What replaced the AI detectors

Between August 2023 and January 2026, the universities named below stopped using software that claims to detect AI writing. This is written for the families who spent the same period worried that a detector would wrongly flag an application essay.

We build an admissions advisory product, so a calmer and better informed parent is good for us. The evidence below is public and linked, and the last two sections are the parts that are bad news for some students.

Seven universities named so far, and one that kept half the tool

Inside Higher Ed reported on 5 August 2026 that a growing number of universities have pulled back from AI detection, naming Yale, Vanderbilt, Johns Hopkins, Indiana, Northwestern, Georgetown and New York University. Treat that as a list of places worth checking rather than a set of matched policies. Checking them individually, the picture is less uniform than the list looks: some of it is university-wide policy, some is a task-force recommendation, and some sits at the level of a single school. If one of these is on your child's list, read that university's own page rather than the roundup. Indiana's Kelley School of Business tells instructors to design assignments around process and reasoning instead.

Curtin University in Australia switched its AI writing detection off across all campuses from 1 January 2026. One detail in Curtin's notice is worth keeping: only the AI detection feature went. The ordinary text matching that catches copied passages stayed on. Plagiarism checking did not go anywhere.

What the Vanderbilt 750 figure actually was

Vanderbilt's number circulates as though 750 students were wrongly accused by a faulty detector. That is not what happened.

Vanderbilt disabled Turnitin's AI detector in August 2023. It gave several reasons: the tool's reliability, the absence of any explanation of how it reached a verdict, and its tendency to flag writing by non-native English speakers. One of the reasons was arithmetic. Turnitin claimed a false positive rate of about 1%. Vanderbilt students submitted roughly 75,000 papers in 2022. Had the detector been running, about 750 of those 75,000 could have been wrongly labelled as partly machine written. The 750 was a projection from the vendor's own accuracy claim, not a count of anything that occurred.

It is a good argument. It is a weaker argument when it is repeated as a body count.

The false positive problem is not evenly distributed

The most-quoted evidence on who gets wrongly flagged is a 2023 study in Patterns by Liang and colleagues at Stanford. They ran seven detectors over 91 TOEFL essays written by non-native English speakers and 88 essays written by US eighth graders. Both samples are small and the finding deserves to be read that way.

The detectors handled the 88 US essays well. On the 91 TOEFL essays they returned an average false positive rate of 61.3%. All seven detectors agreed that 19.8% of those human written essays were machine written, and at least one detector flagged 97.8% of them.

Two caveats usually get dropped when this study is quoted. It covers GPT detectors generally rather than Turnitin specifically, and 91 essays from one forum is a small sample. Later work is mixed, and at least one study using its own carefully built detectors reported no such bias. The exact figure should not be passed around as a measured property of any particular product.

UCAS replaced the personal statement with three questions

For 2026 entry onwards, the UCAS personal statement is no longer a single free-text box. It is still a personal statement and still 4,000 characters. It is now answered as three questions:

• Why do you want to study this course or subject?
• How have your qualifications and studies helped you to prepare for this course or subject?
• What else have you done to prepare outside of formal education, and why are these experiences useful?

The limit is still 4,000 characters in total, with a minimum of 350 characters per answer and the split left to the applicant.

That is a small change to what a student has to say and a large change to what they have to be able to do. A blank box rewards the student who can build a narrative, or who knows an adult with time to help them build one. Three labelled questions hand every applicant the structure that used to be the private advantage. What is left to supply is specific evidence, and evidence is harder to borrow than shape.

The advantage moves, it does not disappear

The Inside Higher Ed piece lists what some institutions are using instead: in-person oral exams, blue book exams written under supervision in class, and assignments that ask a student to show their process rather than only the finished product. Those formats are harder to hand to a language model. They also reward a particular kind of student: one who is comfortable being watched while they think.

That is not the same population as the students who had access to a good editor, and it is not obviously a fairer one. A quiet student, an anxious student, a student who is fluent on paper and halting out loud, each loses ground under oral assessment for reasons that have nothing to do with how well they understand the subject. Swapping an advantage held by families who could buy editing for an advantage held by families who could buy confidence is a trade, not a correction.

What a parent can stop doing this autumn

Stop paying for anything that promises to make writing undetectable. Several of the institutions whose detectors worried families have now switched them off.

Stop treating a detector score as proof on its own. That is the distinction several universities have written into policy: a score may start a conversation, and it is not by itself a finding.

The preparation that helps is unglamorous. Ask your child to talk about one of their projects for five minutes without notes. If they can explain what they chose, what failed, and what they would do differently, they are ready for the kinds of assessment described above. If they cannot, no amount of editing on the written draft closes that gap.