not startedSecurity and maintenance of AI-written code shown to match human-written code

    In a controlled user study, Perry and colleagues (CCS 2023) found that participants with an AI assistant wrote significantly less secure code than participants without one. DORA's 2024 report, published 23 October 2024, estimated that as AI adoption increased, delivery throughput fell by 1.5 percent and delivery stability by 7.2 percent. These are estimated associations, not causal effects. A 2026 preprint found that 1,293 of over 33,000 curated agentic pull requests on GitHub were security-related and that these had lower merge rates and longer review latency. That shows how reviewers handle the changes, not whether the code is more flawed. No source shows that AI-written code matches the security or maintenance record of human-written code. [8][9][10]

  6. Not startedMost software written by AI

    No source measures the share of all software, or of all software projects, that is written by AI. The available figures cover new code at a few large companies. Reaching most software would also need the code to be reviewed, tested and maintained at a standard that the sources reviewed have not shown. [1][3]

Why it matters

If most new code is AI-written, the main work of software teams shifts towards specifying, reviewing and testing code rather than typing it. Flaws in generated code could then appear in many systems at once. Who is accountable for that code, what it costs to maintain, and which skills teams need all depend on how much AI writes and how well that code holds up.

What stands in the way

  • The share-of-code figures come from company statements. Neither Google nor Microsoft has published how it defines or counts code in the sources reviewed, so the numbers cannot be compared with each other or checked.
  • Benchmark results may be inflated by memorisation of benchmark content, which Liang and colleagues argue applies to SWE-bench Verified.
  • The only controlled productivity trial with a clear result found a slowdown, and the 2026 follow-up is too weak to show a speedup. Without a clear measure of productivity, it is unclear how much AI-written code saves.
  • Security evidence is thin. The controlled study tested 2022 to 2023 tools, and the 2026 preprint measures how reviewers handle agent-written pull requests, not whether the code is more flawed.
  • According to The Register's report of Microsoft's remarks, AI works best on entirely new code and is less effective at reworking older code.

Signals to watch

  • Companies publishing the method behind share-of-code figures, including what counts as code and as AI-generated.
  • A replacement for SWE-bench Verified that is shown to resist memorisation.
  • A redesigned controlled productivity study with a clear, pre-stated analysis plan.
  • Security audits of agent-written code merged into large open-source projects.
  • Updates to DORA's annual findings on AI adoption and software delivery stability.

Sources

  1. company announcement
    Alphabet Q3 2024 Earnings Call (prepared remarks)
    Alphabet Inc., 29 Oct 2024.
  2. news
    Google CEO says 75% of company's new code is AI-generated
    Semafor, 24 Apr 2026.
  3. news
    Microsoft and Meta autocoding
    The Register, 30 Apr 2025.
  4. preprint
    SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
    Carlos E. Jimenez, John Yang, Alexander Wettig et al.. arXiv, Oct 2023.
  5. preprint
    The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reason
    Shanchao Liang, Spandan Garg, Roshanak Zilouchian Moghaddam. arXiv, Jun 2025.
  6. preprint
    Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
    Joel Becker, Nate Rush, Elizabeth Barnes et al.. arXiv, Jul 2025.
  7. other
    We are Changing our Developer Productivity Experiment Design
    Joel Becker, Nate Rush, Tom Cunningham et al.. METR, 24 Feb 2026.
  8. peer reviewed paper
    Do Users Write More Insecure Code with AI Assistants?
    Neil Perry, Megha Srivastava, Deepak Kumar et al.. arXiv (published at ACM CCS 2023), Nov 2022.
  9. other
    Announcing the 2024 DORA Report
    Google Cloud (DORA), 23 Oct 2024.
  10. preprint
    Security in the Age of AI Teammates: An Empirical Study of Agentic Pull Requests on GitHub
    Mohammed Latif Siddiq, Xinye Zhao, Vinicius Carvalho Lopes et al.. arXiv, Jan 2026.

Last reviewed 21 Oct 2026.