AI Detection in Education: Use It Carefully
This page argues for restraint about our own product. Detection is useful in education, and it has already caused real harm there — mostly because scores got treated as evidence. The difference is entirely in process.
What happened with text detection
Worth knowing before repeating it with images. AI-text detectors were adopted quickly by institutions and produced documented false accusations against students, with non-native English speakers disproportionately affected because their writing patterns resembled what the tools flagged.
Several major universities subsequently disabled these tools entirely. The lesson was not that detection is worthless — it was that a probability score used as proof produces injustice at scale.
Image detection carries the same structural risk, and the same fix.
Where image detection is genuinely reliable
Photographic submissions. Unedited, full-resolution photos are our strongest case.
Files that still carry metadata. Some generators sign their output; local Stable Diffusion often stores the prompt outright. Check that first — it is free and sometimes ends the matter without any score at all.
Portraits. Face generators leave systematic traces, as covered on our headshot page.
Where it is not
Digital illustration and design work. Clean vector and flat-colour art shares statistical properties with generated images. A student who works cleanly is at higher risk of a false flag than a sloppy one, which should tell you something about the metric.
Anything submitted through a platform. LMS uploads re-compress files and strip metadata, weakening the analysis before you even start.
Screenshots and phone photos of screens. Effectively unanalysable.
A process that is actually fair
1. Never open with the score. Treat it as a reason to ask questions, not as a finding to present.
2. Ask for process evidence. Layered files, version history, sketches, timelapse. This is the strongest evidence of authorship that exists, and it is available to any student who did the work.
3. Talk to them about the work. Someone who made an image can explain the decisions in it. This single conversation resolves most cases.
4. Document what you relied on. If a sanction follows, the reasoning should survive appeal — and “the tool said 87%” does not.
5. Set the policy in advance. Whether AI assistance is allowed, and how it must be disclosed, belongs in the brief rather than in a dispute afterwards.
What to tell students
Two things, ideally before the assignment rather than after.
First, keep your working files. Not because you are suspected, but because process evidence protects you if you ever are. This is now basic professional practice, not paranoia.
Second, understand the disclosure rules for the specific brief. Most conflicts come from undefined expectations rather than deliberate cheating.
If you want to see how the analysis reasons, our technical explainer covers what is measured and where it breaks down.
Key points
- AI-text detection produced documented false accusations — do not repeat the pattern.
- Detection is reliable on unedited photos and portraits, weak on clean digital art.
- Platform uploads strip metadata and compress, weakening analysis before you start.
- A score is a reason to ask questions, never a finding to present.
- Process evidence — layered files, version history — settles authorship properly.
- Set disclosure policy in the brief, not in the dispute.
Frequently asked questions
Check an image yourself
Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.
More from the blog
Try it on your own image
Free account, 3 checks included, no card required.