If you are applying this year, the thing to understand is not whether an AI detector can catch you. It is that several institutions have stopped asking that question, and changed what they collect instead.

Detectors are being switched off

Vanderbilt disabled Turnitin's AI detector in August 2023 and published its reasoning. The university had submitted 75,000 papers to Turnitin in 2022, and at the 1 percent false positive rate Turnitin claimed at launch, roughly 750 of those papers could have been wrongly flagged. That 750 is a projection Vanderbilt did to size the risk, not a count of students it accused, and the distinction matters if you see the number quoted elsewhere.

Turnitin's own published figures are worth reading directly. It gives a document-level false positive rate under 1 percent for documents containing 20 percent or more AI writing, and a sentence-level rate of about 4 percent: roughly one highlighted sentence in twenty-five may be human-written.

The failure is not random

A 2023 Stanford study in Patterns tested seven widely used detectors and found something worse than an even scattering of errors. Across seven detectors, more than 61 percent of TOEFL essays by non-native English speakers were classified as AI-generated on average. The essays were human-written, collected by the researchers from a Chinese forum rather than produced for the study. The same detectors were near perfect on essays by US eighth-graders.

The proposed mechanism is the part worth sitting with. The detectors keyed on low perplexity, meaning text that is grammatically careful and stylistically plain, using a narrower range of constructions. That is what writing looks like in a language you learned deliberately rather than absorbed. The tool was identifying a kind of writer and calling it a machine.

The same study found that simple prompting could get AI text past those detectors. So the tool failed hardest on the honest student writing in a second language, and could be evaded by anyone who tried.

What UCAS actually changed, and why

From 2026 entry, the UCAS personal statement is no longer one free-text box. It is three questions: why you want to study the subject, how your qualifications and studies have prepared you for it, and what else you have done to prepare. The overall limit stays at 4,000 characters, with a 350-character minimum per section.

UCAS did not present this as an answer to AI. Its stated reason is that scaffolding levels the playing field for applicants who have less support with essay writing, and its own release frames it around disadvantaged students. Anyone telling you the personal statement was restructured to defeat chatbots is supplying a motive UCAS did not.

The effect is still worth naming. Three specific questions are harder to answer with a generic paragraph than one open box is, because each answer has to attach to your actual qualifications and your actual experience.

Formats that ask for a decision

Look at what the alternatives to detection have in common. Structured prompts tied to your own record. Writing produced in the room. Interviews where someone asks why you chose that example rather than another.

None of these detects anything. What they do is move the assessment to the part that cannot be handed off: the judgment about what to include, what it meant, and why it belongs in an answer to this question. A finished text can be produced by anything. A decision has to be defended by the person who made it.

That is a better question than the one detection was asking, and it was always the better question.

It also cuts the other way for the essay itself. As AI makes application essays converge on each other, covered in AI is flattening the college essay, the parts of an application that are produced under observation carry relatively more of the signal. A structured prompt tied to your own record is one of those parts.

The replacement is not automatically fairer

A school that retires a biased detector and adopts interviews and in-room writing has not removed bias from the process. It has moved it.

Detection penalised students whose written English was careful and plain. Live formats favour students who are comfortable being watched while they think, who have practised speaking about their own work, and who do not freeze when a stranger asks them to justify a choice. Those are different populations, and neither distribution is just. Congratulating yourself for the first change while ignoring the second is how the next unfairness gets built.

What to do about it this cycle

Answer the question that is actually asked. Three structured prompts punish a general statement in a way one open box did not, and the most common way to lose them is to write about the subject rather than about your preparation for it.

Keep the reasons, not just the results. For anything you put down, be able to say why you chose it and what it changed about how you work. That is what a follow-up question asks for, and it is the part no tool can supply for you.

Write plainly, and stop treating plainness as a risk. If a detector misfires on careful, clear prose, that is a fact about the detector. Writing worse to look more human is advice that fails on both counts.

Every source above, and what it does not cover

Vanderbilt's decision and its 75,000-paper calculation are from its own Brightspace post of 16 August 2023. The false positive rates are Turnitin's own published figures. The 61 percent finding is from Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers", Patterns, volume 4, issue 7, 2023. The personal statement change is from UCAS's own pages for 2026 entry. All checked in August 2026, and each of those organisations is the authority on its own policy rather than this page.

We build SAT and ACT practice questions, and one of our modules trains reading what an application prompt is actually asking for, so we have an interest in the argument above. Our own questions are drafted by language models and then independently verified, which is the same tooling this piece is about. We have no data on machine-written application essays and have made no claim about them here.