The “AI slop” debate runs on anecdotes. Revert rates offer something better: a measure of how often merged code gets undone. Spark data shows what that measure says about AI-assisted vs. human-only code.
The Short Answer
In the Larridin customer examples reviewed, AI-assisted code was reverted less than human-only code. The gap ranged from about 4x (4.2% vs. 18.1% in a 12-week trend) to about 75x (0.2% vs. 15.16% at 30 days). Revert rates can be a useful code-quality signal, but comparisons need consistent measurement periods and context from your own repositories.
Key Wins
What’s a Spark Insight?
A Spark Insight is a data-backed finding Larridin surfaces automatically from how your teams actually use AI, flagging the patterns worth a leadership conversation.
Who This Matters To
CTO · VP Eng · Eng managers
| If you’re asking… | The answer | Go deeper |
|---|---|---|
| Is AI-generated code lower quality? | Not by revert rate, in Larridin data. AI-assisted code was reverted less often than human-only code in every instance reviewed, though the size of the gap varied widely. | The data |
| What is a revert rate? | The share of merged code changes that are later undone. It’s a simple, objective signal that a change didn’t hold up. | Definition |
| How much less often is AI code reverted? | It depends on the org. Gaps ranged from about 4x to about 75x across instances, so no single multiplier applies everywhere. | The data |
| Does AI make software delivery less stable? | DORA’s 2025 research linked AI adoption to both higher throughput and higher instability. Revert rates are one way to check which way your own org is moving. | Why it matters |
Engineering leaders are under pressure from two directions: expand AI coding tools for speed, and protect code quality. Without data, the quality debate turns into opinions about “AI slop.” Revert rates turn it into something you can measure.
| What you’re struggling with | What you actually need | How Larridin helps | Proof from Spark data |
|---|---|---|---|
| “Our engineers say AI code is sloppy.” | An objective quality signal | Compares revert rates for AI-assisted and human-only code | AI-assisted code reverted less often in every instance reviewed |
| “I need evidence before expanding AI tools.” | A comparison from your own codebase | Tracks 30-day and 90-day revert rates by code origin | One org: 30-day revert rate 0.39% (AI) vs. 6.18% (human) |
| “Human code quality seems to be slipping.” | Trend data, not snapshots | Shows revert-rate trends over 12 weeks | One org: human revert rate +102% over 12 weeks, AI +6% |
| “Vendor benchmarks don’t match our experience.” | Your own baseline | Measures the gap in your repos continuously | Gaps range from ~4x to ~75x across instances |
The question you’re asked: “Is AI code safe to ship at scale?”
What to do: Report revert rates by code origin next to adoption. Use your own numbers, not industry multipliers, when making the case to expand.
The question you’re asked: “Should reviewers treat AI PRs differently?”
What to do: Use revert data to set review norms. If AI-assisted changes hold up as well or better, focus review effort on risk, not origin.
For more on connecting AI adoption to engineering outcomes, see AI Monitoring for Engineering Leaders.
The question you’re asked: “Why are our reverts going up?”
What to do: Split the trend by AI-assisted vs. human-only work before drawing conclusions. In one org, the human-only revert rate rose much faster.
Larridin’s own data: In Larridin’s internal engineering data, where 63.7% of code was AI-assisted, the 30-day revert rate was 0.45% for AI-assisted code vs. 18.01% for human-only code (90-day: 0.47% vs. 19.74%).
These are selected observational examples, not an industry benchmark or a causal test of AI code quality. Revert definitions, cohort sizes, task mix, and repository coverage should be checked before comparing organizations.
In Larridin data, AI-assisted code has been reverted less often than human-only code in every instance reviewed.
A revert means a merged change was undone. Comparing revert rates by code origin provides a measurable signal of code durability, though it doesn't capture every type of quality problem.
Across four instances, including Larridin’s own engineering team, AI-assisted code had the lower revert rate every time. That’s a consistent direction, but not a consistent size.
Spark callout: In one org, the human-only revert rate rose 102% over 12 weeks, while the AI-assisted rate rose 6%. That gives engineering leaders a different question to investigate: What's driving the sharper increase in human-only reverts?
Myth: AI-generated code is “slop” that gets thrown away.
Reality: In every instance reviewed, AI-assisted code was reverted less often than human-only code.
The percentages alone can't explain why. Differences in teams, codebases, tasks, and measurement windows can all affect comparisons.
The kinds of tasks engineers hand to AI, how AI-assisted changes are scoped and reviewed, and how much of the codebase AI touches all likely shape the result. Revert rates also don’t capture every quality issue; they show what was undone, not everything that should have been.
That’s why the right takeaway isn’t a multiplier. It’s that the comparison is worth running in your own repos, where the factors behind it are yours.
They measure different things, and both can be true at once.
DORA's 2025 State of AI-assisted Software Development report found that greater AI adoption was associated with higher software delivery throughput and greater instability. That doesn't contradict Larridin's revert-rate examples: the studies measure different outcomes, and instability can involve system-level factors beyond whether individual changes were reverted. Revert rates are one signal; change failure rate and lead time are others.
Possible reasons include the kinds of tasks given to AI, smaller or more focused changes, and extra scrutiny on AI-assisted PRs. The data shows the pattern; your own repos will show which factors apply.
No. It captures changes that were undone, not bugs that were fixed forward or issues that were never caught. Use it alongside change failure rate and code turnover.
It varies widely by codebase and team. In the instances reviewed here, human-only 30-day revert rates ranged from about 6% to 18%.
See revert rates for AI-assisted and human-only code in your repos, and bring them to your next engineering leadership meeting.