An AI detector is a statistical guess with a percentage sign attached. The panel on this page is Clever AI Detector, embedded from a third-party service. It reads your text, compares it with patterns it has learned from generated and human writing, and returns a number. It does not know who wrote your draft, and it cannot prove that a person or a language model produced it.
That is worth saying before the tool, because detector scores now turn up in arguments about essays, job applications and freelance work. A false positive in any of those costs someone real time and trust, and the number is far weaker evidence than it looks.
How this page works
Paste your text into the panel, press Check for AI, and the result appears in the same frame: a percentage labelled AI probability, a short label such as Likely AI-generated, and a HUMAN to AI bar. The panel's own footnote adds the part that matters most: AI detection is an estimate, not proof of authorship.
When this page was checked, the embed set an 80-word minimum and a 3,000-word ceiling, and offered one mode, Normal. Anything shorter than the floor is refused before it is analysed, which is a fair reminder that a sentence or two carries almost no signal. The analysis runs on Clever's servers, not in your browser.
The panel is a cross-origin frame, so this page cannot read what is inside it, point at the sentences that moved the score, or keep a history for you. You work inside the widget's own buttons from start to finish.
What the number actually says
A score of 79% does not mean 79% of the document came from a machine, and it does not mean there is a 79% chance a machine wrote it. It means the text sits closer to the detector's generated-text patterns than to its human-text patterns, measured on a scale the vendor keeps to itself. The percentage looks precise because it is a number. The judgement underneath it is not.
Read the label as well as the figure. Clever changes the wording at points it chooses, so Likely AI-generated and a plain percentage are two different claims. A result sitting in the middle of the scale is the least informative outcome of all: it means the model found evidence on both sides and cannot separate them.
Sample size matters as much as the score. A 100-word paragraph gives the model very little to work with, so small edits can swing the number by tens of points. A few hundred words of consistent prose is a more stable input, which is why the floor exists.
Why careful writing gets flagged
Here is a measured example rather than a general warning. The paragraph below was written by hand for this page, 89 words, no model involved. Pasted into the panel, it scored 79% AI probability and came back labelled Likely AI-generated.
Walking the same route every morning teaches you things a map cannot. You notice which bakery puts bread out at seven and which one waits until nine. You learn that the third crossing is faster in rain because the traffic light there favours pedestrians, and that the short cut behind the school floods after a heavy night. None of this is written anywhere. It arrives slowly, one repetition at a time, and it makes the street feel less like a corridor and more like a place where you live.
That is not a bug so much as the nature of the method. The paragraph keeps an even register, names nothing specific, and runs sentences of similar length. Those are the traits detectors associate with generated text, and they are also the traits of edited, careful prose.
Other things that push human writing toward the AI end of the scale:
- a formal or academic register, passive voice, and paragraph shapes repeated across a document;
- text run through a grammar assistant, which smooths out the irregularities detectors treat as evidence;
- writing in a second language, where careful textbook sentence patterns read as machine-like;
- technical, legal and medical prose, where precision means uniform structure;
- short samples, where there is too little text to find a rhythm in;
- text that was outlined, translated or summarised by a tool and then rewritten by a person.
None of that says anything about authorship. It describes style, and style is what the detector is actually measuring.
How to read a result fairly
Run a baseline first. Paste something you know is human: an old email you wrote, a page from a printed book, a letter. If that scores high as well, the number is telling you about the detector's calibration rather than about your draft.
Give it enough text. A 90-word sample is a coin flip with decimal places. Three hundred to six hundred words of your own writing is a steadier input and still well inside the 3,000-word ceiling. If the result matters, run the same text twice: detectors are not deterministic between runs, and a score that moves on a repeat is a score you should not lean on.
Then read the text itself. Look for what the detector responds to and a human reader would also notice: vague claims, no specifics, repeated sentence shapes, a paragraph that restates its own heading. Fixing those improves the writing whether or not the percentage moves.
Keep the score in proportion. A detector output is weaker evidence than a draft history, a version-control log, your notes, or your ability to explain a sentence you supposedly wrote. When you have those, they settle the question better than any bar chart.
If your text comes back flagged
Do not rewrite to chase the number. Scores move for reasons unrelated to authorship, so editing until the bar turns green teaches you to imitate the detector rather than to write better.
If the text is going to someone who may run the same check, the moves that help are concrete:
- put in the details only you have: names, dates, figures, the thing that was said in the meeting;
- break the rhythm, one short sentence and then a longer one, instead of six of the same length;
- cut the throat-clearing opening and the closing sentence that repeats the title;
- keep your drafts, notes and sources, because they are the evidence a detector cannot see;
- if you used AI for part of the work, say which part. Disclosure is a stronger position than a score.
If a decision about a person hangs on the result, ask for the text to be read by a human. A detector is not a witness, and the vendor says as much inside its own panel.
What happens to the text you paste
Loading the panel makes a request to widgets.cleverhumanizer.ai, which sees your IP address and ordinary request headers. When you press Check for AI, that service receives your text and returns the score.
The panel's Text privacy notice states that text and results are stored for 24 hours and can be viewed by the widget's owner and Clever administrators. The owner here is the ToolsHub4u account that installed the widget, which means what you paste can be read by this site's operator as well as by Clever's staff. We do not copy it anywhere else or use it for anything else, but do not treat the panel as private.
Keep out client documents under confidentiality, personal data, unpublished research, interview transcripts and anything covered by an agreement that does not name this service. The other text tools on this site run inside the page and send nothing: use the word counter or the text comparison tool for a draft you cannot send to anyone.
Limits worth knowing before you rely on it
The panel needs JavaScript and a working connection. A content blocker, a privacy extension or a corporate filter can stop the frame from loading; the link under it opens the same detector in a new tab, which tells you whether the block sits on the page or on the network.
It takes text only. There is no PDF, document or image upload, no batch run over a folder of files, and no export. There is no sentence-level breakdown either, so you cannot see which paragraph carried the result. The embed offers one mode, and two runs on the same text can disagree.
It is not a plagiarism checker and will not tell you whether text matches a published source; that needs a search of the actual passages. It keeps no history here, and the panel is cleared when the page reloads.
When not to use it
Not as the only evidence in a disciplinary or hiring decision. A false positive can follow someone for years, and the score is explicitly an estimate.
Not on non-English text. The embed runs in one mode, and how the model behaves outside English is not something this page can verify.
Not on code, formulas or structured data. Detectors score prose style, and an ordinary script can look mechanical because it is.
And not as a target. If the only reason to change a sentence is that a detector disliked it, the change is probably making the writing worse, which is the one outcome nobody wants.