AI Code Quality Report, September 2026
About 4 in 10 project scans found at least one AI-slop pattern, unchanged from August on twice as many scans.
In September, about 4 in 10 project scans run with aislop found at least one AI-slop pattern. That is the same rate as August, measured on roughly twice as many scans: about 227,000, run by about 22,000 developers.
This is the second monthly report. August’s report is here.
The AI-slop rate held at 4 in 10
AI slop here means the patterns coding agents typically leave behind: comments that restate the code, errors that are caught and ignored, unsafe type casts and helpers left over from an earlier attempt. 40% of scans found at least one, against 40% in August.
More than half of scans found an issue of some kind, and about 1 in 5 found a security issue. A finding is usually a single line, so most scans still scored in the high 90s out of 100.
Most agent edits are fine
aislop can also run automatically each time a coding agent edits a file. In September those scans covered about 46,000 edits where we could compare the file’s score before and after. About 1 in 20 edits (4.7%) left the file scoring worse, usually by 2 or 3 points out of 100.
August’s figure was 12%, but the two are not comparable. aislop 0.15.0 changed how scores are calculated at the end of August, so the drop mostly reflects the new scoring, not better agents. We will compare month to month again once there is a full month on the same scoring.
Limits
Everyone in this data chose to install a code quality tool, so it is not a random sample of all code. One user ran about a fifth of all local scans; leaving them out raises the AI-slop rate to about 45%. Nearly all agent-edit scans came from Claude Code, so we do not compare agents. Telemetry is on by default and can be turned off, so the totals are a floor.
Correction to August
August’s report said an agent edit was about five times more likely to make a file worse than better. That comparison was wrong: aislop versions before 0.14.1 almost never recorded an improvement, so there was nothing reliable to compare against. We have withdrawn it and added a note to the August post. Its 12% figure for edits that made a file worse still stands.
To see where your own project stands, run npx aislop@latest scan, or read how the score is calculated.
Frequently asked questions
How common was AI slop in September 2026?
About 4 in 10 project scans (40%) found at least one AI-slop pattern, the same rate as August. More than half found an issue of some kind, and about 1 in 5 found a security issue.
Do AI coding agents make code worse?
Usually not. About 1 in 20 agent edits left the file scoring slightly worse, typically by 2 or 3 points out of 100. The rest left it the same or better.
Can September be compared with August?
The AI-slop rate can. The agent-edit figure cannot yet, because aislop changed how scores are calculated at the end of August.
Where does this data come from?
Anonymous telemetry from the aislop CLI. It records scores and finding counts, never source code, file names or repository names, and can be turned off with AISLOP_NO_TELEMETRY=1.