AI Code Statistics 2026
Last updated: 2026-11-045 min read
Sources checked 2026-11-0228 sourced figures
These AI code statistics are the published numbers on AI-assisted code, read through the four metrics a team can compute from its own repository: generation-to-verification ratio, review depth, unverified-merge rate and two-week churn. Every figure keeps its source, year and check date; the reading next to it is Reality Graph’s, not the source’s, and says what the figure implies for a team’s ratio, depth, rate or churn. Where a figure bears on none of the four, the row names the gap it describes instead.
Contents
Adoption and the verification gap
| Figure | Evidence base | Reality Graph reading | |
|---|---|---|---|
| Developers using AI tools in development; the survey reports a 14% rise since it began tracking in 2023 | 84%Stack Overflow survey, checked 2 November 2026 · 2025 | Independent surveyEditorial classification · 2026-11 | Read through the generation-to-verification ratio: adoption at this level raises generated volume for nearly every team at once, while reviewer hours do not grow by themselves. A team that adopted AI should expect its ratio to climb and check whether reviewer time moved with it.Reality Graph's reading, not the source's · 2026-11 |
| Developers who do not fully trust that AI-generated code is functionally correct | 96%Sonar, State of Code, checked 2 November 2026 · 2026 | Vendor studyEditorial classification · 2026-11 | No metric applies: distrust is a stated attitude, and no repository records an attitude. The row describes an awareness gap; the 48% below is the figure that turns into a rate a team can measure.Reality Graph's reading, not the source's · 2026-11 |
| Developers who always check AI-assisted code before committing it | 48%Sonar, State of Code, checked 2 November 2026 · 2026 | Vendor studyEditorial classification · 2026-11 | Read through the unverified-merge rate: if only about half of developers always check, a team without its own validation records should assume a rate near one half, not near zero. The value it measures from recorded evidence replaces that default.Reality Graph's reading, not the source's · 2026-11 |
| Developers for whom reviewing AI-generated code takes more effort than reviewing a colleague's | 38%Sonar, State of Code, checked 2 November 2026 · 2026 | Vendor studyEditorial classification · 2026-11 | Read through review depth: when each change takes more effort to review, depth falls first unless reviewer time grows. Tracking substantive comments per 100 changed lines separately for AI-assisted pull requests shows that drop before it reaches churn.Reality Graph's reading, not the source's · 2026-11 |
Read together, the four rows give a team its starting values before it has measured anything: a generation-to-verification ratio that rises with adoption, and an unverified-merge rate of about one half from the 48%. Both are defaults to replace with the team’s own values. The gap between 96% and 48% itself is unpacked in the verification gap.
The review bottleneck
| Figure | Evidence base | Reality Graph reading | |
|---|---|---|---|
| Pull requests merged per developer in high-AI-adoption teams (2025 dataset; the 2026 dataset reports a much smaller rise) | +98%Faros AI telemetry, 2025 dataset, checked 2 November 2026 · 2025 | Vendor telemetryEditorial classification · 2026-11 | Read through the generation-to-verification ratio: twice the merged volume against unchanged review capacity doubles the ratio, which is the warning threshold the metrics page sets. The comparison that counts is the team's own ratio before and after adoption.Reality Graph's reading, not the source's · 2026-11 |
| Median time a pull request spends in review (2025 dataset; the 2026 dataset is the +441% row) | +91%Faros AI telemetry, 2025 dataset, checked 2 November 2026 · 2025 | Vendor telemetryEditorial classification · 2026-11 | Read through the generation-to-verification ratio: time in review is waiting time, not reviewer hours, so its rise does not lower the ratio. It is the queue that forms when generated volume outgrows the hours spent verifying it.Reality Graph's reading, not the source's · 2026-11 |
| Median time a pull request spends in review (2026 dataset, same measure as the +91% row) | +441%Faros AI telemetry, 2026 dataset, checked 2 November 2026 · 2026 | Vendor telemetryEditorial classification · 2026-11 | Read through the generation-to-verification ratio: the same measure rose much further from one dataset to the next, so the queue deepens instead of settling as teams adapt. A ratio that climbs while reviewer hours stay flat predicts this queue before cycle time shows it.Reality Graph's reading, not the source's · 2026-11 |
| Larger pull requests (2026 dataset) | +51%Faros AI telemetry, 2026 dataset, checked 2 November 2026 · 2026 | Vendor telemetryEditorial classification · 2026-11 | Read through review depth: depth counts comments per 100 changed lines, so larger pull requests lower it even when reviewers write as many comments as before. Compare depth at similar pull-request sizes before concluding that review got shallower.Reality Graph's reading, not the source's · 2026-11 |
| Experienced devs slower with early-2025 AI (while feeling faster), 10 July 2025; METR has marked this out of date | −19%METR randomized trial, checked 2 November 2026 · 2025 | Academic RCTEditorial classification · 2026-11 | No metric applies: the trial measured time per task, and none of the four metrics measures speed. The row describes the gap between felt and measured speed, which a repository does not record.Reality Graph's reading, not the source's · 2026-11 |
| Time per task with late-2025 AI, returning / new devs, 24 February 2026; METR calls this data unreliable | −18% / −4%METR follow-up, checked 2 November 2026 · 2026 | Independent study, rated unreliable by METREditorial classification · 2026-11 | No metric applies: time per task is a speed measure again. The row describes the open question whether AI saves a given team time at all, which needs a controlled comparison rather than repository data.Reality Graph's reading, not the source's · 2026-11 |
The four Faros rows describe one mechanism in two datasets: merged volume rose, pull requests grew, and the time a pull request waits in review rose further from the 2025 to the 2026 dataset. Read through the metrics, that is a generation-to-verification ratio climbing while review depth per 100 changed lines falls, the pattern traced in the review bottleneck.
The full METR correction is there too. METR’s study of 10 July 2025 found experienced developers 19% slower with early-2025 AI, and METR has since marked that result out of date on its study page. Its follow-up of 24 February 2026 (57 developers, more than 800 tasks, late-2025 tools) estimates about 18% less time per task for returning developers (interval −38% to +9%) and about 4% less for new participants (interval −15% to +9%), data METR itself calls unreliable, because developers who did not want to work without AI declined to take part or held tasks back.
Code quality and churn
| Figure | Evidence base | Reality Graph reading | |
|---|---|---|---|
| Two-week churn, 2020 to 2024, as AI assistance grew | ~3.1% → 5.7%GitClear, 211M lines, checked 2 November 2026 · 2025 | Vendor-adjacent analysisEditorial classification · 2026-11 | Read through two-week churn: this is the same metric with the same 14-day window, so it is the one baseline a team can compare directly. A value above about 6% sits past the 2024 level; the team's own three-month baseline is still the better comparison.Reality Graph's reading, not the source's · 2026-11 |
| Rise in code blocks with five or more duplicated lines during 2024 | ~8xGitClear, 211M lines, checked 2 November 2026 · 2025 | Vendor-adjacent analysisEditorial classification · 2026-11 | No metric applies: a copied block shows up as churn only once someone edits one of the copies, and none of the four metrics counts copies. The row describes a maintainability gap that a clone detector measures.Reality Graph's reading, not the source's · 2026-11 |
| Share of moved/refactored code | decliningGitClear, 211M lines, checked 2 November 2026 · 2025 | Vendor-adjacent analysisEditorial classification · 2026-11 | No metric applies: less moved code means less consolidation, which none of the four metrics records. Read beside the duplication row, it describes a codebase that grows by addition rather than by reuse.Reality Graph's reading, not the source's · 2026-11 |
Two-week churn is the only figure on this page that matches one of the four metrics one to one, so it is the one published baseline worth comparing directly. Duplication and moved code describe maintainability, which the metrics do not cover. What the churn numbers do and do not say is in AI code churn, and their euro translation in the cost calculation.
Security
| Figure | Evidence base | Reality Graph reading | |
|---|---|---|---|
| AI-generated samples that fail security tests (2025 edition) | ~45%Veracode, 2025 edition, checked 2 November 2026 · 2025 | Vendor studyEditorial classification · 2026-11 | Read through the unverified-merge rate: when close to half of generated samples fail a security test, a team that counts only a passing test suite as validation evidence leaves this class unmeasured. The security check belongs in the evidence the rate counts.Reality Graph's reading, not the source's · 2026-11 |
| Relevant samples failing to prevent XSS (CWE-80), 2025 edition | 86%Veracode, 2025 edition, checked 2 November 2026 · 2025 | Vendor studyEditorial classification · 2026-11 | Read through review depth: a weakness class that fails this often deserves a named review comment, not a general one. On changes that render user input, a substantive XSS check is the comment that depth should count.Reality Graph's reading, not the source's · 2026-11 |
| Failure rate: Java / C# / JavaScript / Python (2025 edition) | 72% / 45% / 43% / 38%Veracode, 2025 edition, checked 2 November 2026 · 2025 | Vendor studyEditorial classification · 2026-11 | Read through the unverified-merge rate: the failure share differed by language in this test, so the rate is worth splitting by language. A Java-heavy repository should not borrow the default of a Python team.Reality Graph's reading, not the source's · 2026-11 |
| Security across model generations while functionality rose (2025 edition; the 2026 edition of 28 July 2026 reports it still flat) | flatVeracode, 2025 edition, checked 2 November 2026 · 2025 | Vendor studyEditorial classification · 2026-11 | Read through the unverified-merge rate: if newer models do not close the security gap, a model upgrade does not lower the rate either. Only recorded validation does.Reality Graph's reading, not the source's · 2026-11 |
The four Veracode rows point at one decision: what a team counts as validation evidence. If a passing test suite is enough, the unverified-merge rate stays blind to the security class that newer models did not improve; a recorded security check makes it visible. The classes and defenses are in security vulnerabilities in AI code.
Supply chain and secrets
| Figure | Evidence base | Reality Graph reading | |
|---|---|---|---|
| LLM-recommended packages that do not exist | 19.7%USENIX Security (Spracklen et al.), checked 2 November 2026 · 2025 | Academic studyEditorial classification · 2026-11 | Read through the unverified-merge rate: an invented dependency passes review when nobody resolves it against the registry. A team that records a registry or lockfile check as validation evidence covers this class; one that does not counts such merges as verified.Reality Graph's reading, not the source's · 2026-11 |
| Unique hallucinated package names observed | 205,000+USENIX Security (Spracklen et al.), checked 2 November 2026 · 2025 | Academic studyEditorial classification · 2026-11 | No metric applies: the count sizes the attack surface for slopsquatting across all models tested, not a team's process. It describes a supply-chain gap outside the four metrics.Reality Graph's reading, not the source's · 2026-11 |
| Hallucinated names repeating in all 10 reruns | 43%USENIX Security (Spracklen et al.), checked 2 November 2026 · 2025 | Academic studyEditorial classification · 2026-11 | No metric applies: repeatability is what makes an invented name worth registering for an attacker. The row describes a supply-chain gap that a registry check closes, not a property of a team's review.Reality Graph's reading, not the source's · 2026-11 |
| New hardcoded secrets on public GitHub in 2025 | 28.65MGitGuardian, checked 2 November 2026 · 2026 | Vendor studyEditorial classification · 2026-11 | No metric applies: the count covers public GitHub as a whole, not one repository. It describes a secrets-hygiene gap that a secret scanner measures.Reality Graph's reading, not the source's · 2026-11 |
| Rise in AI-service secret leaks year over year | +81%GitGuardian, checked 2 November 2026 · 2026 | Vendor studyEditorial classification · 2026-11 | No metric applies: the rise tracks how many AI services a stack connects to. It describes a gap in credential handling rather than in verification.Reality Graph's reading, not the source's · 2026-11 |
| Secret-leak rate of Claude Code-assisted commits vs. the baseline of all public GitHub commits | ~2xGitGuardian, checked 2 November 2026 · 2026 | Vendor studyEditorial classification · 2026-11 | Read through the unverified-merge rate: if assisted commits leak secrets at about twice the baseline, a secret scan belongs in the validation evidence for AI-assisted merges. Without it, a merge counts as verified although no check looked for credentials.Reality Graph's reading, not the source's · 2026-11 |
Most of these rows describe gaps outside the four metrics, in the dependency and secrets layers. The bridge is the evidence definition: a registry check and a secret scan, recorded per change, bring both gaps into the unverified-merge rate. The mechanics are in slopsquatting and what AI tools actually read.
What these numbers do not say
Three limits of the readings. There is no robust industry figure for “what percentage of code is AI-generated”, so no reading rests on one. Several key sources are vendor or vendor-adjacent (Sonar, Faros, GitClear, Veracode, GitGuardian), so the readings use them as defaults and directions, never as a team’s value. And a reading says what a figure implies for a ratio, depth, rate or churn in general; your codebase’s values come from measuring it, via the four metrics, not from ours.
Figure register
Every external figure that three or more pages of this site use lives here once, with its primary source, the year of the edition and the date the source was last opened and compared. Those pages link to the row here instead of to the source, so a corrected value changes on every page at once.
| Figure | Value | Source | Checked |
|---|---|---|---|
| Developers using AI tools in development | 84% | Stack Overflow Developer Survey, 2025 | 2026-11-02 |
| Developers who do not fully trust that AI-generated code is functionally correct | 96% | Sonar, State of Code Developer Survey (press release), 2026 | 2026-11-02 |
| Developers who always check AI-assisted code before committing it | 48% | Sonar, State of Code Developer Survey (press release), 2026 | 2026-11-02 |
| Developers for whom reviewing AI-generated code takes more effort than reviewing a colleague's | 38% | Sonar, State of Code Developer Survey (press release), 2026 | 2026-11-02 |
| Developers who attribute a negative effect on technical debt to AI creating code that looked correct but was not reliable | 53% | Sonar, State of Code Developer Survey report, 2026 | 2026-11-04 |
| 42% | Sonar, State of Code Developer Survey (press release), 2026 | 2026-11-04 | |
| 65% | Sonar, State of Code Developer Survey (press release), 2026 | 2026-11-04 | |
| Pull requests merged per developer in high-AI-adoption teams (2025 dataset) | +98% | Faros AI telemetry, 2025 | 2026-11-02 |
| Median time a pull request spends in review (2025 dataset) | +91% | Faros AI telemetry, 2025 | 2026-11-02 |
| Median time a pull request spends in review (2026 dataset) | +441% | Faros AI telemetry, 2026 | 2026-11-02 |
| Larger pull requests (2026 dataset) | +51% | Faros AI telemetry, 2026 | 2026-11-02 |
| Time per task for experienced developers with early-2025 AI tools; marked out of date by METR | −19% | METR randomized trial, 2025 | 2026-11-02 |
| Time per task with late-2025 AI, returning developers; rated unreliable by METR | −18% | METR follow-up, 2026 | 2026-11-02 |
| Time per task with late-2025 AI, new developers; rated unreliable by METR | −4% | METR follow-up, 2026 | 2026-11-02 |
| Lower end of the interval for returning developers | −38% | METR follow-up, 2026 | 2026-11-02 |
| Upper end of the interval for returning developers | +9% | METR follow-up, 2026 | 2026-11-02 |
| Lower end of the interval for new developers | −15% | METR follow-up, 2026 | 2026-11-02 |
| Upper end of the interval for new developers | +9% | METR follow-up, 2026 | 2026-11-02 |
| Two-week churn in 2020 | 3.1% | GitClear, AI Copilot Code Quality, 2025 | 2026-11-02 |
| Two-week churn in 2024 | 5.7% | GitClear, AI Copilot Code Quality, 2025 | 2026-11-02 |
| Rise in duplicated code blocks | 8x | GitClear, AI Copilot Code Quality, 2025 | 2026-11-02 |
| Generated code samples that introduced an OWASP Top 10 flaw | 45% | Veracode, GenAI Code Security Report, 2025 | 2026-11-02 |
| Relevant samples that failed to prevent cross-site scripting (CWE-80) | 86% | Veracode, GenAI Code Security Report, 2025 | 2026-11-02 |
| Security failure rate, Java | 72% | Veracode, GenAI Code Security Report, 2025 | 2026-11-02 |
| Security failure rate, C# | 45% | Veracode, GenAI Code Security Report, 2025 | 2026-11-02 |
| Security failure rate, JavaScript | 43% | Veracode, GenAI Code Security Report, 2025 | 2026-11-02 |
| Security failure rate, Python | 38% | Veracode, GenAI Code Security Report, 2025 | 2026-11-02 |
| LLM-recommended packages that do not exist | 19.7% | USENIX Security (Spracklen et al.), 2025 | 2026-11-02 |
| Hallucinated package names repeated on every re-run | 43% | USENIX Security (Spracklen et al.), 2025 | 2026-11-02 |
| Hallucinated-package rate of the open-source models studied | 21.7% | Spracklen et al., USENIX Security, 2025 | 2026-07-17 |
| Hallucinated-package rate of the commercial models studied, at least | 5.2% | Spracklen et al., USENIX Security, 2025 | 2026-07-17 |
| New hardcoded secrets on public GitHub in 2025 | 28.65M | GitGuardian, State of Secrets Sprawl, 2026 | 2026-11-02 |
Sources of the register
- Stack Overflow Developer Survey (2025) checked 2026-11-02
- Sonar, State of Code Developer Survey (press release) (2026) checked 2026-11-02
- Sonar, State of Code Developer Survey report (2026) checked 2026-11-04
- Sonar, State of Code report summary (2026) checked 2026-11-04
- Faros AI telemetry (2025) checked 2026-11-02
- METR randomized trial (2025) checked 2026-11-02
- METR follow-up (2026) checked 2026-11-02
- GitClear, AI Copilot Code Quality (2025) checked 2026-11-02
- Veracode, GenAI Code Security Report (2025) checked 2026-11-02
- USENIX Security (Spracklen et al.) (2025) checked 2026-11-02
- Spracklen et al., USENIX Security (2025) checked 2026-07-17
- GitGuardian, State of Secrets Sprawl (2026) checked 2026-11-02
Where Reality Graph fits
Reality Graph cites these numbers; it does not generate its own. There are deliberately no Reality Graph statistics on this page. The product is in private beta, and inventing product-performance figures would violate the same claim-safety rule the rest of the site follows. What Reality Graph does is let a team produce its own version of these numbers, from its own runs, via the evidence reports and metrics - measured, not borrowed.
This reference gives you
- Every key AI-code figure with its source, year and check date
- A reading per figure, mapped to one of four metrics
- The independent-vs-vendor split, stated per source
- A living page updated as the research moves
It does not give you
- A robust 'X% of code is AI-generated' number - none exists
- Any Reality Graph performance statistics
- A universal multiplier for your codebase - measure it
- Certainty from single vendor studies - direction over gospel
FAQ
- How much code is AI-generated, and how many developers use AI tools?
- As of the 2025 Stack Overflow survey, 84% of developers use AI tools in development, a rise the survey puts at 14% since it began tracking in 2023. A reliable industry figure for the share of code that is AI-generated does not exist, and none of the four metrics needs one: each is computed from a team's own merges, reviews and follow-up commits. Read through the generation-to-verification ratio, adoption matters because it raises generated volume for nearly every team at once; whether review capacity followed is what a team's own ratio shows.
- How long do AI pull requests wait, and how much has review slowed?
- Faros AI telemetry reports the median time a pull request spends in review up 91% in its 2025 dataset and up 441% in its 2026 dataset, while pull requests merged per developer rose about 98% in the 2025 dataset. There is no universal wait in hours. Read through the generation-to-verification ratio, the rise is a queue: generated volume grows faster than the hours spent verifying it, and the ratio shows that before cycle time does.
- What is the single most-cited AI code statistic?
- Sonar's 2026 survey figure that 96% of developers do not fully trust AI-generated code while only 48% always check it before committing: the 'verification gap'. Read through the unverified-merge rate, the 48% gives a team without its own records a default of roughly half its AI-assisted merges unverified. That default is meant to be replaced by the share the team measures from its own validation evidence.
- Are these statistics from independent research or vendors?
- Both, and this page marks which. Independent or academic: the METR randomized trial, the USENIX Security 2025 slopsquatting study. Vendor or vendor-adjacent (reliable but with an interest): Sonar, Faros AI, GitClear, Veracode, GitGuardian. In the readings, a vendor figure sets a default or a direction, never a team's value: Sonar's 48% becomes a starting unverified-merge rate to replace with a measured one, and GitClear's churn series a baseline to hold a team's own two-week churn against.
- How current are these numbers?
- Every figure on this page was checked against its primary source on 2 November 2026, and the visible date moved with that check. Where a source has a newer edition, the row names both and the reading follows the newer one. Faros's 2026 dataset shows the review queue deepening rather than settling, which strengthens the generation-to-verification ratio reading; Veracode's 2026 report keeps security flat, which keeps the case for a security check in the unverified-merge rate. METR rates its own 2026 follow-up unreliable, so no metric reading rests on either METR edition.
- Can I cite these figures?
- Yes: for the figure, cite the underlying source and year named in its row, not this page. For the reading next to it, cite this page, because the reading is Reality Graph's and not the source's. Where a figure is a range or a trend, cite it as one; the biggest error in AI-code statistics is quoting a spread as a point.
Keep reading
Sources
- Sonar - State of Code Developer Survey: 96% do not fully trust, 48% always check, 38% more review effort (8 January 2026; checked 2 November 2026)
- Faros AI - telemetry, two datasets: +98% merged PRs and +91% time in review (2025 dataset), +441% time in review and +51% PR size (2026 dataset) (checked 2 November 2026)
- GitClear - AI Copilot Code Quality: churn ~3.1%→5.7%, duplication up ~8x across 211M lines (2025; checked 2 November 2026)
- Veracode - GenAI Code Security Report, 2025 edition: 45% fail, XSS 86%, Java 72%, flat across model generations (2025; checked 2 November 2026)
- Veracode - 2026 GenAI Code Security Report, the newer edition: security pass rate still flat (28 July 2026; checked 2 November 2026)
- Spracklen et al. - package hallucination study: 19.7%, 205k names, 43% persistent (USENIX Security 2025; checked 2 November 2026)
- GitGuardian - State of Secrets Sprawl 2026: 28.65M new secrets, AI-service leaks +81%, ~2x leak rate of Claude Code-assisted commits vs. all public commits (17 March 2026; checked 2 November 2026)
- METR - RCT: experienced developers 19% slower with early-2025 AI while feeling faster (10 July 2025; marked out of date by METR on its study page in February 2026; checked 2 November 2026)
- METR - We are changing our developer productivity experiment design (24 February 2026; late-2025 follow-up, 57 developers, 800+ tasks; rated by METR as an unreliable signal; checked 2 November 2026)
- Stack Overflow Developer Survey: 84% AI-tool use, a 14% rise since tracking began in 2023 (2025; checked 2 November 2026)