8 Texts, 1 Detector: A Human Paragraph Scored 99% AI

Our glossary of AI detection terms makes a repeated argument: a detector score is probabilistic evidence, not proof of authorship. That is easy to assert. It is harder to feel how wide the error bars are until you watch a detector label a paragraph you know to be human-written as "99% AI probability."

So we ran a small, fully disclosed test. Four AI-generated texts and four human-written texts, all between roughly 110 and 160 words, went through one detector - the same Clever AI Detector whose free widget is embedded on our own AI detector page. Every score below is real, recorded on October 17, 2026, including the one that should make anyone cautious about treating detector output as evidence.

Full disclosure before the results

Three things, because honest research states its constraints up front:

• The detector is one we embed. The Cleverhumanizer widget powers the free detector on our site. We did not run this test to promote it or to criticize a competitor; we ran it on the tool our readers actually use, and we report what it did - including a false positive on human text.
• Eight texts is a tiny sample. This is an illustration, not a benchmark study. It cannot produce an accuracy rate; it can show, concretely, the failure mode the research literature warns about.
• The AI texts came from us. They were generated by MiMo V2.6, the assistant model used in RewriteAI's editorial workflow, at default sampling settings, for the specific purpose of this test. That disclosure is itself the policy we tell writers to follow.

The eight texts

The AI texts cover common use cases: social-media marketing prose, a product description, an internal policy email, and an opinion-essay intro on social media and teenagers. All were written by the model in a single pass, unedited.

The human texts were chosen to span different registers of genuine human writing, all license-safe to quote:

• Wikipedia (CC BY-SA): the opening of the article on photosynthesis - formal, encyclopedic, human-edited prose.
• arXiv: the abstract of "Attention Is All You Need" (Vaswani et al., 2017) - dense, modern scientific writing by human authors.
• Jane Austen (public domain): the opening of Pride and Prejudice, from Project Gutenberg.
• Henry David Thoreau (public domain): the "I went to the woods" passage from Walden, from Project Gutenberg.

Nothing was edited, translated, or smoothed. Each text went into the detector exactly as shown, within the widget's stated 80-to-3,000-word range.

The results

Two clean findings and one uncomfortable one:

• Every AI text was flagged - scores from 90% to 99%. For ordinary, single-pass model output on everyday topics, this detector works as advertised.
• Three of four human texts were correctly identified - scores from 1% to 12%, including two centuries-old literary passages.
• One human text was called 99% AI. The Wikipedia introduction to photosynthesis - written and curated by humans, licensed CC BY-SA, with the same score as our most obviously machine-flavored marketing sample. This is a textbook false positive, and it is not a rounding error: the detector was as confident about it as it was about any AI text in the set.

Why did the Wikipedia paragraph fail?

The glossary's vocabulary explains it better than intuition can. Encyclopedic prose is optimized to be low-perplexity: neutral register, standard constructions, terminology used in its most statistically expected sense, sentences with uniform length and structure, content words distributed with unusual evenness - low burstiness. Those are precisely the statistical properties detectors learn to associate with machine output, because they are what model-generated text looks like by default.

In other words, the detector plausibly was not detecting "AI authorship" so much as detecting "textbook register." The Wikipedia paragraph shares its statistics with machine text; Austen and Thoreau do not, and the arXiv abstract - human, dense, but written with a distinctive authorial hand - landed at 12%. This failure mode is exactly what Liang et al. documented at scale when detectors systematically mis-flagged non-native English writers: constrained, careful prose gets read as machine prose, and the writers least able to defend themselves absorb the penalty.

What this means in practice

• Never treat a score as a verdict. A 99% score attached to a human Wikipedia paragraph is the clearest possible demonstration that confidence percentages describe the detector's model, not the writer's laptop.
• Know what the register trap is. Formal, edited, terminology-dense human writing - academic abstracts, encyclopedic entries, corporate documentation, careful non-native English - sits dangerously close to machine text statistically. If you write in those registers, expect noisy scores and insist on human context before any decision.
• Cross-test before believing. One detector, one opinion. A score that says 99% on one tool and 20% on another is worth exactly neither number - it is a reason to read the text instead.
• The safe workflow is review, not screening. This is the same conclusion Google's publisher guidance reaches and the same one our humanizer workflow assumes: read the text, verify the claims, judge the authorship from evidence beyond a probability.

Limitations, plainly stated

This test has real constraints and they cut against over-reading it in both directions. The sample is eight texts. The AI and human groups cover different topics and registers - which is part of why the Wikipedia text failed, but also means we cannot separate "detector bias" from "topic effects" here. Each text was run once, on one day, through one detector's free tier; detector outputs can vary with model updates. Literary passages from the 1800s are an easy human class that says little about how the detector treats modern blog prose. And yes - the detector is embedded on our own site, which is why the disclosure sits at the top of this article rather than in a footnote.

What the test does establish is the thing worth establishing: on this detector, on this day, a confidently scored human paragraph scored 99% AI. That single data point, honestly reported, is worth more than any accuracy percentage printed on a marketing page.

Frequently asked questions

Can an AI detector give a human text 99% AI?

Yes. In our test on October 17, 2026, the Clever AI Detector scored the human-written Wikipedia introduction to photosynthesis at 99% AI probability - the same score it gave an obviously machine-flavored marketing paragraph. Published research explains why: formal, low-variance, encyclopedic writing shares statistical properties with model output, and detectors trained on those statistics misread the register as authorship.

Are AI detectors accurate on AI-written text?

On our four AI-generated texts, the detector flagged all four, with scores from 90% to 99%. That matches the pattern in the research literature: detection of straightforward, single-pass model output is usually good, and reliability degrades with edited, paraphrased, or unusually styled text - and on constrained human writing, in the other direction.

Should students worry about false positives?

They should understand them. Research by Liang et al. (2023) found widely used detectors consistently mis-classified non-native English writers' work as AI-generated, which is why the authors advised caution about detector use in evaluative or educational settings. If a score is being used against you, ask for the evidence behind it: a percentage is an estimate, not proof, as the detector's own interface states.

Does this test prove detectors are useless?

No, and we are not claiming that. All four AI texts were caught. The point is narrower and sturdier: a score is one probabilistic signal with documented, reproducible failure modes on human text - including a 99% confident error in our own small sample. Use detectors as one input among several, never as the decision itself.

Sources and text provenance

• Wikipedia: Photosynthesis - opening text, licensed CC BY-SA
• Vaswani et al.: Attention Is All You Need (arXiv:1706.03762) - abstract
• Project Gutenberg: Pride and Prejudice (public domain) - opening passage
• Project Gutenberg: Walden (public domain) - "I went to the woods" passage
• Liang et al.: GPT detectors are biased against non-native English writers (2023)
• RewriteAI: AI detection glossary - perplexity, burstiness, and the rest