And I kept the passages unchanged when moving between tools. Otherwise, I would have been comparing different inputs as well as different detectors.
Fixing the inputs
I ran the checks on October 2–3, 2026, using six English passages: an AI essay, business memo, paraphrase, fictional personal story, a Lewis Carroll excerpt, and that excerpt lightly edited by AI. Their lengths ranged from 271–328 words.
Some passages were related, and the Carroll selection was a narrow human control. I counted only completed checks. I also kept confidence, probability, and estimated AI share separate when reading the reports. Those labels did not describe the same measurement.
Working through the submissions
The generated passages produced quite different outcomes. I completed the entire set without paying or registering on the Clever AI Detector.
It identified all four generated texts. Pangram caught the generated samples I completed there, but its allowance ended after five checks. Originality.ai correctly flagged its completed AI samples with full confidence. I had selected its 15% AI Allowance, so that confidence concerned a threshold-based verdict, not the proportion of generated words. AI or Not also correctly labeled its completed generated samples.
Sapling caught the essay and story, although it processed only 2,000 of the essay’s 2,070 characters. The story check covered the complete passage. GPTZero caught my first three AI inputs at 100% AI. When I replaced the text afterward, the previous score remained visible; I did not treat that as another result.
Grammarly and Quetext each caught the essay before registration interrupted further checking. Quetext let me inspect individual sentences. Neither gave me a completed human-control result.
Other completed checks went wrong. QuillBot labeled every generated passage 100% human-written. SciSpace called the essay Essentially Human, with 2% AI Detected. Proofademic’s public demo settled at zero AI-likeness for the essay, and Surfer assigned no chance of AI writing to either generated passage I submitted.
When I checked Carroll, Clever, Pangram, Originality.ai, AI or Not, Sapling, QuillBot, Proofademic, and Surfer accepted it as human. ZeroGPT instead returned 99.3% AI, higher than any generated passage in its completed set. NoteGPT also gave Carroll its strongest AI probability. Similar numbers across products did not establish shared technology.
The lightly edited Carroll passage complicated the distinction further. Clever returned Human 99%, and Pangram treated it as entirely human. I could not count that passage as another pure-human control because AI had participated in it. Clever’s sentence colors sometimes disagreed with its headline, and some explanation tags were unhelpful.
Recording the limits
As submissions stopped, I separated access problems from detection errors. QuillBot ended my guest session before the mixed passage, despite advertising a larger daily allowance. SciSpace requested signup after the essay. Proofademic’s results covered only its public demo, not authenticated reports.
Decopy opened a login dialog instead of returning a score. Smodin displayed highlights but blurred its headline percentages, so I had no readable numeric report to assess. AI Detector Pro required registration. Winston AI presented trial registration and terms acceptance before scanning; I stopped without a result.
Copyleaks loaded a guest interface, but I stopped over its applicable-use terms. Turnitin required institutional access and a license add-on. Neither belonged in an accuracy ranking from these attempts.
Hive’s inspected demo accepted media rather than prose; I did not install its text extension or configure its API. ContentDetector.AI led to a Moxby extension listing, which I did not install. Humanly remained an inspected Apple store listing, with a free download and in-app purchases but no verified scanning allowance.
I retained the screenshots below as records of results and access screens, not interchangeable evidence of accuracy. Vendor previews did not represent checks of my passages. I had not tested paid workflows, multilingual coverage, or a broad sample of modern human writing.
On the Scribbr follow-up, after an earlier attempt produced no usable result, both the generated essay and Carroll received human-written labels. That accepted the control but missed the essay. I also saw QuillBot branding there, so I would not automatically count the two sites as independent votes.