Used AI to Cut Your Essay to the Word Limit, Now It's Flagged?

Because the shortened version is text the chatbot wrote, and a detector scores the text in front of it, not the history behind it. When you ask for 900 words to become 650, you get back a new, shorter text that says roughly what yours said, in sentences the model produced. Turnitin describes its own detector as learning statistical patterns: its classifiers "are trained to detect these differences in word probability" between model writing and human writing. Your original sentences are no longer there to be scored.

What happens next depends on the tool, because the four detector makers covered here describe AI-edited writing in different ways. None of them uses the word "shorten"; Turnitin mentions summarizing once, as a Grammarly feature.

So the same shortened essay can come back as original work with a polish tag from one tool and as likely AI from another. The number on your screen is one tool's reading, not a verdict on how the essay was made.

Turnitin answers the grammar question directly in its AI writing FAQ, and the full sentence ends by naming what it does target: "other AI content written by LLMs such as GPT-3.5." Its statement about summarizing is about Grammarly's generative features specifically. Turnitin does not say the same about every chatbot, but a feature that writes a summary for you is doing what a chatbot does when you ask it to shorten. Two other articles here, on whether Turnitin detects Grammarly and on AI-assisted versus AI-generated writing, read that FAQ answer in full.

GPTZero is the only one of the four that calls AI-polished work original. Its June 2025 post introducing AI Polished describes the tag as covering text written by a human and then polished or rewritten by AI, and says texts with "low edit distance from the original text" were rejected as too close to human writing, "indicating they had not been changed enough to be considered edited." The same post calls AI Polished a "more advanced form of writing enhancement (like the use of Grammarly)." GPTZero made a newer model its default on 20 September 2026, so check the definition panel on your own result screen; we explain that label and panel in a separate walk-through.

Originality.ai now lets the person scanning choose how much AI to tolerate. Its AI Allowance, announced on 21 September 2026, sits alongside its classic model-based scan rather than replacing it, and offers 0%, 5%, 15%, 25% or 40%. For students it suggests "Check assignment or school policies (typically 0% AI or 5% AI)". If someone else ran the scan, it is worth asking which mode and which setting.

Pangram is in the middle of changing its labels. Pangram 3.0 separated light assistance ("surface-level changes that do not affect the underlying ideas, structure, or content of the text") from moderate assistance ("changes where AI may have rewritten significant portions of the text or added content of its own"). Pangram 4 merges them: "AI-Assisted replaces the previous Lightly AI-Assisted and Moderately AI-Assisted labels." Its migration guide says: "Pangram 3.3.2 is still the default and will remain live until September 30, 2026." So which label you see depends on when, and through what, the text was scanned.

A detector reading is one question. Whether you were allowed to use the tool is another, and it is the one a misconduct process asks.

The University of York's policy on acceptable assistance, linked from its current assessment support page, is the plainest example we have found of a rule that names shortening. Among the kinds of help it says are not appropriate for summative work: "Generating or re-writing (including shortening or summarising) any of the student’s sentences or sections of work;" The same policy allows tools that suggest edits for the student to accept or reject, asks for an acknowledgement when proofreading or editing help was used, and exempts the spelling and grammar checking built into Word or Google Docs from that acknowledgement. It applies to York's taught programmes, not to other universities, and which writing tools count as AI assistance sets several other rulebooks side by side.

If the essay is for an application rather than a course, the admissions body's rules are the ones that apply. UCAS, for example, lists Microsoft Copilot summarising a first draft that is over the character count among the ways AI can help, and also asks applicants to declare that the statement was not copied or provided by another source, including AI software. Its list of don'ts has a single entry, "Copy and paste from AI tools and use as final text." and the same page warns: "Anything not in your words could get flagged as fraud by UCAS, and could in turn mean you don't receive any offers or have offers withdrawn."

1. Go back to the version before the shortening. If you drafted in Google Docs or Word online, it is in the history; what Google Docs version history records shows where.
2. Find out what the limit actually counts. References, footnotes, headings and appendices are counted by some courses and not others, and word processors disagree with each other. See whether the reference list counts, why Word, Google Docs and Turnitin give different counts and how much over the limit is tolerated.
3. Delete first, then tighten. Remove whole sentences the argument can stand without, then tighten what remains, yourself or with a tool your rules allow. How to shorten an essay to a word limit goes through the order, and a worked abstract example shows the tightening step.
4. Check anything a tool produced against your original. Quotations should be word for word, every citation should still be attached to its claim, and numbers and qualifiers should be unchanged.
5. Acknowledge the help if your rules ask for it. Requirements like York's apply whatever a detector says.

Do not expect the percentage to fall by a set amount. Turnitin's percentage is taken over the prose it counts as qualifying text, not the whole paper, so what moves it is which sentences are gone and whether they were among the flagged ones; does cutting words change your AI score explains why.

HumanPen rewrites English sentences. By York's wording, that is the kind of assistance York rules out for summative work, and if your course forbids AI rewriting, that settles it.

Where AI-assisted editing is allowed, what we can describe is scope, not outcome. You import the Turnitin or iThenticate report and only the passages it flags are rewritten; you see that scope before anything runs, and every paragraph you did not confirm stays as you wrote it. We do not predict what any detector will say about the result.

Does Grammarly's summarising feature count differently from grammar fixes? For Turnitin, yes: summarizing is one of the generative features whose output it says will likely be flagged, while spelling and grammar corrections are not what it targets. The other three do not mention summarising at all.

What about machine translation? Pangram 3.0 listed translation under light assistance, and Pangram 4 has since merged its grades. Turnitin reads the English you submit and does not separate translated text from generated text. York lists translating whole or significant sections of summative work among the kinds of assistance it rules out.

If I put my original sentences back, will the score go back down? Nobody can promise a number: it depends on which sentences end up in the file and how the tool reads them. What you can be sure of is that every sentence is yours, which is what a misconduct process will ask about.

My essay is for a college application, not a course. Does any of this apply? The detector part does; the rules part is the admissions body's, not a course's. If the tool that flagged you was GPTZero, note what its own post says about AI Polished: that it "indicates the work is original, unlike texts generated entirely with AI."

• Turnitin Guides, Turnitin's AI writing detection capabilities FAQs, read 25 September 2026.
• GPTZero, Behind the Scenes: Introducing "AI Polished" (5 June 2025).
• Originality.ai, Introducing AI Allowance (21 September 2026).
• Pangram, Introducing Pangram 3.0 with AI assistance detection (11 December 2025), the Pangram 4 model card and the Pangram 4 migration guide.
• University of York, Policy on Acceptable Assistance with Assessments ("Last updated: September 2025"), reached from York's assessment support page on 25 September 2026.
• UCAS, A guide to using AI and ChatGPT with your personal statement, read 25 September 2026.