Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
No one has directly measured the share of code written by AI outside GitHub, so there is no reliable single percentage for private repositories, other hosting platforms, or software built outside public code archives. The closest broad figure is a 2026 self-reported survey by JetBrains. Among professional developers who answered, the average share of their work code from the previous month was about 47% fully generated by AI agents and about 38% written by the developer with some AI assistance. Those are reported experiences, not a census of codebases, and they cannot be added together into an “85% AI-written” total.
Why there is no single figure
Every published percentage answers a slightly different question. Some ask developers what they produced last month. Some ask what share of a startup’s existing codebase came from AI. Some count what developers commit. Some classify code found in public GitHub projects, and one reports what an AI model authored inside a single company. Outside GitHub, the only sources that can see code at all are the people who wrote it, or the organizations that hold it, so the figures that exist are surveys and internal company accounts. Classifiers that infer authorship from code artifacts work only where the code is visible, which is largely public repositories.
Figures from the main sources
The table below lists each source with what it counted. Read each row as a separate measurement.
| Source and date | Population | What was counted | Reported figure | Evidence type |
|---|---|---|---|---|
| JetBrains Research, Developer Ecosystem Survey 2026 (fielded May–July 2026) | More than 15,000 professional developers worldwide, reweighted to the global developer population; roughly 90% in developer, programmer or software engineer roles | Share of work code produced in the previous month that was fully agent-generated, written with AI assistance, or fully manual | Averages of about 47% agent-generated, 38% AI-assisted and 27% fully manual | Self-reported survey using percentage bands |
| Supabase, State of Startups 2026 | Surveyed startups | Share of the startup’s existing codebase written by AI | 61% said more than half; 40% said 76–100%; 2% said zero | Survey of startup respondents; sample size and sampling method not stated on the publisher page reviewed |
| Sonar, 2026 State of Code Developer Survey (summary dated January 8, 2026) | Developers surveyed | Share of code they commit that is AI-generated or AI-assisted | 42%. Separately, 38% said reviewing AI-generated code took more effort than reviewing colleagues’ code | Self-reported survey; combines generated and assisted code |
| Science study (published 2025) | More than 30 million GitHub commits by 160,097 developers in six countries, 2019–2024; the figure given concerns Python functions in the United States | Functions that a classifier estimates were written by AI | 29% of Python functions in the United States, estimated | Classifier inference from public GitHub code |
| Anthropic company report (May 2026) | Code merged into Anthropic’s own codebase | Code merged that was authored by Claude | More than 80% | Company-reported internal figure |
| GitHub with Wakefield Research (fielded February 26–March 18, 2024) | 2,000 non-student, non-manager respondents at companies with at least 1,000 employees; 500 each in the U.S., Brazil, Germany and India | Use of and perceptions about AI coding tools at work | More than 97% had used AI coding tools at work at some point. No share of code generated is reported | Adoption survey |
Reading the JetBrains averages correctly
JetBrains is the broadest of these measures, so it is the one most often quoted. Its respondents chose from percentage bands (0%, 1–20%, 21–40%, and so on up to 100%, plus “I don’t know”), not exact numbers. To compute averages, the publisher used the midpoint of each band. The result is an approximation, and the publisher’s methodology notes say so directly:
#1 Best Overall
“The averages across the three categories of how code is written within the same group (e.g. seniors) could exceed 100% because of the bucketed nature of the answers, and respondents’ self-reports may not always be fully accurate.”
That is why the three averages sum to more than 100%. Keep these rules when you use the figures:
- Use “roughly 47%” and “roughly 38%” as separate categories. The first means the code was generated by an agent with no developer editing; the second means the developer wrote it with AI help.
- Do not add them into a combined share. The categories were measured separately, and the total is not a measured quantity.
- State the unit: work code from the previous month, as reported by the developer, not code in a repository or a company’s total output.
Startup, committed-code and company figures
Startup codebases
The Supabase figure measures a different thing from the JetBrains one. It asks about the share of an entire codebase, which accumulates over the life of a product, so a high figure says less about current daily output than about how much of the existing software an AI tool has touched. It describes a group of startups that answered the survey. It is not an audit of any startup’s code and should not be generalized to all startups or all software teams.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Committed code
The Sonar figure concerns code that developers commit, and it counts AI-generated and AI-assisted code together. Because it mixes the two categories, it cannot be compared directly with JetBrains’ agent-only figure. The review-effort result is a separate point: it is a developer perception about the cost of checking AI output, not a measurement of code quality.
Company-reported code
Anthropic reports that more than 80% of code merged into its own codebase, as of May 2026, was authored by Claude. This is an internal accounting for one company, and it reflects that company’s tools and workflows. It is useful as an example of how far AI authorship can go in one organization, but it is not an industry-wide estimate.
What GitHub-based estimates can see
The Science study is the main source that measures code directly rather than through self-reports, and it is the most useful for understanding the limits of “outside GitHub” estimates. Its classifier looks at public code in GitHub projects. A classifier can only estimate authorship where it has code to examine, so its result describes the studied GitHub projects and languages. It does not describe private repositories, code on other hosting platforms, or software that never reaches a public archive. The 29% figure for Python functions in the United States should be read with those boundaries in mind.
Adoption figures are not volume figures
A common mistake is to treat a survey about tool use as evidence of code volume. The 2024 GitHub and Wakefield Research survey found that more than 97% of enterprise respondents had used AI coding tools at work at some point. That measures how many people have used the tools, not how much of their code the tools wrote.
Code volume also does not measure value. Anthropic itself cautions that “Lines of code is an imperfect measure, as it measures quantity over quality.” A high share of AI-written lines does not show that the software is better, worse, or cheaper to produce.
Best Value
How to check a code-share claim
When you see a percentage of AI-generated code, verify it with these steps before repeating it:
- Identify the population: professional developers, startup teams, enterprise respondents, one company, or a set of public repositories.
- Identify the unit: recent work output, committed code, lines or functions in a repository, or an existing codebase.
- Check the definition: does “AI-generated” mean agent-only output, or does it include AI assistance, suggestions and edits?
- Check the period: previous month, an ongoing codebase, a survey fielding window, or historical commits.
- Check the evidence type: self-report, classifier inference, or internal company accounting.
- Check the coverage: languages, geography, public or private code, and company size.
If a claim cannot answer most of these questions, treat it as a rough signal rather than a measurement.
Where the evidence stops
- No audited, universal figure for AI-written code outside GitHub has been published. The available numbers come from surveys and company reporting.
- Survey figures depend on what respondents know about their own code, and the JetBrains averages depend on band midpoints.
- The Science figure is taken from the study’s published abstract, and it is limited to GitHub projects and to the populations and languages it covers.
- Any combined “AI-written” total across the sources is not a measured quantity, because each source uses a different population and definition.
What the evidence does support is narrower and more useful: a large share of developers now report that a substantial part of their recent work code was produced with AI, and the way that share is counted determines what the percentage means.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
