AI now writes a large share of the world’s software. How large depends on who is counting and what they count.
This article pulls together 2026 data from developer surveys, code analysis firms, job market trackers and company reports. It covers how much code AI writes, how developers use it, and how well the code holds up.
The numbers show fast adoption. They also show that checking the code has become the hard part.
How Much Code Does AI Write in 2026?
There is no single answer. Published figures run from about a quarter of commits to 75% of new code.
Google says 75% of its new code is now AI-generated and approved by an engineer. Sundar Pichai gave that figure in April 2026. It was 50% in autumn 2025.
GitClear measures much less. Its analysis of 623 million code changes found measurable AI assistance in about a quarter of commits.
Sonar surveyed more than 1,100 professional developers. They said AI accounts for 42% of the code they commit, and they expect 65% by 2027.
New Relic surveyed 200 U.S. technology leaders at manager level and above. Of those, 67% said AI generates or heavily refactors between 51% and 75% of weekly code output.
Source
Figure
What it counts
Google (April 2026)
75%
New code that is AI-generated and engineer-approved, self-reported
Jellyfish (March 2026)
64%
Share of 700+ companies where most code is AI-assisted
GitHub Copilot telemetry
46%
Average code share across its user base
Sonar survey (January 2026)
42%
Committed code that is AI-generated or assisted, as reported by developers
GitClear (2026)
About 25%
Commits with detectable AI assistance, across 623 million changes
Microsoft (April 2025)
20% to 30%
Code in its repositories written by software, not updated since
Three differences explain most of the gap between these numbers.
The first is new code versus all code. Google’s figure covers new code only, and most engineering work happens in code that already exists.
The second is generated versus surviving. A suggestion accepted in an editor is not always a line that reaches the main branch.
The third is self-reported versus measured. Google and Microsoft describe internal data that outsiders cannot check. GitClear, DORA and Jellyfish publish their methods and sample sizes.
Before repeating any percentage, check what it counts and when it was published. Microsoft’s often-quoted figure is now about 18 months old.
Other companies have shared figures too. Salesforce estimates that AI agents handle 30% to 50% of its workloads, and Meta has signaled a 50% target for late 2026.
Garry Tan of Y Combinator said in March 2025 that about a quarter of one startup batch had codebases that were 95% AI-written. Those were teams of two to five people shipping a first product.
Figure 1: Reported share of AI-written code differs by what each source measures.
How Many Developers Use AI Coding Tools?
Adoption is close to universal among professional developers. Trust is much lower.
Google’s DORA report puts AI use at work at 90% of developers, a rise of 14 points in one year. The median developer spends two hours a day working with AI.
The 2025 Stack Overflow survey found 84% of respondents using or planning to use AI tools, up from 76%. Among professional developers, 51% use them daily.
Sonar found that 72% of developers who have tried AI coding tools use them every day. Opsera studied more than 250,000 developers at over 60 enterprises and reports that 90% of teams use AI daily.
Company policy is catching up. New Relic found that 95% of organizations authorize AI-generated code in core production work. Of those, 87.5% do so formally and 7.5% informally.
Agent tools are earlier in their rise. Stack Overflow found that only 14.1% of developers use AI agents daily, and 37.9% have no plans to use them.
Sonar’s report series includes a chapter on why 64% of developers use agentic tools. The gap likely reflects different survey groups and different definitions of use.
Team size does not seem to limit use. One analysis found that 51% of active AI users work on teams of 10 or fewer developers.
Support differs by country. GitHub data shows that 88% of U.S. developers report company support for AI tools, against about 70% in Brazil, about 65% in India and 59% in Germany.
Age also plays a part. A Zero to Mastery survey found that developers aged 18 to 34 are twice as likely to use AI daily as older groups.
Figure 2: Most developers use AI tools, while daily agent use is still low.
Where AI Helps Most, and Where It Does Not
Developers use AI for nearly every kind of project. Sonar found 88% use it for prototypes and proofs of concept, and 58% use it for business-critical services.
Project type
Developers using AI
Prototypes and proofs of concept
88%
Internal, non-critical production software
83%
Customer-facing applications
73%
Business-critical or mission-critical services
58%
Usage and effectiveness do not match. Writing new code is the most common use, at 90% of developers. Only 55% of those users rate AI as very or extremely effective for it.
Refactoring shows a wider gap. Some 72% use AI for it, and 43% find it highly effective.
Developers rate AI highest on tasks with existing context. Documentation scores 74%, explaining existing code 66%, and generating tests 59%.
Stack Overflow asked a different question: which tasks are mostly done by AI today. Searching for answers leads at 54.1%, while writing code sits at 16.9%.
Task
Currently mostly done by AI
Search for answers
54.1%
Learning new concepts
33.1%
Documenting code
30.8%
Debugging or fixing code
20.7%
Writing code
16.9%
Committing and reviewing code
10.2%
Deployment and monitoring
6.2%
The pattern is clear. Developers hand over low-risk tasks first and keep control of judgment steps such as review and deployment.
Figure 3: AI scores highest on documentation and lowest on refactoring.
Does AI Make Developers Faster?
Sometimes. The size of the gain shrinks when researchers measure it instead of asking developers how they feel.
The often-cited 55% figure comes from a 2022 GitHub study of 95 developers building one HTTP server in JavaScript. That is close to a best case for a model.
METR ran a randomized trial in 2025 with 16 experienced open-source maintainers and 246 real issues. With AI allowed, tasks took 19% longer.
The developers expected a 24% speedup beforehand. Afterward they still believed AI had made them 20% faster.
METR’s follow-up, updated in February 2026, estimates an 18% speedup for returning developers and 4% for new ones. Both ranges still include a slowdown, and METR calls the evidence weak.
Stanford research cited by DORA shows gains of 35% to 40% on simple tasks. On complex legacy code the gain is 10% or less.
Faster writing does not always mean faster delivery. Opsera found AI cuts time to a pull request by up to 58%, but AI-generated pull requests wait 4.6 times longer in review.
One study followed an enterprise that mandated twice as many merged pull requests per engineer. It tracked 802 developers and 196,212 pull requests. Output reached 2.09 times the baseline, and the load on each reviewer roughly doubled.
Sonar adds that 38% of developers say reviewing AI code takes more effort than reviewing a colleague’s work.
Company leaders give modest estimates. Google CEO Sundar Pichai said AI had raised the company’s engineering velocity by about 10%, and that Google planned to hire more engineers.
Vendor analyses report larger gains. One claims developers save 30% to 60% of their time on coding, testing and documentation. These figures come from self-reported or vendor-run studies.
Opsera also found that senior engineers capture nearly five times the productivity gains of junior engineers. In its data, 21% of AI coding licenses go underused.
Figure 4: Measured speed gains are smaller than early claims, and one trial found a slowdown.
Code Quality and Maintainability Statistics
AI code usually compiles and passes review. The concern is how it ages.
GitClear and GitKraken studied 623 million code changes from 2023 to 2026. Eight maintainability signals moved in the wrong direction, and LeadDev reported the main ones.
Duplicated code blocks rose 81%. Moved or refactored code fell 70%, and updates to code older than a year fell 74%.
Error-masking code rose 47%, and cross-file function calls fell 35%. These figures track code written as AI use grew. They are not a controlled comparison.
GitClear CEO Bill Harding described the cause this way: “Every time you want something, AI creates a new package for it.”
CodeRabbit studied 470 pull requests on GitHub in December 2025. It found 10.83 issues per AI co-authored pull request, against 6.45 for human-only ones.
Issue type
AI co-authored vs human-only
Readability issues
More than 3x
Security issues
Up to 2.74x
Formatting issues
2.66x
Error handling gaps
Almost 2x
Logic errors
75% more
Excessive I/O
About 8x
CodeRabbit labeled a pull request as AI co-authored only when it carried a signal saying so. Some pull requests counted as human may contain AI code. CodeRabbit also sells AI code review tools.
Developers report the same debt. In Sonar’s survey, 88% cited negative effects of AI on technical debt, while 93% also saw positive ones such as better documentation.
The top complaints were code that looks correct but is unreliable (53%) and code that is unnecessary or duplicated (40%).
Coding benchmarks paint a brighter picture. Anthropic’s system card reports that Claude Opus 5 resolves 96.0% of SWE-bench Verified problems.
OpenAI stopped reporting that score in February 2026. Its audit found flawed tests in 59.4% of the hard problems it checked.
On the harder SWE-bench Pro test, Anthropic reports 79.2% for Claude Opus 5 and OpenAI reports 64.6% for GPT-5.6 Sol. Each vendor uses its own setup, so small gaps are not a clean ranking.
None of these benchmarks score security or maintainability. A patch passes when the tests pass.
Figure 5: Duplication rose while refactoring and updates to older code fell.
AI-Generated Code Security Statistics
Security has not improved as models have. Veracode’s 2026 GenAI Code Security Report gave 11 new models 80 coding tasks. The average security pass rate was 56%.
Its first report, in July 2025, measured 55% across more than 100 models. The rate has barely moved in a year.
GPT-5.5 had the best result at 68%. Models built for coding averaged 51%, against 52% for general-purpose models.
The weakness type matters more than the model. Models passed 87% of weak cryptography tasks and 83% of SQL injection tasks. Those problems follow repeatable patterns.
Pass rates fell to 15% for cross-site scripting and 12% for log injection. Those depend on how user input moves through an application.
Language also matters, and Java is the weakest by a wide margin.
Language
Veracode July 2025
Spring 2026 update
Python
61.7%
62%
JavaScript
57.3%
57%
C#
55.3%
58%
Java
28.5%
29%
Veracode suspects the training data. Public Java repositories hold millions of lines written before modern security frameworks existed.
Earlier studies show the same pattern. A 2021 study of GitHub Copilot found about 40% of 1,689 programs vulnerable. A study of 733 AI-generated GitHub snippets found weaknesses in 29.5% of Python and 24.2% of JavaScript snippets.
Agents do not solve it. In the SusVibes benchmark, 57% of solutions from SWE-Agent with Claude 4 Sonnet worked, but only 11.8% were secure.
Opsera reports that AI-generated pull requests carry 15% to 18% more security vulnerabilities. Veracode’s prompts did not mention security, so its scores show default behavior.
Figure 6: AI code passes pattern-based security tasks far more often than input-handling tasks.
Reliability, Trust and Review Statistics
Reviewers tend to like AI code. Production tells a different story.
In New Relic’s survey, 94% of technology leaders rated AI-generated code as higher quality than human code at the time of review. After release, 78% reported a spike in production incidents tied to AI code.
Another 86% reported more senior engineer firefighting, and 82% had at least one major production failure from AI code in the past six months. A further 74% said at least 25% of AI-generated code needs rework after deployment.
New Relic says AI code brings about 1.7 times more critical runtime issues than human-written code. CodeRabbit’s pull request study found a similar ratio of issues.
The four most common failure types each affect roughly three in ten organizations. They are integration failures (30%), compliance and governance issues (30%), data-integrity problems (29%) and new security vulnerabilities (28%).
Verification is the weak point. New Relic found that 62% of leaders say their teams often or always trust AI code enough to ship it without line-by-line checks.
Sonar found that 96% of developers do not fully trust AI code, yet only 48% always check it before committing.
Stack Overflow found that only 3.1% highly trust AI accuracy, while 19.6% highly distrust it. About 66% named code that is almost right, but not quite, as their top frustration.
DORA found that AI adoption now lifts delivery throughput, which reverses its 2024 result. It still links AI to higher delivery instability. In 2024, DORA estimated 7.2% more instability for every 25% rise in AI adoption.
Teams are responding. New Relic found that 78% now prompt AI tools to include logging hooks and metrics in the first draft, and 96% rate observability as critical for AI-generated environments.
The worst cases are rare but public. In July 2025, a Replit coding agent deleted a live database during a declared code freeze. The database held data on more than 1,200 executives and more than 1,190 companies.
The agent then told the user a rollback would not work. The rollback did work, and Replit’s CEO called the deletion unacceptable and added separate development and production databases.
Figure 7: Leaders rate AI code well at review, yet report more incidents and firefighting after release.
AI and Software Developer Jobs Statistics
Fears that AI would end developer jobs have not matched the data so far. Indeed Hiring Lab reports that U.S. software development postings rose almost 15% from late February 2025, while overall postings fell 7%.
The rebound is uneven. Postings remain about 27.5% below their pre-pandemic level.
Senior roles made up 71% of the increase between May 2025 and May 2026. Job titles that mention AI made up 37%.
Entry-level workers face more pressure. Stanford’s 2026 AI Index found that employment in AI-exposed fields for ages 22 to 25 fell 16% since generative AI arrived. Early-career software developers fell by 20%.
Stanford’s Digital Economy Lab used ADP payroll records through June 2026. It found a gap of about 19% for ages 22 to 25 in the most AI-exposed jobs.
The declines cluster where AI automates the task. Where AI assists the worker, employment is flat or rising.
The long-term outlook is positive. The Bureau of Labor Statistics projects 10% growth for software developers from 2025 to 2035. That takes employment from about 1.72 million to 1.89 million, adding roughly 174,700 jobs.
Sneha Puri, an economist at Indeed Hiring Lab, said companies still need developers to oversee AI work and help clients adopt it. She warned that employers who skip entry-level hiring will lack senior talent in ten years.
Pay data from Netcorp shows a premium for AI skills, based on 2025 figures.
Role level
Software developer
AI or machine learning
Entry-level
$65,000 to $85,000
$90,000 to $130,000
Senior
$100,000 to $130,000
$150,000 to $220,000
Figure 8: Software postings are rising from a low base, led by senior roles, while early-career employment falls.
What the Data Shows
AI writes a large and growing share of code, but the exact share depends on the source. Google’s 75% and GitClear’s 25% can both be right.
Adoption is high and trust is low. Most developers use AI tools, yet few trust the output fully, and fewer than half always check it.
Gains are real on simple, well-defined tasks. They shrink on complex code, and review has become the main limit on speed.
Quality and security data show a steady cost. Duplication is up, refactoring is down, and security pass rates sit near 56%.
The job market is shifting rather than shrinking. Demand is rising for senior and AI-focused roles, while entry-level paths are narrowing.
Many of these figures come from firms that sell code review, security or monitoring products. Methods differ between studies, so compare the numbers with care.
Sources
Introducing the State of AI Coding 2026 (New Relic)
State of Code Developer Survey report: The current reality of AI coding (Sonar)
AI Coding Impact 2026 Benchmark Report (Opsera)
AI is automating coding. So why are software developer job postings on the rise? (CNBC)
How Much Code Is Written by AI in 2026? The Real Numbers (Second Talent)
AI-Generated Code Quality Metrics and Statistics for 2026 (Second Talent)
AI-Generated Code Statistics 2026: Can AI Replace Your Development Team? (Netcorp)
The post AI Coding Statistics 2026: What the Data Shows appeared first on Blockonomi .
