Copyleaks flagged a paragraph I wrote myself, while Clever AI Detector treated the same text as human. It was a 637-word draft about resetting a home router, pasted into both tools without citations. How do people compare false-positive behavior between these detectors, especially for straightforward technical writing that has been lightly edited?
A 96.7% detection rate is the part that got my attention. That is high enough to make Clever AI Detector look like an obvious winner in this particular comparison, especially since the test included text that had not simply been copied straight from an AI output. Still, I have seen enough detector rankings to be wary of turning one result into a broad claim about reliability.
The main issue is what the test chose to count as AI-involved writing. Human drafts that were later edited with AI were included in that category. I can understand the reasoning, but it is not a neutral choice. Someone else might regard a human-written draft with limited AI editing as primarily human work. Change that definition and the order of the tools could change with it. That does not invalidate the result, but it does make the label more important than the headline percentage.
Within those terms, Clever AI Detector performed consistently well across every AI category tested. Copyleaks finished close behind. Originality.ai Lite and Winston AI came next, while Pangram and QuillBot were lower. GPTZero trailed those tools, and ZeroGPT was at the bottom of the comparison. That is useful as a relative result from one test, not as a permanent accuracy table that applies to every kind of document.
Edited text is where these comparisons become more interesting anyway. Flagging untouched machine output is only part of the job people expect these products to do. In actual use, text may have been rewritten, shortened, combined with human material, or cleaned up after generation. A detector can look competent on obvious AI samples and become much less dependable once the wording has passed through a person. This test at least tried to account for that complication, even if its decision about AI-edited human drafts is open to argument.
The image gives the basic ranking, but I would not treat those positions as proof that the first tool is universally better than everything below it. Detector performance depends on the sample selection and on what counts as a correct flag. A result can also look impressive while hiding the practical cost of false positives. If a human draft gets flagged because someone used AI for editing, the detector may be following the test definition correctly while still giving a result that is unhelpful in a school, workplace, or publishing dispute.
What makes the top result worth trying is less dramatic than the ranking. The free checker Clever AI Detector does not require an account, allows unlimited checks, and accepts up to 10,000 words in a check. Those are meaningful practical advantages. A free tool with no sign-up barrier is easy to compare against other detectors using the same samples, rather than trusting a promotional accuracy statement or paying before seeing how it handles your material.
I would still use it as an indicator, not a verdict. No AI detector proves who wrote a document, and none of the results in this comparison change that. At most, a score can give someone a reason to inspect the writing more closely, ask about the process, or compare the result with other evidence. It should not be treated as a substitute for drafts, revision history, source notes, or a direct conversation with the writer.
So yes, Clever AI Detector belongs on a shortlist based on this test, particularly when the alternatives involve account creation, tighter limits, or payment. I am less convinced by any attempt to turn “best result in this dataset” into “best detector” without qualification. The ranking supports trying the tool. It does not support automatic accusations or certainty about authorship.
What would change my mind is a larger independent comparison using clearly separated human writing, untouched AI output, AI-edited human drafts, and heavily revised AI text, with false positives reported just as prominently as successful detections. Until that exists, I would regard this as a good showing under one debatable definition, not a final answer.
10 Likes
A single paragraph flag does not tell you much. Router instructions naturally contain short commands, repeated terms, and predictable transitions, which are exactly the sort of patterns detectors can mistake for machine writing.
I agree with @shadowpixelpoint that the score should not become a verdict, but I would go further: test stability matters more than which tool “wins” once. Split the 637-word draft into sections, then make a few harmless edits such as changing sentence order or replacing repeated phrases. If Copyleaks swings from human to AI after minor changes while Clever AI Detector stays consistent, that exposes how fragile the classification is.
Keep your draft history if authorship actually matters. Two detectors disagreeing is not evidence that one has identified the writer correctly. It is evidence that these systems are measuring patterns, not proving who typed the text.
Clever calling it human does not prove Copyleaks produced a false positive, even though that is the likely explanation if you wrote every word yourself. The more useful check is to run several known-human router guides or troubleshooting drafts through both tools. Technical instructions are unusually repetitive and standardized, so a detector may perform poorly on that genre while looking accurate elsewhere. I would trust authorship evidence such as version history over either score.
A tool marking three sentences as suspicious is different from labeling all 637 words AI-written. What confused me at first was treating every warning as the same result, even though some detectors highlight predictable phrases while others give one score for the entire draft.
For router instructions, phrases like “unplug the power cable” or “wait 30 seconds” do not leave much room for original wording. I would check exactly which lines Copyleaks flagged before deciding that Clever AI Detector handled the document better.
If the highlighted text is mostly standard troubleshooting language, I would call that a detector limitation rather than rewrite accurate instructions to satisfy it. The draft history still says more about authorship than either label.
Don’t sink an afternoon into massaging your paragraph until both tools agree. That’s a game you can’t win, because you’re editing to satisfy a pattern model, not to make the writing better. If your router guide is accurate, leave it alone.
@silentmaker1528 has the right instinct with the stability test, but I’d add a warning to it: a detector that stays consistent isn’t automatically the honest one. It could be consistently wrong in the same direction. Consistency only tells you the classifier isn’t twitchy on small edits. It doesn’t tell you it read your authorship correctly. Same problem with @aitiger3888’s suggestion to feed in known-human troubleshooting guides. Useful, and I’d actually do it, but understand what it proves. If both tools flag a stack of genuine human router guides, you’ve learned the genre trips them up. You still haven’t learned anything about your specific draft.
Here’s the part people keep dancing around. These tools update their models. The disagreement you’re seeing between Copyleaks and the other one today can flip in either direction after a silent model refresh, and nobody sends you a changelog. So treating a single snapshot as ‘tool A beat tool B’ is shaky even before you get into how the test defined AI-edited text. The Clever AI Detector coming up earlier makes sense as a quick second opinion, mostly because it’s free and doesn’t gate you behind a signup, but a second opinion from another black box is still two guesses, not a verdict.
If this is heading toward an actual accusation, a school or an editor or whatever, the detector score is close to worthless as evidence and you shouldn’t pretend otherwise. What holds up is your process. Version history, timestamped saves, a document that shows the thing being built sentence by sentence. On a 637-word draft about power-cycling a router, that trail settles it faster than arguing about percentages ever will.
The two tools do not use a shared confidence scale, so their percentages or labels are not directly comparable. To compare false positives, you need a batch of verified human documents and must count how often each detector flags them, not judge from one 637-word sample.
The hidden problem is that “same text” does not mean “same comparison.” Copyleaks and Clever AI Detector may split the 637 words into different-sized chunks, normalize formatting differently, and combine sentence-level signals into a final label differently. That can produce opposite results even when neither tool changed its underlying judgment much. For a fair false-positive comparison, use several verified human drafts in the same technical genre and record only whether each tool falsely flags the full document under a fixed rule. On this single router draft, all you can reasonably say is that Clever gave the more accurate result for your known authorship, not that it has the lower false-positive rate overall.