AI detectors: tools, scores and evidence
Understand the detectors named in our scorecards, what their numbers mean and what evidence you need before comparing the results.
AIHT currently publishes stored detector figures, without a complete set of source passages, outputs and run logs. We cannot substantiate a new catch-rate leaderboard or monthly trend from those summaries. Provider descriptions below were checked on October 7, 2026; that is a source-review date, not a test date.
The detector directory
Compare the type of output and the details to record. These are product descriptions, not an accuracy ranking.
| Detector / source | What to look at | AIHT evidence | Before comparing |
|---|---|---|---|
| GPTZero | Text checking with document-level results and highlighted passages. | Summary only | Check how the result relates to the passage and which scan mode was used. |
| Originality.ai | AI checking alongside plagiarism and publishing-oriented tools. | Summary only | Record the detector model and settings; avoid mixing results from different modes. |
| Copyleaks | AI text detection with integrations and API options. | Summary only | Keep the returned classification, confidence information and integration settings together. |
| Winston AI | AI checking with a human-score display and a prediction map. | Summary only | Read the direction of the score: a human score differs from an AI-likelihood score. |
| ZeroGPT | Browser-based text checking with highlighted passages. | Summary only | Distinguish the document percentage from the individual highlighted passages. |
Summary only = stored values exist, but their underlying runs cannot currently be reproduced from the published materials. It does not mean a detector scored zero.
Three numbers that answer different questions
A tool’s assessment of a particular text. Check whether a higher score means more human-like or more AI-like.
Flagged AI-origin samples ÷ all evaluated AI-origin samples. A rate needs a threshold, sample count and test conditions.
Human-written samples incorrectly flagged ÷ all evaluated human-written samples. A humanizer-only test cannot establish this.
Rates from different sample sets or thresholds are not directly interchangeable. A low detection rate also does not establish factual accuracy, originality or writing quality.
What is available today
| Evidence | Status | What that means |
|---|---|---|
| Published detector summaries | Available with limits | Read each page’s labels. Some older templates use different datasets. |
| Original text, rewrites and detector logs | Not published | No reproducible run-level evidence package is available here. |
| Monthly detector changes | Not established | No verified time series supports an up/down movement claim. |
| Human-written control results | Not established | We cannot calculate a supported false-positive rate. |
Official sources
The descriptions below refer to the providers’ own product pages. Their performance claims are not independent AIHT test results.
GPTZero
Text checking with document-level results and highlighted passages.
Check how the result relates to the passage and which scan mode was used.
Official product source ↗Originality.ai
AI checking alongside plagiarism and publishing-oriented tools.
Record the detector model and settings; avoid mixing results from different modes.
Official product source ↗Copyleaks
AI text detection with integrations and API options.
Keep the returned classification, confidence information and integration settings together.
Official product source ↗Winston AI
AI checking with a human-score display and a prediction map.
Read the direction of the score: a human score differs from an AI-likelihood score.
Official product source ↗ZeroGPT
Browser-based text checking with highlighted passages.
Distinguish the document percentage from the individual highlighted passages.
Official product source ↗What about Grammarly?
Grammarly has a separate detector guide on this site. It is not part of our five-detector summary dataset.
Read the Grammarly guide →Before acting on a result
Use the score as a reason to inspect the writing and its history. It is not proof of authorship or a substitute for checking facts, sources and the rules that apply to your work.