What Percentage of Software Is AI-Generated in 2026?

Researchers and companies are measuring narrower things instead: AI-generated code, AI-assisted code, accepted suggestions, agent activity, code changes inside particular companies, and developer adoption. These measurements are useful, but they are not interchangeable.

This article examines the strongest available evidence as of September 17, 2026 and separates what is measured from what remains unknown.

1. The Short Answer

We do not currently know what percentage of all software in existence is AI-generated in 2026.

No global dataset tracks every line of software written by companies, governments, open-source projects, students, hobbyists, and private developers and records how each line was created.

What we can measure is much narrower. Several 2026 studies and company reports show substantial AI involvement in new code production:

• GitLab reports 34% AI-generated, 37% written from scratch, and 29% copied from another source in its code-source measurement.^[1]
• JetBrains reports an average of approximately 47% of developers' code as fully generated by AI agents, with another roughly 38% written with some AI assistance.^[2]
• Google reported that 75% of its new code was AI-generated and approved by engineers in April 2026.^[3]
• Stack Overflow found that 59% of respondents were using AI agents at work at some frequency in its April 2026 pulse survey.^[4]
• Anthropic found that roughly 56% of Claude Code sessions involved writing, fixing, testing, or orchestrating code-related work.^[5]

These are different measurements, not competing estimates of one global percentage.

The strongest defensible conclusion is: AI is responsible for a substantial share of new code production in several measured populations and organizations, while the AI share of all software in existence remains unknown.

2. Why There Is No One Universal Percentage

"AI-generated software" can refer to several different measurements.

AI-generated code

AI directly produces an implementation, such as a function, component, test, or file change.

AI-assisted code

A developer writes code while receiving suggestions, corrections, explanations, or partial implementations from AI.

AI-modified code

Code originally written by a person is later changed by an AI system, such as during refactoring, debugging, migration, or rewriting.

AI-assisted development

AI can also help with planning, documentation, debugging, code review, testing, research, or understanding an unfamiliar system without necessarily contributing many lines to the final source code.

That is why a percentage of AI-generated code is not the same thing as a percentage of software development performed by AI.

3. How Much Code Do Developers Say AI Generates?

JetBrains' 2026 Developer Ecosystem Survey asked more than 15,000 professional developers worldwide in May–July 2026 what percentage of the code they had produced for work the previous month was fully generated by AI agents, AI-assisted, or fully manual.^[2]

Respondents selected percentage ranges, and JetBrains used each range's midpoint to calculate averages. The resulting figures were approximately:

• 47% fully generated by AI agents
• 38% written with some AI assistance
• 27% fully written manually

These figures should not be added together. JetBrains explains that the separate midpoint calculations can produce totals above 100%, and the results are self-reported estimates.

The survey also found that more than half of developers reported writing less than 20% of their code manually, roughly one in five reported writing zero code without AI assistance, and about 22% relied on agents for more than 80% of their code.^[2]

The overall picture is therefore one of widespread but uneven AI use: developers are not all delegating the same amount of coding work to agents.

4. How Many Developers Use AI Coding Tools?

Adoption is easier to measure than the total percentage of AI-generated software.

JetBrains' January 2026 AI Pulse survey, covering more than 10,000 professional developers worldwide, found that 90% regularly used at least one AI tool at work for coding and development tasks, while 74% had adopted specialized developer AI tools.^[6]

Its later May–July 2026 Developer Ecosystem Survey found that 90% of professional developers were using AI coding agents at work at least weekly, with 68% using them daily.^[7]

Adoption, however, does not tell us how much code AI produces. A developer can use an AI tool every day while still writing most implementation code manually.

Adoption measures how widespread AI use is; code-share measurements estimate how much code developers say AI produces.

5. GitLab's 2026 Measurement of AI-Generated Code

GitLab's 2026 Global DevSecOps Report surveyed 3,266 DevSecOps professionals.^[1]

Its code-source measurement reports:

• 34% AI-generated
• 37% written from scratch
• 29% copied from another source

The categories total 100%, making this one of the more direct code-source measurements available for 2026.

But the result describes the respondents and the code they work on; it is not a census of all software produced worldwide. It also measures code source rather than the percentage of complete applications that are AI-generated.

6. Google Reports 75% of Its New Code Is AI-Generated

In April 2026, Google CEO Sundar Pichai said that 75% of all new code at Google was AI-generated and approved by engineers.^[3]

Google said that figure had increased from 50% the previous fall.^[3]

This measures new code inside Google, not every line in its existing codebases and not the software industry as a whole.

The engineer-approval language is also important: the reported figure is not a claim that 75% of Google's software is produced autonomously.

7. AI Agents Are Common, but Full Autonomy Is Different

Stack Overflow's April 2026 pulse survey received responses from 1,100 developers and working professionals.^[4]

59% said they were using AI agents at work at some frequency, compared with 31% in Stack Overflow's 2025 Developer Survey.^[4]

At the same time, 63% rarely or never let agents run entirely on autopilot. The survey also found that 60% blocked agents from making unapproved system changes, while 68% preferred predictable single-agent setups over more complicated multi-agent configurations.^[4]

These findings show why agent adoption should not be confused with full autonomy. AI can be heavily integrated into development while humans retain review and approval responsibilities.

8. What Happens During Real AI Coding Sessions?

Anthropic analyzed approximately 400,000 Claude Code sessions from about 235,000 people, covering October 2025 through April 2026.^[5]

About 56% of sessions involved writing code, fixing code, testing code, or orchestrating code-related work. The remaining sessions involved activities such as operating software, planning, exploring existing systems, analyzing data, or producing documents.^[5]

This is a measure of session activity, not a percentage of AI-generated source code.

Anthropic also found that people generally made most planning decisions while Claude handled more execution decisions, illustrating another distinction between AI-assisted development and fully autonomous software engineering.^[5]

9. Production-Scale Evidence From Microsoft Research

Microsoft Research analyzed sampled GitHub Copilot traces from June 2026 covering approximately:

• 3.2 million users
• 13 million sessions
• 761 million LLM calls
• 95 trillion tokens

This is telemetry from production use of a coding-agent system rather than a developer opinion survey.^[8]

The researchers studied the characteristics of agentic coding workloads; they did not calculate what percentage of source code was generated by AI.^[8]

The scale of the dataset therefore demonstrates extensive AI-agent activity, but cannot be converted into a statement such as "X% of GitHub code is AI-generated."

10. How Much Control Are Developers Giving AI?

A Microsoft Research study published in July 2026 examined 448 professional developers at Microsoft and found that acceptance of AI-generated work varied by task. Developers generally accepted AI working under human oversight, with lower acceptance for some design-oriented and human-facing tasks.^[9]

Another Microsoft Research study presented at the 2026 International Conference on Software Engineering used paired interviews with 54 developers across 27 teams and found substantial differences in how developers used generative AI tools.^[10]

These studies do not measure AI-written code. They address a different question: how much autonomy developers are willing to give AI during software work.

11. How Companies Measure AI-Generated Code

GitHub's Copilot documentation illustrates why "AI-generated code" is not a single platform metric.

GitHub separately tracks:

• code suggestions and accepted suggestions
• lines suggested and lines actually added
• agent-initiated and user-initiated code changes
• pull-request activity

GitHub defines an agent-initiated code change as lines automatically added to or deleted from the codebase by an agent on behalf of a user. It separately records user-initiated changes made through completions, chat, and other Copilot interactions.^[11]

The distinction matters because a displayed suggestion, an accepted suggestion, and an agent-made repository change are different events.

12. Why Attribution Is Becoming a Problem

GitLab's 2026 AI Accountability Report surveyed 1,528 DevSecOps professionals and technology buyers across six countries.^[12]

43% said their organizations could not reliably distinguish AI-generated code from human-written code in their own codebases. Another 39% cited systems that do not track code origin as a barrier to control and traceability.^[12]

GitLab also reported that 85% of respondents agreed AI had shifted the bottleneck from writing code to reviewing and validating it.^[12]

Together, these findings highlight two measurement problems: organizations may not know the origin of all existing code, and code generation is only one part of the surrounding engineering workflow.

13. What About Open-Source Software?

Microsoft Research's March 2026 study examined self-admitted use of generative AI in open-source projects.^[13]

The researchers mined GitHub for signs of ChatGPT and GitHub Copilot use in commits, issues, and pull requests, then manually coded the evidence into a taxonomy covering 64 AI-use tasks grouped into seven categories.^[13]

The study shows that AI contributions can leave traces in software artifacts beyond simple code-generation counters.

It still cannot establish the percentage of all software worldwide that is AI-generated because public repositories represent only part of software development and substantial commercial and internal code remains private.

14. Why AI-Generated Code Is Easier to Measure Than AI-Generated Software

Imagine an application has one million lines of existing code. Developers add another 100,000 lines during 2026, and AI generates 70,000 of those lines.

AI generated approximately 70% of the new code, but the AI-generated share of the entire codebase would be much smaller because most of the code predates that work.

Real software is more complicated still: old code is modified, rewritten, deleted, copied, generated, reviewed, and regenerated over time.

That creates at least three distinct questions:

1. What percentage of new code is generated by AI?
2. What percentage of an organization's current codebase originated from AI?
3. What percentage of the overall software-development work is performed with AI assistance?

A number answering the first question cannot simply be used as the answer to the second or third.

15. AI-Generated Code Is Not the Same as AI-Generated Software

Software development includes much more than writing source code: requirements, design, architecture, testing, debugging, security, deployment, monitoring, maintenance, incident response, and product decisions.

AI can contribute to several of these activities without those contributions appearing as AI-generated source code.

Microsoft Research's April 2026 study of 860 Microsoft developers examined where developers wanted AI support beyond traditional coding assistance. The researchers found demand for AI to absorb surrounding assembly work while maintaining boundaries around higher-level human responsibilities.^[14]

This reinforces the distinction between code production and software engineering as a whole.

16. Historical Context: How the Measurements Have Changed

Historical comparisons are useful, but these figures do not form a single continuous time series. The sources measure different things, populations, and periods.

The figures provide historical context, but they should not be combined into a single growth curve.

17. What We Can Actually Estimate for 2026

We can measure AI-generated code inside specific populations

JetBrains and GitLab provide survey-based estimates of how much code developers report AI generating or assisting with.

We can measure new-code production inside companies

Google reported that 75% of its new code was AI-generated and approved by engineers in April 2026.

We can measure AI-agent activity at very large scale

Microsoft Research's Copilot analysis measured millions of users and millions of coding-agent sessions.

We can measure how AI is used during software work

Anthropic's Claude Code analysis provides evidence about the kinds of activities agents are performing.

We cannot measure all software globally

No trustworthy dataset identified in the research reviewed here covers every private codebase, open-source project, internal tool, application, embedded system, government system, student project, and legacy system.

18. What the Numbers Do Not Tell Us

They do not tell us whether AI-written code is good

The percentage of generated code says nothing by itself about reliability, security, maintainability, or correctness.

They do not tell us how much human work disappeared

A high percentage of AI-generated code does not by itself tell us how much developer work was eliminated or shifted elsewhere in the workflow.

GitLab reported that 85% of respondents agreed AI had shifted the bottleneck from writing code to reviewing and validating it.^[12]

They do not tell us how much of a software product was created by AI

A code-generation statistic usually refers to a particular time period or set of changes. It does not automatically describe the full history of a software product.

They do not tell us how much AI contributed to decisions

Planning, architecture, debugging, research, review, and other activities can involve AI without appearing as AI-generated source code.

19. Are Companies Measuring This?

Increasingly, yes.

GitHub exposes separate Copilot metrics for generated output, accepted suggestions, lines added, agent-initiated changes, pull requests, and other forms of activity. GitLab is researching AI-code governance and provenance, while Google has published a direct internal measurement for new code.^[11]

At the same time, GitLab found that 43% of respondents could not reliably distinguish AI-generated code from human-written code in their own codebases.^[12]

That means even organizations using detailed tooling may not have a complete historical record of where every line originated.

20. The Most Accurate Answer for 2026

So what percentage of software is AI-generated in 2026?

The answer depends on which software and which measurement you mean.

For all software in existence worldwide: unknown.

No credible source currently provides a global percentage for the AI authorship of every existing software codebase.

For newly produced code: the available 2026 evidence shows that AI is responsible for a substantial share in several measured populations and organizations, with some reporting majority-AI production.

The strongest 2026 measurements include:

• 34% AI-generated in GitLab's code-source measurement
• approximately 47% fully agent-generated in JetBrains' self-reported developer code-production averages
• over 60% of the code committed to Braze's main repositories was reported by its CTO as written with AI
• 75% of new code reported as AI-generated at Google

These are not four estimates of one global percentage. They describe different populations, workflows, companies, and methodologies.

21. What We Know, What We Can Estimate, and What Remains Unknown

What we know

• AI coding tools and agents are used by a large share of professional developers.
• AI-generated code represents a substantial share of newly produced code in several 2026 measurements.
• Some companies have publicly reported that AI generates a majority of new or committed code within their organizations.
• Human review and decision-making remain important parts of many professional workflows.

What we can estimate

• How much AI-generated code developers report producing.
• How much new code is generated by AI inside companies that disclose internal measurements.
• How common AI coding agents are.
• How developers use AI during different parts of software development.

What remains unknown

• The percentage of all software in existence that is AI-generated.
• The exact percentage of all newly created software worldwide that is AI-generated.
• A universal authorship standard for code that is repeatedly generated, edited, copied, and refactored by both people and AI.
• The percentage of the overall software-development process that is performed by AI rather than humans.

22. Conclusion

In 2026, AI is a substantial part of modern software development. Surveys, company disclosures, and large-scale telemetry all show significant AI involvement in coding work.

But the evidence does not justify assigning one percentage to all software.

JetBrains measures developers' reported code production. GitLab measures reported code-source shares. Google measures its own new code. GitHub measures Copilot activity through several separate metrics. Anthropic and Microsoft Research measure AI coding activity at large scale.

AI is generating a substantial share of newly produced software code in several measured populations and organizations in 2026, and some organizations report that AI generates most of their new code. But nobody currently has a reliable percentage for how much of all software is AI-generated.

"75% of new code at Google is AI-generated" is a specific company-reported measurement. "75% of software is AI-generated" would be a different claim, and the available evidence does not support it.

The most useful way to think about AI-generated software in 2026 is therefore as a set of measurable layers: how much developers use AI, how much code AI produces, how much of that output is accepted or changed, how much new code companies generate with AI, and how much of the broader engineering process AI touches.

We can measure parts of that picture increasingly well. The global percentage of all software that is AI-generated remains unknown.