AI Code Quality Statistics 2026: What the Evidence Shows
AI tools raise output, some of the quality issues they introduce stick around, and automated review still trades coverage against noise. These are the studies behind those claims.
Five recent studies say AI-authored code can introduce detectable issues, some of which persist, and that neither human reviewers nor AI reviewers catch all of it. None of them supports a single headline percentage for "AI code".
For each study: what it measured, what it found, and how far the result reaches.
More than 15% of AI-authored commits introduced a detected issue
Debt Behind the AI Boom ran static analysis over 302,600 verified AI-authored commits in 6,299 GitHub repositories and attributed 484,366 issues to them.
More than 15% of commits from each of the five assistants introduced at least one issue. Most findings (89.3%) were code smells, and 22.7% of tracked issues survived to the latest revision.
How far it reaches: the unit is a commit, and only issues the chosen analyzers detect are counted.
Cleaning up agent search residue cut CodeSlop by 18% to 33%
The TRIM paper defines CodeSlop as functionally unnecessary edits an agent leaves behind from paths it tried and abandoned.
Its trajectory-minimisation method reduced measured CodeSlop by 17.9% to 32.9% across the agent scaffolds tested, with negligible loss in task performance.
How far it reaches: the figures apply to the paper's definition, tasks and models, not to every AI patch.
Open-source maintainers are merging fewer first-time pull requests
"AI Slop is DDoSing Open Source" analysed more than two million pull requests and issues across 294 repositories, alongside practitioner research.
The authors estimate one-time contributors saw an 18.18% drop in pull-request merge rate against the counterfactual. They then tested 11 response strategies with 229 open-source practitioners.
How far it reaches: this is observational analysis of public projects, so it does not show AI caused each rejection or that private teams see the same drop.
AI reviewers working from the diff miss most issues
SWE-PRBench tested eight automated review systems on 350 real pull requests with human-labelled issues.
Given only the diff, the systems found roughly 15% to 31% of issues. Every model got worse as more repository context was added to the prompt.
How far it reaches: these are benchmark setups; a product in your repository may do better or worse.
More review comments do not mean better review
CR-Bench scores automated reviewers on issue coverage and false positives together. A tool that comments everywhere finds more real issues but adds noise that costs trust and verification time.
GitHub reported that 67.3% of Code Quality preview findings in its own repositories were resolved before merge.
How far it reaches: the GitHub figure is a vendor-reported internal result, not an independent benchmark across companies.
Limits
Most of these papers are 2026 preprints. Each measures a specific unit (commits, patches, pull requests, benchmark issues) with specific tools, so none gives a universal rate for AI code. They do not show that AI code is worse in general, that a fixed share of it is insecure, or that any one tool can replace review.
Measure your own repository
Treat these studies as a hypothesis and set your own baseline: findings per thousand changed lines, reviewer dismissal rate, escaped defects, median review time, and the share of findings fixed before merge. Run a short pilot and keep only the checks that save more work than they create.
See the complete guide to AI-slop detection for a workflow. npx --yes aislop@latest scan runs locally as one deterministic input, not proof that the code is correct.
Primary sources
Frequently asked questions
What percentage of AI-generated code has quality issues?
There is no universal figure. One large study found that more than 15% of verified AI-authored commits from each assistant introduced at least one issue its static analyzers could detect.
Does AI-generated code create more technical debt?
Studies show agents can leave unnecessary edits behind and that some detected issues persist in repositories. That supports stronger checks, not a claim that every AI-assisted change is worse than human code.
Can AI code review replace human review?
No current benchmark supports that. Automated reviewers find useful issues but still miss many and raise false positives, so teams still need tests, deterministic checks and human judgment.