A student's desk by a window with handwritten draft pages, a marked-up printout and a closed laptop, in soft morning light
AI-generated image, not a photograph of a real person or place.
aimodelrate

WHAT THE RESEARCH FOUND

A score is
not proof.

ERROR RATE read it before you trust it

  • A guess, not a match
  • False flags happen
  • Look at the process
  • Evidence first, score second

AI detectors guess; they do not check.

Tools that claim to spot AI-written text look at how predictable the writing is. Studies find they miss edited AI text and wrongly flag real human writing, most often by people writing in a second language.

Updated

Check a case

THE SHORT ANSWER

Detectors make
two kinds of mistake.

They can miss AI text, and they can flag human text.
Only one of those mistakes lands on an innocent writer.

HOW THEY WORK

A statistical guess

A plagiarism checker matches text against text that exists elsewhere. An AI detector instead guesses from patterns, so there is nothing to check its answer against.

Check a case

FALSE FLAGS

Real writing, wrongly flagged

Seven detectors flagged an average of 61% of genuine TOEFL essays as AI-written, while scoring US pupils' essays almost perfectly.

See the numbers

WHAT WORKS

Look at the process

Universities that have stepped back from detectors point to drafts, conversation and assessment design instead of a single score.

What to do instead

A QUICK CASE CHECK

How far can you
trust the score?

Pick a situation to see what the research found about detector errors. The bar shows how much doubt the studies cast on a detector's verdict in that case.

Case check

Choose a situation
School essay, first languageMostly right

In one 2023 study, seven detectors were near-perfect at recognising US 8th-grade essays as human-written. Even so, a score on its own is never proof.

Essay in a second languageOften wrong

The same detectors wrongly flagged an average of 61.22% of 91 genuine TOEFL essays, and 89 of the 91 were flagged by at least one detector.

AI text, then rewordedOften missed

When AI-written college essays were put through one more rewording prompt, detection fell from 100% to 13%. Light edits can push a score to zero.

A whole year of papersSmall rate, many people

A 1% false positive rate sounds small. Vanderbilt worked out that on its 75,000 papers a year it could mean around 750 wrongly labelled.

Sources: Liang et al., GPT detectors are biased against non-native English writers (2023); Vanderbilt University guidance on AI detection (2023).

Figures come from published studies of specific tools at a specific time. Results vary with the tool, the text and the settings.

TWO NUMBERS WORTH KNOWING

The errors are
not random.

One figure on who gets wrongly flagged,
one on what a small rate means at scale.

≈ 61%

Genuine TOEFL essays flagged as AI

Seven widely used detectors misclassified an average of 61.22% of 91 human-written TOEFL essays as AI-generated. The authors link this to the narrower word choice of people writing in a second language.

FALSE FLAGS · LIANG ET AL. 2023

≈ 750 papers

Papers a 1% rate could mislabel

Vanderbilt submitted 75,000 papers in 2022. At the 1% false positive rate claimed at launch, around 750 could have been wrongly labelled, so it switched the detector off.

AT SCALE · VANDERBILT

A 2023 test of 12 public tools and two commercial systems by researchers across Europe and Mexico concluded that the detectors were neither accurate nor reliable, and leaned towards calling text human-written.

FOUR STEPS THAT HELP

Make the work
speak for itself.

A tutor and a student, seen from behind at a distance, talking through a printed essay at a table in a quiet library
AI-generated image, not a photograph of a real person or place.

None of these needs a detector. They are the ordinary signs of real work that university guidance keeps coming back to.

  1. Keep your drafts

    Notes, early drafts and version history show how a piece of work grew. They are evidence a score cannot give.

  2. Know the rules for each task

    Ask what AI use is allowed, and say how you used it when use is permitted, citing it as the course asks.

  3. Be ready to talk it through

    Being able to explain your argument and choices is strong evidence of your own thinking.

  4. Treat a score as a question

    For staff, a high score is a reason to look more closely and talk to the student, not a verdict.

WHY IT IS HARD

Good writing can
look like AI.

Two hands holding a printed essay covered in pencil notes and crossings-out, with earlier drafts stacked beside it
AI-generated image, not a photograph of a real person or place.

Many detectors rate how predictable the wording is. Plain, careful, conventional phrasing reads as more predictable, and that is often how people write in a second language or while still learning academic style.

Oxford's AI Competency Centre says the University does not currently endorse any AI detector for academic decisions, and argues the better question is not whether AI was used but whether the student learned.

Questions

Questions people ask

Can AI writing detectors prove someone used AI?

No. They make a statistical guess from patterns in the text. Oxford's AI Competency Centre notes that, unlike a plagiarism match, there is no way to check a detector's answer on real student work.

Why are non-native English writers flagged more?

Liang and colleagues found the detectors' errors were linked to lower perplexity, meaning more predictable wording. Writers using a narrower range of words and structures look more 'AI-like' to the tools.

Can AI-written text slip through?

Yes. In the same study, one extra rewording prompt cut detection of AI-written college essays from 100% to 13%. A 2023 study by researchers in Europe and Mexico also found rewording and obfuscation made the tools perform much worse.

How many students use AI for assessed work?

HEPI's Student Generative AI Survey 2026 of 1,054 UK full-time undergraduates found 95% use AI in some form and 94% use generative AI to support assessed work.

What should I do if my work is flagged?

Ask what other evidence is being used, and share your drafts, notes and version history. Follow your school or university's academic misconduct procedure and ask for the chance to talk through your work.

Sources

Where the facts come from

  1. Liang et al. (2023): GPT detectors are biased against non-native English writers (arXiv)
  2. Weber-Wulff et al. (2023): Testing of detection tools for AI-generated text (arXiv)
  3. University of Oxford AI Competency Centre: The question of AI detection
  4. Vanderbilt University: Guidance on AI detection and why we are disabling an AI detector
  5. HEPI: Student Generative Artificial Intelligence Survey 2026
  6. HEPI blog: AI detectors and the fairness gap

ABOUT THIS NAME

Interested in leasing aimodelrate.com?

aimodelrate.com is part of the DoodleSale portfolio. It isn't in an auction right now, but you can ask about leasing it.

Ask about this name