AI content detection is one of the fastest-growing tool categories of the last two years. ZeroGPT and GPTZero are the two most visited detectors, both built around the same task: estimate whether prose was written by a model. This post compares that approach with code-aware detection on technical text - no ranking, just differences.
What General Detectors Optimize For
ZeroGPT and GPTZero score prose signals: perplexity, burstiness, sentence-length uniformity, phrasing statistics. On essays and articles this is exactly the right feature set.
The Technical-Text Problem
Developer text is adversarial input for a prose-statistics model. Code has low burstiness by nature - uniform lines, consistent formatting, repeated structure. The documented failure mode: a genuine, human-written technical document scores as AI-generated because its code blocks look statistically machine-like.
| Input type | Prose-statistics detector | Code-aware detector |
|---|---|---|
| Essay / article | Full-featured scoring | Same prose signals applied |
| Code block in a doc | Often inflates the AI score | Forced to p=0 - code is not prose |
| Commit messages, commands | Frequently flagged | Recognized as code/commands |
| Technical docs (mixed) | Score dominated by code share | Prose judged, code excluded |
Two Different Jobs
General detectors answer: was this essay written by AI? Code-aware detectors like Lint AI Detector answer: which parts of this technical document read as AI-generated prose, with code excluded from judgment? On a README or tutorial, the second question is usually the one you actually have.
Also Worth Knowing
Every detector - including this one - is a heuristic. False positives and negatives happen on all of them. Treat scores as one signal among several, especially on short texts.
👉 Try Lint's code-aware AI detector - sentence-level report, code never flagged, free tier.
Top comments (0)