Are AI Detectors Accurate on Short Essays?
UC PIQs cap at 350 words; supplement limits we checked run 50 to 200. Why AI detectors are unreliable on short essays, and how to read a flag on one.
Are AI Detectors Accurate on Short Essays?
Not reliably, and the published warning is about exactly your length. ChatGPT's developer said its own classifier was "very unreliable on short texts (below 1,000 characters)", roughly 170 to 200 words, then withdrew it. On a 150-word supplement, read the highlighted sentences, not the percentage.
Pages that answer this question usually review detectors, or quote one overall accuracy figure. Neither helps with an application, because an application is mostly short answers. A UC applicant writes four answers of up to 350 words each. Stanford's short questions stop at 50 words. This page puts each answer length next to the published short-text warning, then gives a procedure for a high score on a short answer.
Every college and platform page quoted below was checked on October 30, 2026.
The length math for your application
To compare word limits with a character threshold, we converted words to characters. The two constructed answers on this page run 5.0 and 6.0 characters per word, including spaces (591 characters for 117 words; 667 characters for 112 words). The table uses that range. It is our arithmetic, not a published conversion.
| Answer | Limit, as published | About this many characters (our arithmetic) | Against the 1,000-character line |
|---|---|---|---|
| Common App personal essay | "no more than 650 words" (2026 application) | 3,250 to 3,900 | Above |
| UC personal insight question | "a maximum of 350 words" each (UC) | 1,750 to 2,100 at the limit | Above at the limit; a 160-word answer is below |
| Common App challenges question | "Max word count: 250" (2026 application) | 1,250 to 1,500 | Just above |
| Yale academic interest answer | "(200 words or fewer)" (Yale, 2026-2027) | 1,000 to 1,200 | At the line |
| Penn short answers | "(50-200 words)" (Penn) | 250 to 1,200 | Below to at the line |
| Caltech community answer | "(Min: 100 / Max: 150 words)" (Caltech) | 500 to 900 | Below |
| Stanford short questions | "a 3-word minimum and a 50-word maximum for each question" (Stanford) | 250 to 300 at the limit | Far below |
Penn's page is labelled 2026-27 and Yale's 2026-2027; Stanford's page states no cycle year. For more limits, see the 2026-2027 supplemental essay word limits guide.
Our view: most of what you write after the personal essay sits at or below the length where the only detector maker's length warning we found said its tool was "very unreliable." Basis: the table above. The countercase: a full 350-word PIQ or a 650-word essay is well above that line. For those, the length warning does not apply, but the published false-positive rates still do.
What detector makers say about length
The clearest statement we found comes from ChatGPT's developer. On January 31, 2023, it launched an AI text classifier and wrote:
"The classifier is very unreliable on short texts (below 1,000 characters). Even longer texts are sometimes incorrectly labeled by the classifier. Sometimes human-written text will be incorrectly but confidently labeled as AI-written by our classifier."
It also wrote that "Our classifier's reliability typically improves as the length of the input text increases," and reported that in its tests it labeled human-written text as AI-written "9% of the time". The page now carries this update: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." We read the archived copy; the live page returned HTTP 403 to our request.
That is one tool from 2023, not a measure of today's detectors. It is still the maker's own statement about its own tool. We tried to check a second vendor's published minimum length and could not load its help page, so we do not quote one. For independent error-rate data across tools, see AI detector false positive rates compared.
Why short answers look AI-like even when a person wrote them
Three things work against short answers:
- Less text, less evidence. The classifier's own wording, that reliability "improves as the length of the input text increases", says the same thing from the other side.
- Short prompts invite formula. "Why this major" and "list your favorites" questions push everyone toward the same moves: a childhood spark, a list of qualities, a closing about making a difference. Formula is not AI, but it reads as generic to people, and the classifier's maker warned that "Text that is very predictable cannot be reliably identified."
- Non-native English was flagged more in one study. Liang and colleagues (arXiv 2304.02819, published in Patterns, 2023) found that the detectors they tested "consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified."
There is also a floor no tool can get below. Our page on why AI detector false positives are mathematically unavoidable explains the base-rate problem, and what flagbait is lists the phrases that tend to trip detectors.
What to do with a high score on a 150-word answer
Run this procedure once, then stop.
- Ignore the overall percentage. On an answer under about 1,000 characters, it is the number the classifier's maker called "very unreliable."
- Look at which sentences are highlighted. If the tool highlights nothing specific, there is nothing to fix. Go to step 5.
- For each highlighted sentence, ask: could another applicant have written it word for word? If yes, replace its category words ("passion", "difference", "community") with one concrete detail only you know: a date, a place, an object, a number.
- Read the answer aloud. If a sentence sounds like a brochure, cut it. Short answers have no room for it.
- Stop. Do not run three more checkers. Scores on short text will vary, and chasing a lower one leads to worse writing.
The test in step 3 is a reader's test, not a detector's. That is the point: the fix that lowers suspicion is the same fix that makes the answer better.
Constructed examples: two answers to "Why this major?"
Two constructed examples, written by us, not real applicant statements. Both answer an invented prompt: "Why are you interested in mechanical engineering?" Both are under 1,000 characters.
Answer A, formulaic but human (112 words, 667 characters):
Ever since I was young, I have been fascinated by how things work. This curiosity led me to take every science class my school offered. I discovered that engineering combines my passion for problem-solving with my desire to make a difference in the world. Through my coursework and activities, I have developed strong analytical skills and a collaborative mindset. Studying mechanical engineering will allow me to grow as a learner, a leader, and a member of my community. I am excited to explore new challenges, connect with peers who share my interests, and use my education to create solutions that improve people's lives. This major is the right fit for my goals.
Answer B, specific (117 words, 591 characters):
In March, our school's 3D printer jammed, and the teacher who knew how to fix it had left. I took the print head apart on a lab bench, pausing a repair video every ten seconds. The fault was a worn plastic gear the size of my thumbnail. I printed a replacement in nylon, and it lasted until June. That repair is why I want mechanical engineering. I like problems I can hold in my hand, and I want to know why that gear wore down on one side only. In my first year I want to take the statics and materials courses that would let me answer that question, and design a gear that would not fail.
What a reader would question in Answer A: "Ever since I was young, I have been fascinated by how things work" (any applicant could sign it); "combines my passion for problem-solving with my desire to make a difference in the world" (two abstract nouns, no evidence); "grow as a learner, a leader, and a member of my community" (a three-part list with nothing behind it); "create solutions that improve people's lives" (a closing that fits every major).
Why Answer B holds up: every sentence has something checkable in it: a month, a broken part, a material, a date it lasted until, a question about wear. No other applicant has that gear. Both answers are human. Only one of them gives a reader anything that looks like a particular person.
For more on this prompt, see how to answer "what do you want to study and why".
If your short answer was flagged and you wrote it
Keep your drafts and your version history; what to do if your essay was flagged as AI but you wrote it has the step-by-step. UC applicants should also read UC's written rule on ChatGPT and the PIQs.
Questions about AI detectors and short essays
How many words does an AI detector need?
No detector maker we could load publishes a number for application answers. The one published warning we found sets the line at 1,000 characters, which is about 170 to 200 words at the rates measured on this page.
Are UC PIQs too short for AI detection?
A full 350-word PIQ is about 1,750 to 2,100 characters, above the 1,000-character line. A shorter PIQ answer of 160 words or less falls at or below it. UC's integrity statement says UC "conducts regular screenings to verify the integrity of the responses"; it does not say how.
Is a supplemental essay AI checker worth running?
Only if you use the highlighted sentences. On a short answer the overall score is the least reliable number on the screen. The rewrite test in the procedure above helps whether or not you run a checker.
Get a free review of your supplemental essay
Get a free Quick Review of your supplemental essay against GradPilot's personal short-answer rubric. It reads whether your answer addresses the prompt you chose, uses material that fits the format, and supports its claims with your own details. For "why this major" answers, use the why-major rubric; UC applicants answering the "what else makes you stand out" question can use the UC PIQ rubric for that question or start from the UC PIQ review page.
Sources
Retrieved October 30, 2026, except where noted.
- ChatGPT's developer, New AI classifier for indicating AI-written text (January 31, 2023, with July 20, 2023 update; archived copy, live page HTTP 403): short-text warning, 9% figure, withdrawal.
- Liang et al., arXiv 2304.02819 (2023; published in Patterns): misclassification of non-native English writing.
- Common App, 2026 first-year application (© 2026): 650-word essay; 250-word challenges question.
- University of California, Personal insight questions: four of eight, 350 words each.
- University of California, Statement of Application Integrity (last updated August 2023): screenings.
- Yale, Essay topics (2026-2027): 200-word answer.
- University of Pennsylvania, Writing (2026-27): 50-200 words.
- Caltech, Supplemental application essays: 100-150 word community answer.
- Stanford, Apply (no year stated): 3 to 50 words per short question.
Quick AI Check
See AI-detection signals and rubric feedback before you submit.