Accuracy and methodology
We used 100 code snippets in a blind test. Human-written and AI-generated solutions addressed the same problems, with the source labels hidden during scoring.
Blind-test results
| Model | Accuracy | Scope |
|---|---|---|
| gpt-6.2-sol | 96% | Full blind test |
| Gemini | 90% | Full blind test |
| deepseek-v4.1-flash | 88% | Tested subset |
| GPTSniffer (open source) | 54% | Full blind test |
Source: DetectCodeAI project inception blind-test results. Flash’s subset score and the full-set model scores use different sample scopes.
Where Flash makes mistakes
- False positives on human-written code
- 8%
- Missed AI-generated code
- 16%
- Detection of GPT-generated code
- 75%
The false-positive and miss rates describe different groups in the Flash test. GPT-generated code was a weaker part of that benchmark. Review the snippet and context before acting on a result.
What published studies found
Pan and colleagues found that detectors designed for natural-language text struggled with code. Suh and colleagues reported that the text detectors they evaluated mostly scored below 0.6 accuracy on source code.
How a check works
- We identify the language automatically or use the language you select, then apply your plan’s length limit.
- Flash estimates likely authorship from observable patterns such as naming, structure and comments.
- Sol automatically reviews estimates below 66 or with conflicting signals across all plans. When review runs, its estimate determines the verdict.
- A model confidence of 66 or higher produces a Likely AI-written or Likely human-written verdict. Lower scores stay in the Possibly AI-assisted band. Signals explain the estimate.
This result is a probability estimate, not proof. Never use it as the sole basis for academic or employment decisions. Detection is probabilistic — always review the signals yourself before acting.
Try the code detector