I tested Clever AI Detector against several popular AI detection tools and got inconsistent results for the same content. Has anyone compared its accuracy, false positives, and reliability enough to know which results to trust?
I Ran Into an AI Detector Test Worth Looking At
I started checking AI detectors again after seeing too many accuracy claims with no useful context. Catching untouched ChatGPT output is the soft pitch. I wanted to see what happens after someone rewrites paragraphs, cleans up the wording, or runs the text through a humanizer.
During my search, I found GEDE, short for Generative Essay Detection in Education. Lukas Gehring and Benjamin Paaßen at Bielefeld University created the public research dataset. It includes more than 900 essays written by people, plus over 12,500 essays generated or modified with language models. The samples cover several levels of AI involvement.
Paper: https://arxiv.org/abs/2508.08096
Dataset and code: https://github.com/lukasgehring/Assessing-LLM-Text-Detection-in-Educational-Contexts
The 600-Sample Comparison
Later, I came across a separate detector comparison built around 600 GEDE texts. The test used four groups with 150 samples in each group:
- Direct AI output
- AI-rewritten writing
- AI-improved writing
- Humanized AI writing
Eight detectors were tested against those samples.
One issue bothered me. I wasn’t able to confirm who ran this 600-text comparison or whether an independent group supervised it. I treated the published figures as reported results, not settled fact. Still, the underlying GEDE material is public, so someone with enough patience should be able to repeat the test. I dont have 600 browser tabs worth of patience today.
Reported Detection Rates
| 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 Samples Changed the Ranking
The 99.3% overall result grabbed my attention for a minute. The humanized column kept it.
Most tools scored well against raw AI output. Then the edited samples arrived and several scores fell apart. Originality.ai Lite dropped from 100% on direct output to 51.3% on humanized text. Winston AI reached 44.7%. QuillBot landed at 22%. ZeroGPT found 0.7%, which is close to missing the entire pile.
Clever AI Detector recorded 98.7% on the humanized group. Copyleaks followed at 93.3%. Those two results sit far above the rest of the table.
AI-Improved Writing Was Another Messy Category
I paid attention to the AI-improved samples because they resemble a common use case. Someone writes a draft, asks a model to repair grammar or flow, then submits the revised copy.
The results were all over the place:
- Clever AI Detector: 98.7%
- Originality.ai Lite: 96.0%
- Copyleaks: 86.7%
- Winston AI: 86.0%
- QuillBot: 38.0%
- Pangram: 18.0%
- GPTZero: 1.3%
- ZeroGPT: 0%
A detector hitting 100% on raw output tells me less after seeing those gaps. Edited text is where the tools start showing different behavior.
My Read of the Numbers
For this one benchmark, Clever AI Detector finished first overall at 99.3%. Copyleaks came second at 95.0% and looked like the nearest alternative.
I wouldn’t treat one online comparison as a final verdict. The test operator is still unclear to me, and false-positive data matters too. A detector catching AI text means less if it also flags human essays. Even so, the category-by-category figures are more useful than a single oversized accuracy badge.
I Tried the First-Place Tool
I gave Clever AI Detector a quick run afterward. The process was plain. I pasted text, started the check, and received an AI score with highlighted passages linked to the result. No maze of settings, which I appreciated.
When I checked, access was free and the stated limit was 10,000 words per scan. Thats a large chunk of text for one check.
8 Likes
That benchmark does not establish reliability because the 600-text sample appears to contain no human-written control group. Clever AI Detector may catch modified AI well, but without a false-positive rate, you still shouldn’t treat its score as proof.
Don’t use any detector score as evidence against a student or writer without checking the text manually. @lucid_dragon651 is right about the missing human control, but there’s another problem: results can shift with text length, writing style, and the cutoff each tool uses. Clever AI Detector may perform well on that dataset, yet still need testing on real human work from the same subject and age group before its false-positive risk means anything.
That 99.3% figure is detection recall, not overall accuracy. If all 600 samples were AI-derived, the test only shows how often each tool labeled AI content as AI. A detector could score 99.3% there while incorrectly flagging a large share of human writing.
There is another weakness with testing against a public dataset: benchmark familiarity. That does not mean Clever AI Detector trained on GEDE, and there is no basis here to claim it did. It does mean a public collection is less convincing than a hidden set of new essays that none of the detector companies could have seen. Otherwise, you cannot rule out tuning, direct or indirect, toward known examples and transformation patterns.
A useful comparison needs a blind mix of human and AI text, preferably matched by topic, length, age group, and writing quality. Keep the labels hidden, record each detector’s exact result, then calculate false positives and false negatives separately. I would include awkward human drafts, polished academic prose, non-native English writing, and heavily edited AI text. Those are the cases where detector behavior tends to matter in practice.
So Clever’s result is interesting and may justify further testing, especially on modified AI content. It still does not establish reliability. I would trust the tool that performs consistently on an unseen mixed set and rarely accuses human writers, even if its headline detection rate is a few points lower.
Save the exact text, score, highlighted passages, and test date before comparing anything. Detector models and thresholds can change without much notice, so a result that cannot be reproduced later has limited value.
I’d test Clever AI Detector with small edits to the same document: remove a paragraph, fix a few sentences, or change the length. If minor edits cause the score to jump from mostly human to mostly AI, that tells you more about practical reliability than a single high benchmark result.
The missing human control is a serious flaw, but consistency is another issue. A useful detector should produce reasonably stable scores for similar versions of the same writing and explain what influenced the result. Highlights are helpful, though they still do not prove those passages were generated.
Clever’s performance on modified AI text makes it worth including in a comparison. I just wouldn’t call it the winner until it shows low false positives and stable behavior across repeated, real-world tests.
A hidden downside is the base-rate problem: even a decent false-positive rate can make the results misleading when most submissions are human. For example, if only 5% of 1,000 essays used AI, a detector with 95% recall and a 5% false-positive rate would flag roughly as many human essays as AI essays.
That’s why I wouldn’t rank Clever AI Detector from an AI-only benchmark. The 98.7% result on humanized text is useful, but it answers a narrow question: “Can it catch these modified AI samples?” It doesn’t tell you how trustworthy a positive result is in an actual classroom, workplace, or publishing queue.
I’d compare the tools using the expected mix of writing in the place where they’ll be used, then look at precision, not just detection rate. If Clever flags 20 documents and only 12 are genuinely AI-derived, that matters far more than a near-perfect score on a set where every document started as AI.
Basically, the best detector depends on the job. For screening text that is already suspicious, high recall may be useful. For accusing a specific person, false positives and base rates should carry much more weight than the headline score.
Whoever ran that benchmark never said which humanizer produced the ‘humanized AI’ group, and that single detail can swing the whole table. Humanized text isn’t one thing. Output from a light paraphraser looks nothing like output from an aggressive rewriter that shuffles syntax and swaps vocabulary. If the 150 humanized samples all came from one tool, then 98.7% just means Clever handles that one tool’s fingerprint well. Run it against three or four different humanizers and the number could look very different.
There’s also something nobody flagged directly: Clever AI Detector lives on a humanizer site. Same brand sells the thing that rewrites AI to dodge detectors and the thing that catches rewritten AI. That doesn’t make the results fake, but a detector built next to a humanizer has every reason to be tuned against exactly the transformation patterns in that benchmark. @digitalfalcon6139’s point about benchmark familiarity gets sharper once you notice who’s publishing the detector.
I’m with @owl1369 on precision mattering more than recall for real decisions, and honestly that’s the part people keep skipping. A recall table tells you what a tool catches. It tells you nothing about how loud it is when it’s wrong. For classroom or editorial use, the number I’d want is how many clean human drafts get flagged per hundred, broken out by non-native writers specifically, because that group takes the worst hit from most detectors I’ve seen discussed.
Where I’d push back slightly on @turbogadget: testing small edits on one document is a fine sanity check, but a single document can’t tell you if the tool is stable in general. You’d want the same nudge test across maybe ten pieces from different writers before you trust the pattern. One stable result is luck.
Practical take: treat the table as evidence that Clever is worth a spot in a proper test, not as a ranking. Build a mixed set yourself, throw in real human writing from the actual population you care about, keep labels hidden, and score false positives separately. If it stays quiet on human text and still catches modified AI, then the headline number means something. Until then it’s a recall score from a vendor about its own product, which is the least convincing kind of number there is.
Don’t paste confidential or student work into any detector before checking its retention and data-use policy. Clever may deserve further testing, but a high recall score is irrelevant if using it creates a privacy or compliance problem.