Home / Guides

Do AI detectors work on short text?

Mostly no. Short text is where detectors fail hardest: under a couple hundred words, scores swing wildly and mean very little. Here is why length matters and what to do instead.

Updated October 31, 2026 · By the IsThisAI? team

The short answer

Mostly no. Short text is where AI detectors fail hardest. Under a couple hundred words, scores swing wildly and mean very little. The shorter the text, the less you should trust the number.

Why length matters so much

Detectors look for statistical patterns across many words: how predictable the word choices are, how varied the sentence rhythms, how "bursty" the structure is. These are aggregate signals. A 30-word paragraph simply does not contain enough of them to measure. It is like trying to judge someone's vocabulary from a single sentence.

This is not a flaw in one particular tool. It is structural. Every published evaluation of detectors finds the same cliff: accuracy degrades sharply as text gets shorter. A detector that is reasonably calibrated on a 1,500-word essay is guessing on a 100-word discussion post.

What counts as "short"

• Under ~100 words: scores are close to noise. Tweets, headlines, and short answers cannot be meaningfully scored.
• 100 to 300 words: weak signal at best. Treat any score in this range with heavy skepticism.
• 300 to 1,000 words: the gray zone. Scores start to mean something but remain unreliable as evidence.
• Over 1,000 words: where detectors do their best work, and still only as a probability.

The traps short text sets

Short, polished writing is exactly what triggers false positives. A carefully written three-sentence email, a concise product description, or a non-native speaker's short paragraph can all score high for being "too clean." Meanwhile a short AI output that was lightly edited by a person can score low. On short text, both kinds of error are common.

Discussion posts are the classic problem case. A 150-word forum reply is short, often formulaic ("Great point, I agree because..."), and exactly the kind of text detectors misread. Schools that run detectors over discussion boards generate false accusations at scale.

What to do instead

If you only have a short passage, do not run it through a detector and act on the result. Either gather more of the person's writing and check the longer sample, or skip the detector entirely and judge the work on its merits. More words will not make a detector perfect, but fewer words reliably make it worse.

Bottom line

Length is the single biggest factor in whether a detector score deserves any attention. Short text, short trust. If the passage fits in a screenshot, the score is not evidence.

Why vendors stay quiet about this

Detector companies market accuracy numbers measured on long, clean test sets: full essays, articles, and papers. Those numbers do not transfer to short text, but the marketing rarely says so. Some tools refuse to score very short inputs at all, which is the honest choice. Others return a confident-looking percentage on 50 words, which is worse than refusing: it dresses a guess up as a measurement.

If a tool gives you a precise score on a tweet-length passage, that precision is theater. Ask the vendor how their accuracy was measured and on what text lengths. The serious ones publish this. The ones that do not are telling you something too.

How much text is enough, exactly?

There is no magic threshold, but the research points to a rough ladder. Below about 100 words, treat scores as noise. From 100 to 300 words, you have a weak hint, useful only alongside other evidence. From 300 to 1,000 words, scores start to stabilize but remain unreliable as standalone evidence. Above 1,000 words of connected prose, detectors do their best work, and even then the result is a probability with a real error rate.

Genre matters as much as length. A thousand words of bullet points and headings is harder to score than a thousand words of essay prose. When you can, check the most prose-like, most personal part of a document: the introduction someone wrote freely beats the templated middle every time.

Sources

1. Weber-Wulff et al., "Testing of detection tools for AI-generated text," International Journal for Educational Integrity 19, 26 (2023). https://link.springer.com/article/10.1007/s40979-023-00146-z
2. Liang, Yuksekgonul, Mao, Wu & Zou, "GPT detectors are biased against non-native English writers," Patterns 4(7), 2023. https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7
3. OpenAI, "New AI classifier for indicating AI-written text" (Jan 2023; withdrawn July 2023). https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/

This guide is general information, not legal or hiring advice. AI detection results, including ours, can be wrong.