How accurate are AI detectors on short texts like a 150-word discussion post?
Short texts are where detectors are weakest. A 150-word post is only around 800 to 1,000 characters, and vendors themselves say accuracy improves with length. OpenAI withdrew its own classifier over low accuracy, and GPTZero says document-level results beat paragraph and sentence level. Treat any score on a post this size as a weak hint, never as evidence.
Cite
Global 100 Forum, "How accurate are AI detectors on short texts like a 150-word discussion post?", https://global100.org/forum/q/how-accurate-are-ai-detectors-on-short-texts-like-a-150-word/, accessed 2026-11-10.I TA for a writing course, and weekly discussion posts are capped at about 150 words. An instructor in another section has started pasting every post into a free detector and sending students a "please explain" email when the tool highlights something. A few of my own students have now asked me whether a post that short can be assessed at all.
I wanted a rough sense of it before answering, so I ran two posts I wrote myself years before any of these tools existed. One came back clean. The other had a sentence highlighted as likely AI, and that sentence was just me restating the reading in the reading's own vocabulary.
Two questions. Is there any published guidance, from vendors or researchers, on how much text a detector needs before its output means anything? And if there is a floor, what do I tell students whose posts sit well under it?
Write something first.
Give people something to work with: at least 30 words on what happened and what you tried.
That is too long. Keep it under 6,000 characters.
Write the question as the title, 15 to 140 characters, no links.
Pick a category.
Add a name (2 to 40 characters, no links).
That email address does not look right.
That was quick. Read the thread, then try again.
The form expired. Reload the page and post again.
Something went wrong with the form. Reload and try again.
Please complete the check and post again.
Limit reached for now. Try again later.
This thread is closed to new replies.
This thread no longer accepts replies.
Something went wrong with the form. Reload and try again.
Post a reply or question first (name and email), then this browser can vote, edit and accept answers.
You cannot vote on your own post.
Only the author (within 30 days) or the forum team can do that.
That email already has an account. Use your sign-in link instead.
New accounts are paused for the moment. Try again later.
That was already posted.
7 answers
- Accepted answerMaya LindqvistStaffEdits the AI text detection and academic integrity sections ·
Short texts are where detectors are weakest. A 150-word post is only around 800 to 1,000 characters, and vendors themselves say accuracy improves with length. OpenAI withdrew its own classifier over low accuracy, and GPTZero says document-level results beat paragraph and sentence level. Treat any score on a post this size as a weak hint, never as evidence.
Here is what the vendors actually say. When OpenAI launched its text classifier in January 2023, it stated that the tool was "very unreliable on short texts (below 1,000 characters)" and "should not be used as a primary decision-making tool." Its own challenge-set results were that it flagged 26% of AI-written text as likely AI and wrongly labelled 9% of human text as AI. That is an old model, so please do not read those two figures as the performance of any tool today, but the short-text caveat is exactly your situation. OpenAI took the classifier down in July 2023 "due to its low rate of accuracy."
GPTZero's FAQ makes the same point from the other direction. It says accuracy "increases as more text is submitted," so document-level classification is better than paragraph level, which is better than sentence level. It also says the classifier can flag "highly procedural text" and that results "should not be used to punish students."
Your highlighted sentence fits that pattern: a restatement of a reading, in the reading's vocabulary, inside a word cap, is low-information text. My advice is not to screen discussion posts with a detector at all. If a post looks off, compare it with the same student's earlier posts and ask them to talk through the argument. See also how accurate AI text detectors are, really.
Sources:
- Dan OkaforMember ·
Secondary school, so my "short texts" are exit tickets and three-sentence responses, and I can tell you what happens. The kids who write in the plainest, most on-topic way get flagged most. A student summarising a paragraph in the paragraph's own words looks a lot like a language model summarising it.
I stopped running anything under a full page. At that length there is nothing for the tool to work with, and I am left with a number I cannot defend to a parent.
- Lena B.Member ·
Product descriptions, 100 to 150 words, plain and factual. That is most of what I write for some clients. Those get flagged too.
Short, tidy, on-brief copy is what these tools read as machine-like. I now keep my drafts and send them with every delivery. Boring, but it ends the argument in about one email.
- Sam WhitlockMember ·
Dry observation: a detector looking at 150 words is estimating from very few observations, so the error bars are wide before any model quality enters. A confident-looking single number hides that. If a tool does not show you its uncertainty, assume it is large on short inputs and act accordingly.
- Hana SatoMember ·
@Aisha R. I would not rely on that. I am not aware of any published threshold like 50 words, and the two vendor pages Maya quotes say the opposite: reliability rises gradually with length and short text is the weak spot. If your instructor has a source for 50 words, ask to see it. It would be useful to read.