AI Detector Short Text Accuracy: When 100 Words Is Not Enough

A single paragraph is usually too little text for an AI detector to classify reliably. Scores on short samples swing with every sentence, so treat a 100-word result as a weak signal, not a verdict. Build a longer sample from your own writing before you act on any number.

Key takeaways

• Short text gives detectors too few signals to separate human and machine patterns, so the score is unstable.
• The same paragraph can score differently each time you paste it, because the sample is near the detector’s decision boundary.
• Assemble at least a few full sections from your own drafts before you judge anything.
• A 100-word flag is not proof of authorship and should not be treated as one.
• If someone judges you on a short-text score, ask what sample length and what threshold produced it.
• Sentence-level views tell you which lines carry the signal, which matters more than the headline number.

Why does a 100-word sample give a detector so little to work with?

Detectors do not read for meaning. They estimate the statistical shape of your text: how predictable each next word is, how much sentence length varies, how often certain connective phrases appear. Those are distributional signals, and distributions need volume. A hundred words is roughly one short paragraph, maybe five or six sentences. That is a very small sample to estimate a distribution from.

Think about what a detector is trying to measure. Two of the common signals are perplexity (how surprising the word choices are to a language model) and burstiness (how much sentence length and structure vary). A human writer might produce one long, winding sentence followed by a short one. An unedited model tends to produce sentences of similar length and rhythm. But in a single paragraph, one short sentence can look like a burst, and a run of medium sentences can look flat, purely by accident. The signal is real, but the sample is too small to see it clearly.

There is also a boundary problem. Detector output is a probability estimate, not a measurement. When the true signal is weak, the estimate sits near the decision threshold, and small changes push it across. Add one clause and the score moves. Remove one em dash and it moves again. That is not the detector being broken. It is what happens when you ask a statistical model to classify a sample that does not contain enough evidence.

Our own instrumentation shows how noisy this gets. On 2026-09-19 we scored two corpora with the DeAIze detector (local ensemble, lite tier) at a 30% flag threshold. A human-written sample of n = 46 hand-written DeAIze guides had a mean AI score of 27% and a median of 25%, with a range from 0 to 55%. At that threshold, 17 of 46 were flagged. Unedited model output, n = 12 passages generated from fixed prompts, had a mean of 35% and a median of 34%, range 25-49%. Notice the overlap. Even with much longer samples than a single paragraph, the two distributions touch. Shorten the sample to 100 words and that overlap widens.

How does sample length change the score on the same text?

Length does not just make the score more precise. It changes which signals dominate.

A short sample is dominated by local features: a repeated transition word, a uniform sentence rhythm, a phrase that happens to be common in model output. A long sample lets the detector average those out and look at broader tendencies across sections. The same writer can look machine-like in one paragraph and clearly human across five pages, because the human variation lives in the longer arc.

The practical consequence is that a short-text score tells you about that paragraph, not about you. If you want a number that reflects your writing, you need enough of it.

Sample size	What the score mostly reflects	How much to trust it
One paragraph (about 100 words)	Local word choices and one or two rhythm patterns	Low. Treat as a hint, not a result.
A few paragraphs	A mix of local habits and some variation	Moderate. Useful for spotting flagged sentences.
A full section or more	Broader rhythm, vocabulary range, structural habits	Higher. Worth acting on.
A whole document	The overall writing profile	Highest available, still a probability estimate.

That last row matters. Even a full document score is a probability estimate. It is not proof that a machine wrote the text, and no detector should be treated as if it were.

If you want to see how length changes the picture on your own writing, run a paragraph and then a full section through the AI detector and compare the sentence-level highlights. The difference between the two runs is the clearest lesson in short-text accuracy you will get.

How do you assemble a longer honest sample from your own writing?

The honest way to test your writing is to test your actual writing. That sounds obvious, but people often do the opposite: they paste a paragraph they wrote under pressure, get a bad score, and conclude something about themselves.

Here is a method that works.

1. Pick three to five pieces you wrote without AI assistance. Old emails, a report, a blog draft, lecture notes, a long message to a friend. Anything you typed yourself.
2. Strip out quoted material, citations, code blocks and lists. Detectors read prose differently from structured content, and quoted text is not yours to defend.
3. Combine the prose into one document of at least several hundred words. Keep the original punctuation and sentence lengths. Do not tidy it up.
4. Run the combined sample through a detector that shows sentence-level output. A single number hides the useful part.
5. Note which sentences are flagged. Are they the ones you wrote quickly? The ones you edited heavily? The ones you drafted with AI help?

Step five is where the value is. A flagged sentence is a specific, local thing you can look at. A flagged document is a vague accusation. If you are being judged on a short sample, being able to point to the specific lines and explain how you wrote them is far more useful than arguing about the headline number.

One more thing: keep the sample. If a score is ever questioned, you have a dated file showing what your unaided writing looks like. That is the closest thing to a baseline you can have.

When should you ignore a short-text score entirely?

There are clear cases where a short-text result carries almost no information and you should set it aside.

• The sample is under a few hundred words. There is not enough distribution to estimate. The number is mostly noise.
• The text is heavily quoted or cited. Proper nouns, citations and technical terms are unusual to a language model, which can push the score in either direction for reasons that have nothing to do with authorship.
• The text is a list, a table or a set of bullet points. Detectors are trained on prose. Structured content is out of distribution.
• The writer is non-native. Writers working in a second language often produce more uniform sentence structures and more common word choices, which can look machine-like to a detector. This is a known pattern, not a personal failing. If this is your situation, the guide on why an AI detector flags non native English writers is worth reading before you change a word.
• The text has been translated. Machine translation leaves its own artefacts, and those artefacts are not authorship signals. The translated text AI detection walkthrough covers how to separate the two.
• The detector gives a different answer on a second run. If the same text scores differently on consecutive runs, the sample is too close to the boundary to mean anything.
• The score is being used as proof. It is not proof. It is an estimate, and treating a short-text estimate as evidence is a category error.

If any of these apply, the right response is not to rewrite until the number drops. The right response is to get a longer, cleaner sample and re-run.

What do you tell someone who is judging you on a 100-word result?

This is the awkward part, and it deserves a direct answer rather than reassurance.

Start by asking what the score actually is. Which tool produced it? What sample length was used? What threshold counts as flagged? Many people quoting a number do not know its provenance. Asking is not defensive; it is the first step of any honest evaluation.

Then explain the sample problem in plain language. A detector estimates the statistical shape of text. A hundred words is too little to estimate much of anything. You are not disputing that the tool produced a number. You are pointing out that the number is not strong enough to support a conclusion.

Offer a longer sample. Volunteer to run a full section or a whole document through the same tool. If the result holds at length, that is a real conversation. If it does not, you have shown why length matters.

Finally, be honest about your process. If you used AI to draft or edit part of the text, say so, and describe what you changed. Most policies that allow AI assistance require disclosure, not perfection. The academic integrity policy guide explains what many institutions actually permit. The ideas, data and conclusions still have to be yours.

What you should not do is treat a short-text score as a verdict and start rewriting your voice out of your own writing to satisfy it. That trades a real thing for a number that was never reliable in the first place.

What does a sentence-level view give you that a single score does not?

A single number collapses everything into one figure. A sentence-level view shows you where the signal actually lives, which is the part you can act on.

On a short sample, sentence-level output is often the only useful thing available. If two of five sentences are flagged and three are not, you have learned something specific: those two sentences share a pattern. Maybe they are the same length. Maybe they both open with the same kind of transition. Maybe you wrote them last, quickly, at the end of a long day. That is a concrete editing target.

On a long sample, sentence-level output tells you whether the flags are clustered in one section or spread evenly. Clustered flags usually mean something local happened in that section. Spread flags suggest a broader pattern in how you write, which is a different problem with a different fix.

DeAIze detection works this way: the whole document gets a score, and each sentence is highlighted by AI probability, so you can see which lines carry the signal. That structure is more useful for short-text questions than a single percentage, because it lets you check the detector’s reasoning instead of accepting or rejecting it wholesale.

For a deeper look at what the underlying signals mean, the how AI detection works explainer covers perplexity, burstiness and the rest without overclaiming.

How should you act on a short-text score?

Treat it as a triage signal, not a diagnosis.

If a short sample comes back high, do three things. First, check the sentence-level output and see whether the flags cluster on lines you can explain. Second, build a longer sample from your own unaided writing and re-run. Third, compare the two results. If the longer sample drops, the short result was noise. If it holds, you have something worth editing.

If a short sample comes back low, do not treat that as clearance either. A low score on a small sample is just as weak as a high one. It means the detector did not find much signal, not that the text is certified human.

This is the honest position: detector output is a probability estimate, and the strength of that estimate depends on how much text it was computed from. Short text, weak estimate. Long text, stronger estimate. Neither is proof.

Ready to see which sentences read machine-made?

Paste a paragraph and you will get a number. Paste a full section and you will get something you can actually use: a document score plus sentence-level highlights showing which lines carry the signal.

Start with the AI detector to see the sentence-level breakdown on your own writing. If the flags cluster on lines you wrote with AI help, run them through the humanizer and check the before-and-after scores. The site reports an average AI-score reduction of 60%+ after humanizing, best case 71%, and on the example paragraph used on the site a 92% AI score becomes 4% AI. Those are the product’s own predicted scores, not a third-party verdict.

You can try it free with 200 words on signup, no credit card. If you need more, pricing is pay-as-you-go, credits never expire, no subscription. Uploads accept .txt, .docx and .pdf, and rewriting a Word document preserves fonts, styles and tables. When you are ready, create an account and start with the sample you already have.

The point is not to chase a number. It is to know which sentences read machine-made, decide whether you agree, and fix the ones you do.