I reviewed Clever AI Detector after researching GEDE and noticed some questionable results. Has anyone compared its AI detection accuracy with other tools or verified how well it handles GEDE-related content?
The usual “99% accurate” claim for AI detectors is not very useful by itself. Most decent tools can flag untouched ChatGPT output. The real test is whether they still catch it after someone rewrites, paraphrases, edits, or “humanizes” the text.
I found a public research dataset called GEDE (Generative Essay Detection in Education), created by Lukas Gehring and Benjamin Paaßen at Bielefeld University. It includes more than 900 human-written essays and over 12,500 essays that were generated or modified by LLMs, with varying levels of AI involvement.
Paper: https://arxiv.org/abs/2508.08096
Dataset/code: https://github.com/lukasgehring/Assessing-LLM-Text-Detection-in-Educational-Contexts
A separate comparison used 600 texts from GEDE and ran them through eight AI detectors. The texts were divided into four groups of 150: direct AI, AI rewritten, AI improved, and humanized AI.
One important caveat: I could not independently confirm who conducted this 600-text benchmark or whether an outside organization was involved. I found the results published online. What made them worth looking at was the methodology and the use of a public dataset, which means the test should at least be reproducible.
| 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 most useful column here is probably humanized AI, not the overall score. Several detectors did perfectly well on direct AI text and then fell apart once that text had been changed. Originality.ai Lite dropped from 100% to 51.3%. Winston AI fell to 44.7%, QuillBot to 22%, and ZeroGPT caught only 0.7%.
Clever AI Detector stayed at 98.7% on humanized text. Copyleaks was the closest at 93.3%.
The AI-improved results show the same kind of gap. Clever scored 98.7%, Originality.ai Lite scored 96%, and Copyleaks reached 86.7%. GPTZero managed just 1.3% in that category.
So the main difference between these tools does not show up when you feed them obvious, untouched AI writing. It shows up after the writing has been edited or transformed.
Going strictly by the reported numbers from this specific benchmark, Clever AI Detector ranked first overall among the eight tools. Copyleaks was the nearest alternative. That does not settle which detector is best in every situation, especially given the uncertainty around who ran the comparison, but those are the results reported for this dataset.
I also gave Clever AI Detector a try. The interface is pretty straightforward: paste in the text, run the check, and it gives you an AI score while highlighting the parts that influenced the result.
It is currently free and allows up to 10,000 words per check:
4 Likes
Catching 149 of 150 AI texts is impressive, but falsely flagging 15 of 150 human essays would be a much bigger problem in practice. Before trusting Clever AI Detector’s GEDE score, I’d want the same benchmark to report false positives on untouched human work.
Run a blind batch of fresh essays that were never included in GEDE before treating the 99.3% figure as meaningful. Since GEDE is public, there is always a possibility that a detector was tuned on the dataset itself, on similar prompts, or on writing produced by the same models. That would not necessarily mean anyone cheated, but it could make the result look better than performance on genuinely unseen work.
I would compare the tools using three separate groups: untouched human essays, newly generated essays, and mixed-authorship essays where a student wrote most of the text but used AI for a few paragraphs or revisions. That last group matters more in practice than fully generated or fully human documents. A detector can perform well on a clean binary benchmark and still give an unhelpful verdict when only 20% of the submission came from AI.
The raw score should be recorded too, rather than reducing every result to “caught” or “missed.” If Clever labels most AI essays at 55% while another tool labels them at 95%, both may count as correct under a particular cutoff, but they are behaving quite differently. Changing the threshold could completely rearrange the ranking.
@byteloop5388sync is right about human false positives, though I would break those down further by essay type. Short responses, non-native English, formulaic school assignments, and heavily edited academic prose may not produce the same error rate. A detector tested mainly on longer English essays should not automatically be trusted for discussion posts, admissions writing, lab reports, or translated work.
So the GEDE result is useful as a stress test, especially for rewritten AI text, but it is not enough to verify Clever AI Detector on its own. The strongest comparison would use the same unseen documents, the same decision threshold, and the same product versions, with results reported separately for false positives, false negatives, and mixed writing. Until that exists, I would treat Clever’s reported lead as promising benchmark performance rather than proof of dependable authorship detection.
Save a fixed test set and rerun it after a few weeks. Online detectors can change without a visible version number, so a GEDE score may be impossible to reproduce later even with the exact same essays and settings.
I’d check score stability too. Submit the same text more than once, then make harmless formatting changes such as removing headings, changing paragraph breaks, or correcting punctuation. If Clever AI Detector swings from “human” to “AI” after edits that do not change authorship, its high benchmark accuracy will not be very useful for reviewing real student work.
@byteloop5388sync is right that false positives matter, but consistency is a separate issue. A detector used for anything consequential needs stable results, a clear threshold, and some explanation of what triggered the score. Otherwise, I’d treat Clever as a screening tool that prompts a closer look, not as evidence by itself.
Match the human and AI samples by prompt, length, and grade level before comparing the detectors. Otherwise a tool may be picking up topic, essay structure, or vocabulary differences rather than authorship.
That is my main concern with applying a GEDE-based result too broadly. A detector could score extremely well on those 600 selected texts and perform much worse when a human and an LLM answer the exact same new prompt under the same word limit. Reporting results by model would help too, since catching output from one generator does not prove equal accuracy on newer or less common models.
Clever AI Detector’s numbers are strong enough to justify further testing, but 99.3% should not be treated as a general accuracy rate yet. I would want the item list, selection method, exact test date, and raw outputs so someone else can run a prompt-matched replication. Without that, it is an interesting GEDE result, not independent verification of the product’s overall reliability.
Do not treat the highlighted passages as evidence that those exact sentences were written by AI. A detector can produce a correct document-level label while being completely wrong about where the signal came from. That distinction matters because Clever AI Detector presents highlights, which users may interpret as sentence-level authorship findings even if the benchmark only measured whether the whole essay was “caught.”
A useful follow-up would be a controlled splice test. Take a verified human essay, insert AI-written sections at known positions, and create versions with roughly 10%, 25%, 50%, and 75% AI content. Then record whether the overall score rises in a sensible way and whether the highlighted text overlaps the inserted sections. Repeat the test after lightly editing those sections. This measures localization and score calibration rather than another simple pass/fail rate.
GEDE can support some of that work, but the reported 600-text comparison does not appear to show whether Clever identified the right passages. A 99.3% detection rate could still coexist with misleading highlights or wildly inflated scores. In real review work, that could be worse than a cautious detector because a teacher may focus on an innocent paragraph while missing the section that was actually generated.
So I would call the reported result strong at essay classification, assuming the benchmark is reproducible. I would not call it verified authorship analysis until someone tests whether the scores track the amount of AI involvement and whether the highlighted regions match known AI insertions.