Why Corra will never tell you a student used AI
By The Corra team · · 2 min read · Assessment
Corra will not include an AI-writing detector, and this is a permanent product decision rather than a gap in the roadmap. Detectors return a probability dressed up as a verdict, that verdict is wrong often enough to damage real students, and the students it is most likely to be wrong about are the ones with the least ability to argue back.
The accuracy problem is structural, not temporary
Detectors work by measuring how predictable a text is, and predictable prose is not evidence of a machine. Careful, formulaic, well-drilled writing scores as machine-like. So does writing by someone using a limited vocabulary with great precision. A better detector does not fix this, because the signal it measures is not the thing anyone actually wants to know.
Who gets hurt
Researchers at Stanford, publishing in the journal Patterns in 2023, tested several GPT detectors on essays by non-native English writers and found they were flagged as AI-generated at far higher rates than essays by native speakers. If you teach multilingual learners, a detector is not a neutral tool in your classroom; it is a tool with a known bias against your students.
- Multilingual writers are disproportionately flagged.
- Students with a plain, careful writing style score as machine-like.
- The accusation lands on the student, while the uncertainty stays with the software.
What a false positive actually costs
An integrity accusation is not a low-stakes event. It is a meeting, a mark on a record, a conversation with a family, and for a student who did the work honestly it is a lesson that effort is not what gets recognised. Software that produces that outcome with an unquantified error rate is not a feature we are prepared to ship, whatever the demand for it.
What Corra does instead
The useful response to AI in student work is assessment design, not surveillance. Corra’s assessment module is built to produce tasks that are harder to outsource to a chatbot, because they are anchored to things a model does not have access to.
- Tasks anchored to a specific lesson, discussion or local context.
- Process evidence — drafts, revisions, a short oral defence — weighted alongside the product.
- Prompts requiring a student to argue against their own earlier position.
- Variants generated per class so an answer copied from elsewhere does not fit the question asked.
The honest limit of this position
None of this prevents a determined student from using a chatbot, and we are not claiming it does. It shifts the effort so that the shortcut stops being cheaper than doing the work, and it keeps the teacher from having to act on a number that cannot support the weight being put on it.
Common questions
- Will you add a detector if they get more accurate?
- No. The problem is what the measurement is, not how well it is measured. A more confident wrong answer is worse, not better.
- What do I do if I genuinely suspect AI use?
- Talk to the student about their process and ask them to walk through their draft. A conversation produces evidence a probability score never can, and it treats the student as a person rather than a flag.
- Is there a source for the bias claim?
- Yes — Liang, Yuksekgonul, Mao, Wu and Zou, “GPT detectors are biased against non-native English writers”, published in Patterns in 2023. It is worth reading in full rather than taking from a product page.