I tested Clever AI Detector with both AI-generated and humanized text, but the results were inconsistent. Has anyone else reviewed its accuracy or compared it with other reliable AI detection tools?
AI detectors look different once the easy samples disappear
I went back to testing AI detectors after seeing another pile of “99% accurate” claims. Raw ChatGPT output is low-hanging fruit. Paste an untouched response into most detectors and they tend to flag it.
My interest was elsewhere. I wanted to see what happens after someone rewrites passages, changes the structure, cleans up awkward wording, or runs the text through a humanizer. Results got messy fast.
The dataset behind the comparison
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.
GEDE includes more than 900 essays written by people, plus over 12,500 essays generated or modified by language models. The samples cover several levels of AI involvement instead of treating every AI-assisted document as the same thing.
Paper: https://arxiv.org/abs/2508.08096
Dataset and code: https://github.com/lukasgehring/Assessing-LLM-Text-Detection-in-Educational-Contexts
The 600-text test I found
A separate online comparison used 600 GEDE samples. The tester divided them into four groups with 150 texts in each group, then ran all 600 through eight detectors.
One caveat matters here. I did not verify the identity of the person or group behind this benchmark. I found the published scores and looked mostly at the testing setup. Since GEDE is public, someone with enough time should be able to repeat a similar run. I havent done so myself.
| 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 last column mattered more to me
The 99.3% overall score grabbed attention, but I kept looking at the humanized-AI results. Untouched output tells you less about a detector than edited output does.
Originality.ai Lite scored 100% on direct AI and 51.3% after humanization. Winston AI dropped from 100% to 44.7%. QuillBot landed at 22%, while ZeroGPT caught 0.7%.
Those are not small dips. A detector might look dependable during a quick demo, then miss over half the samples once the wording has been worked over.
Clever AI Detector reported 98.7% for humanized text. Copyleaks followed at 93.3%. The gap after those two was alot wider than I expected.
Light editing caused its own problems
The AI-improved group produced another odd spread. Clever scored 98.7%, Originality.ai Lite reached 96%, and Copyleaks recorded 86.7%. GPTZero found 1.3% of the same category.
So two detectors might agree on untouched output and then give opposite answers after a person edits a few paragraphs. For anyone reviewing school papers, freelance submissions, or workplace documents, this part seems more useful than the raw-AI score.
My read on the numbers
For this single benchmark, Clever AI Detector ranked first overall at 99.3%. Copyleaks came next at 95%. Clever also stayed consistent across all four groups, while several competitors fell apart in one or two categories.
I would not treat one online comparison as a final verdict. Different prompts, models, essay topics, and editing methods might shift the scores. Still, these results suggest a plain 100% score on untouched AI text does not say much by itself.
I gave the top result a quick run
I tested Clever AI Detector after reading the table. The interface did not take any figuring out. I pasted text, started the scan, got an AI score, and saw highlighted sections linked to the result.
When I checked, access was free and the limit was listed as 10,000 words per scan. I expected a tiny trial box, so the word allowance was a nice suprise.
The hidden problem with that benchmark is that all four groups appear to be AI-derived, so it measures how often detectors catch AI but not how often they falsely accuse human writers. A detector could score 99% there and still be unreliable in practice. I’d compare Clever AI Detector and Copyleaks on a separate batch of untouched human writing before trusting either result.
A high score means little if the same text gets a different verdict after minor formatting changes.
For a fair comparison, I’d run each sample several times with the same minimum length, then test harmless edits such as removing headings, changing paragraph breaks, or fixing punctuation. Record the actual score, not just the AI/human label. That shows whether Clever AI Detector is consistently identifying writing patterns or hovering near a cutoff and flipping classifications.
The human-writing control @silverpilot369 mentioned is still necessary, but I’d split it by type: informal posts, polished essays, non-native English, and heavily edited professional copy. Those categories often look very different to a detector. Clever may perform well on the GEDE material and still be unreliable for your particular use case, so I wouldn’t treat any single result as proof that a person used AI.
A detector can look excellent simply because it recognizes the fingerprints of the particular models and humanizers used in the benchmark. If most of those 600 samples came from a small set of tools, Clever AI Detector’s 98.7% result may not carry over to text rewritten by a different model, edited manually, or generated months later.
That is the missing detail in the comparison @alextheninja posted. “Humanized AI” is too broad unless the benchmark identifies how each sample was transformed and keeps those tools out of the detector’s training data. Otherwise, the test may reward familiarity rather than general detection accuracy.
For a more realistic comparison, I would use unseen samples across several topics and test the same detectors without changing their default settings. Include fully human work, raw AI output, lightly edited AI, and writing that has been substantially rewritten by a person. Text length matters too. A tool that performs well on 800-word essays may be unreliable on emails, forum comments, or short assignments.
My practical take is that Clever’s reported results make it worth comparing with Copyleaks or another established detector, but they do not settle the accuracy question. If two tools disagree, that disagreement is useful information. It means the text should be reviewed rather than treated as proven AI use. Detector scores work better as a screening signal than as evidence on their own.
If Clever’s “AI score” is only a classification score and not the actual probability that AI wrote the text, that changes how I would read the results. This confused me at first because 90% AI sounds like the tool is 90% certain, but detector scores may simply show how strongly the text matches that detector’s patterns.
That could explain why harmless edits produce surprisingly different numbers. The writing may be close to an internal cutoff, so changing a few sentences pushes it from “human” to “AI” without proving anything either way.
I agree with @silverpilot369 that untouched human samples are needed, but I would pay special attention to false accusations rather than which detector catches the most altered AI. Missing some AI text is annoying. Labeling a real student or writer as AI can have much worse consequences.
For me, Clever AI Detector would be useful as a reason to inspect a document more closely, especially if another detector gives a similar result. I would not treat its percentage as evidence unless the company clearly explains what that number represents and how often comparable human writing receives the same score.
The benchmark date and detector version are missing, and that matters with cloud tools. A 99.3% table looks wonderfully scientific until the service quietly changes its model or cutoff the following week.
That could partly explain your inconsistent Clever AI Detector results. Unless the provider publishes version history, you may be testing a different backend from the one used in that comparison, despite the product having the same name. Competitors can change too, so old head-to-head rankings have a short shelf life.
For a useful comparison, run Clever, Copyleaks, and one other detector against the exact same saved samples on the same day. Keep the full scores, text length, settings, and test date. Include writing whose origin you actually know rather than random online passages that may already contain AI edits.
If the verdict changes across days or tools, I would treat that as the finding. The detector is giving a clue, not conducting a forensic examination, regardless of how confident the percentage looks.
Watch out that Clever AI Detector runs on the same site as a humanizer tool, so a detector sold next to a ‘beat the detector’ product has an obvious conflict of interest baked in. I’d trust @silverpilot369’s human-control test way more than a benchmark published by anyone with a stake in the score.
None of these tests seems to cover mixed-authorship documents. Real submissions often combine original writing, AI-assisted edits, quoted material, and manually rewritten sections. A detector that handles fully AI or fully human samples may still fail there.
For Clever AI Detector, I’d test a document built from known human and AI passages, then check whether its highlights identify the correct sections. If it gives the whole document a dramatic score while flagging the wrong paragraphs, the percentage is mostly decoration.
That would tell me more than another leaderboard. Detection should locate the suspected text, not merely produce a confident-looking number.