AI images and academic integrity | Best AI Image Detector

A score cannot support a misconduct finding. Detection has a published error rate, no chain of custody and no examiner, and a student cannot cross-examine a model.

Why a score cannot decide a hearing

Academic misconduct procedures are quasi-legal. They require evidence a student can examine and challenge, applied consistently, with a route to appeal. A detector output fails on every count.

There is a known error rate. Around one image in eleven is called wrong under laboratory conditions, and worse on submitted work that has been exported, compressed and uploaded through a portal.
There is no chain of custody. The file was processed on somebody's laptop with no record of what was run or which model version produced the number.
The reasoning is not examinable. A student cannot question a model about why it produced a figure, and the institution usually cannot explain it either.
The errors are not evenly distributed. Students who use assistive editing tools, or who work on lower-specification equipment that compresses exports, are flagged more often.

That last point is the one institutions underestimate. If a detector systematically flags a group of students more often, using it in decisions creates an equality problem on top of a fairness one.

What to ask for instead

The question is not whether an image looks generated. It is whether the student did the work. Process evidence answers that directly, and it is much harder to fabricate after the fact than a finished file.

Require working files with the submission
Layered documents, project files, raw captures or version history. Ask for them at submission rather than during an investigation, so nobody has to assemble them under suspicion.
Ask for intermediate stages
Three progress exports with timestamps demonstrate development. This is the single most useful requirement an institution can add, and it costs students almost nothing.
Hold a short viva on the work
Five minutes asking why a decision was made settles most cases. A student who made the image can explain the choices. One who did not, usually cannot.
Use a detector only to decide where to look
A flag is a reason to ask for the working files. It is not a finding, and it should never appear in an allegation letter.

The false positive problem in coursework

Student submissions are close to a worst case for detection. Work is exported for a portal, compressed on upload, and often produced with upscaling, denoising or generative fill used legitimately as part of a taught workflow.

One image, four stages of the same submission

Original export

Compressed for upload

Portal re-encode

Marker screenshot

Illustrative figures for a genuine student photograph exported and uploaded normally.

How much an identical piece of work can move on its way through a submission system.

Nothing was done to that image except submitting it. A marker checking a screenshot from a grading interface is checking the most degraded copy in the chain, and getting the least reliable answer available.

If you are a student who has been accused

Being asked to prove you made your own work is difficult and increasingly common. These are the things that carry weight with a panel.

The working file, with layers or history intact. This is the strongest evidence available and it is difficult to fake convincingly.
Dated intermediate versions, including the ones you were unhappy with.
The original capture, straight from the camera or phone, rather than the export you submitted.
A record of the tools you used, including any assistive features, stated openly rather than discovered later.
A request for the specific evidence against you. An institution should be able to say what it relies on beyond a number.

Writing a policy that holds up

The institutions handling this best changed their submission requirements rather than buying a detector. That approach costs less, produces fewer disputes, and survives contact with an appeals panel.

Require working files at submission for any assessment where authorship matters. Students who know the requirement plan for it, and the material arrives as a matter of course rather than being demanded from somebody already under suspicion.

Say which tools are permitted and require a short declaration of what was used. Most disputes involve a student who used a tool the course allowed and did not think to mention it, which a declaration field removes entirely.

State plainly that automated detection does not on its own constitute evidence of misconduct. Putting that in the policy protects the institution as much as the student, because it stops a panel being asked to rely on something it cannot properly examine.

Questions people ask

Can a university use an AI detector on submitted images?

As a triage signal, with care. As evidence in a misconduct case, no. The error rate is published, the reasoning is not examinable and a student cannot challenge a model. Any finding needs evidence about the process by which the work was made, not about the finished file.

What if a student cannot produce working files?

That is a question about your submission requirements rather than about the student. If working files were not required at submission, asking for them afterwards is unfair and often impossible to satisfy. Change the requirement for the next cohort rather than penalising this one.

Are some students flagged more than others?

Yes, and this is the risk institutions miss. Students using assistive editing tools, working on lower-specification equipment, or submitting through systems that recompress heavily are all more likely to be flagged, and none of that correlates with misconduct.

Should image AI be banned in coursework?

That depends on what the course teaches, and a blanket ban is usually unenforceable. Clear rules about which tools are permitted, plus a declaration requirement, produce far fewer disputes than a prohibition nobody can verify.

What should an allegation letter say?

What evidence is relied upon, in terms the student can respond to. A letter that says an automated check indicated possible AI generation is not a charge somebody can answer, and a panel is likely to say so.

Can students check their own work before submitting?

Yes, and it is worth doing on the exact file being submitted. A high score on genuine work is a prompt to keep the working files and note the tools used, rather than a reason to change anything about the piece.