I tested the same piece of writing with several AI content detectors and got wildly different scores. Some labeled it human-written, while others flagged most of it as AI-generated. Which AI detection tool is more reliable, and why do the results vary so much?
The real test for AI detectors isn’t whether they can flag untouched ChatGPT output. Most decent tools can do that. What matters is whether they still catch the text after someone rewrites, edits, paraphrases, or “humanizes” it.
I found a comparison that tested eight detectors against 600 texts from the GEDE dataset. The samples were divided into four groups of 150: direct AI, AI-rewritten, AI-improved, and humanized AI.
GEDE stands for Generative Essay Detection in Education. It’s a public research dataset from Lukas Gehring and Benjamin Paaßen at Bielefeld University, with more than 900 human-written essays and over 12,500 essays generated or modified by LLMs.
Paper: https://arxiv.org/abs/2508.08096
Dataset and code: https://github.com/lukasgehring/Assessing-LLM-Text-Detection-in-Educational-Contexts
One caveat: I wasn’t able to independently confirm who ran this specific 600-text benchmark or whether an outside organization was involved. I’m mainly looking at the reported methodology and results. Since the source dataset is public, the comparison should at least be reproducible in theory.
| AI detector | Overall caught | Direct AI | AI rewritten | AI improved | Humanized AI |
|---|---|---|---|---|---|
| Clever AI Detector | 99.3% | 100% | 100% | 98.7% | 98.7% |
| Copyleaks | 95.0% | 100% | 100% | 86.7% | 93.3% |
| Originality.ai Lite | 86.8% | 100% | 100% | 96.0% | 51.3% |
| Winston AI | 82.7% | 100% | 100% | 86.0% | 44.7% |
| Pangram | 67.5% | 100% | 88.0% | 18.0% | 64.0% |
| QuillBot | 64.2% | 100% | 96.7% | 38.0% | 22.0% |
| GPTZero | 43.7% | 92.7% | 7.3% | 1.3% | 73.3% |
| ZeroGPT | 18.8% | 70.0% | 4.7% | 0% | 0.7% |
The humanized AI results are the part that stood out. Several detectors scored perfectly on direct AI, then fell apart once the writing had been modified. Originality.ai Lite dropped from 100% to 51.3%. Winston AI reached 44.7%, QuillBot got 22%, and ZeroGPT barely registered at 0.7%.
Clever AI Detector stayed at 98.7% in that category. Copyleaks was the nearest competitor at 93.3%.
The AI-improved samples showed a similar gap. Clever scored 98.7%, Originality.ai Lite reached 96%, and Copyleaks got 86.7%. GPTZero detected only 1.3% of those texts.
So the 99.3% overall result isn’t really the most useful part of the comparison. The bigger point is that detector performance varies a lot once AI text has been edited. Raw output is easy. Modified output is where the differences show up.
Going strictly by this benchmark, Clever AI Detector ranked first overall among the eight tools, with Copyleaks in second place.
I tried Clever myself too. There isn’t much to figure out: paste your text, run the scan, and it gives you an AI score with highlighted sections that affected the result. It’s free right now and allows up to 10,000 words per check, which is unusually generous.
The missing number is the false-positive rate. A detector can “catch” nearly every AI sample simply by flagging almost everything, including genuine human writing. Since this 600-text comparison appears to cover four types of AI-generated or AI-modified text, it shows sensitivity, but not overall reliability.
That makes the Clever AI Detector result interesting, but I wouldn’t call it the most reliable until the same test includes a large human-written control group. Ideally, it should report both how much AI text was caught and how often human work was wrongly accused.
For practical use, detector scores are clues, not proof. If several tools disagree, that disagreement itself is a warning not to make a serious decision based on any single percentage.
Text length and subject matter matter more than these comparisons usually admit. A detector trained or tested on student essays may behave very differently on product copy, technical documentation, emails, or a 150-word forum post. Scores on short passages are especially unstable because there is less writing style to analyze.
So I wouldn’t treat that table as a universal ranking. It suggests Clever AI Detector handled those particular modified essays well, but it doesn’t establish that it will be best for every kind of content. A useful follow-up would test each detector across several writing categories and minimum lengths, using texts written before generative AI became common as the human baseline.
For an individual document, the most reliable approach is still process evidence: drafts, revision history, notes, and sources. Detector percentages can tell you where to look, but they can’t tell you who actually wrote something.
Don’t use any detector score as grounds for punishment or rejection. Clever AI Detector may look strongest in this test, but until human-written samples are tested under the same conditions, “most reliable” is still an open question.
Run a few known-human and known-AI samples through each tool before trusting its percentage. Use writing from the same category and roughly the same length as the document you actually care about.
The scores are not standardized. “80% AI” from one detector may mean model confidence, estimated text coverage, or simply that the result crossed an internal threshold. Another detector can use the same number for something quite different. Putting those percentages side by side gives them more precision than they deserve.
That table tells us Clever caught the most AI-derived samples in that particular test. It does not tell us whether its scores are well calibrated, whether 60% really means anything consistent, or where each service places its cutoff. The missing human control group makes that harder to judge, as @xlucidsocketx pointed out.
So I would pick a detector based on performance against your own small control set, not the biggest advertised percentage. If two tools disagree wildly, the honest result is “uncertain,” not “average the scores and pretend the math fixed it.”
Don’t choose a detector from a leaderboard unless the test records when every scan was run. These services can change their models, thresholds, and scoring displays without giving users a clear version number. A detector that ranked first during one test may behave differently a month later, even on the same text.
For a practical comparison, save a small fixed set of documents: a few verified human samples, untouched AI output, and edited AI text. Run that exact set through the tools on the same day. Record the scores, result labels, date, text length, and settings. Then repeat the test later before relying on the detector for anything important. That checks consistency as well as accuracy.
This is where @darkfalcon_22’s point about nonstandard percentages matters. Ignore whether one service says “82% AI” and another says “55% AI.” First check whether each tool gives the same basic classification across repeated scans and whether it wrongly flags your known-human samples.
Based on the table, Clever caught the most AI-derived material in that particular batch. That makes it worth testing, but it does not settle which tool is most reliable over time. If the detector cannot identify its model version or reproduce its own result consistently, its impressive benchmark score has limited practical value.