Why You Should Never Screenshot a Detection Score and Email It to a Student as Proof

A red number on a screen is persuasive. “98% AI” looks like a verdict, and the impulse to grab a screenshot, attach it to an email, and send it to the student feels like efficient, evidence-based teaching. It isn’t. That one habit can turn a fixable misunderstanding into a formal dispute you’re likely to lose.

This is about a specific moment in an educator’s day: the moment you decide how to act on a detection score. Send the screenshot, and you’ve documented a conclusion before you’ve had a conversation. Here’s why that backfires, and what works better.

Key takeaways

• A detection score is a probability estimate, not a finding of fact — a screenshot of one is not evidence.
• Emailing the score creates a written record of a flawed process before the student has said a word.
• Detectors produce false positives and, according to research, disproportionately flag non-native English writers, plain prose, and short passages.
• The same passage scores differently across tools and even between runs, so one number proves nothing alone.
• Lead with a conversation and process evidence — drafts, version history, the student explaining the work — not a chart.

What a detection score actually is

A detection score is a model’s estimate of how likely a passage is to have been machine-generated, expressed as a percentage or a label. It is not a measurement of cheating. It’s a statistical guess about word patterns.

Most detectors look at how predictable the text is. AI models tend to pick high-probability words and keep sentence rhythm even, while human writing wanders more. That’s the core idea behind the statistics behind perplexity and burstiness. The trouble is that plenty of human writing is also smooth and predictable — a careful student editing for clarity, a non-native speaker leaning on safe sentence structures, a tightly formatted lab report. The model can’t tell “wrote it carefully” from “didn’t write it at all.”

So “98% AI” doesn’t mean “98% sure this student cheated.” It means the text shares statistical features the model associates with AI output. Those are very different claims, and the gap between them is where students get hurt.

Why emailing the score is the part that bites you

Even if you suspect something, putting the score in an email is the move that creates real exposure. Three reasons.

It documents a conclusion before any process

An email saying “the detector says this is 98% AI, please explain” reads as an accusation that’s already been decided. If the score later turns out to be a false positive — and they happen — that message becomes part of the record. In an appeal or a meeting with an administrator, it shows the case started from a number, not from the work.

It anchors everyone to the wrong thing

Once a percentage is in writing, the whole conversation orbits it. The student argues with the number. You end up defending the tool. Meanwhile nobody is talking about the actual assignment: the drafts, the sources, the thinking. A score is the weakest part of any integrity case, and leading with it teaches everyone to fight over the weakest evidence.

A screenshot has no context and no audit trail

A screenshot strips away which tool ran, which version, what settings, how long the passage was, and when. Short passages are notoriously unreliable — too few words for the statistics to mean much. Run the same paragraph twice and you can get different results. A static image of a number can’t show any of that, yet it looks authoritative. That mismatch — looks certain, is fragile — is exactly what gets institutions into trouble.

A mini-scenario: the email that became the problem

Here’s a composite of the kind of situation that plays out on campuses. A composition instructor runs a detector on a discussion post from Mateo, an international student. It comes back 89% AI. The instructor screenshots it and emails: “This was flagged as 89% AI-generated. Per the syllabus, this is a zero. Let me know if you’d like to discuss.”

Mateo wrote every word. English is his third language, so he writes in short, clear, careful sentences — exactly the pattern detectors over-flag. He has the drafts. He has the browser history showing his research. He forwards the email to an advisor.

Now the instructor is defending a screenshot. The advisor asks reasonable questions: Which tool? What’s its false-positive rate? Was the 150-word post long enough to score reliably? Why a zero before a conversation? The instructor has no good answers, because the screenshot never contained them. The case collapses, and the relationship with the student is damaged over something that was never cheating.

Compare that to the email the instructor wishes they’d sent: “Hi Mateo, I’d like to talk through your discussion post and your process — can you stop by office hours or send me your drafts?” Same concern, no premature verdict, no screenshot to defend.

What the evidence says about detector reliability

This isn’t a fringe worry. A few grounding facts worth keeping in mind:

• OpenAI retired its own AI text classifier in July 2023, citing a low rate of accuracy. The company that builds the models couldn’t reliably detect them.
• A 2023 Stanford study (Liang et al., published in the journal *Patterns*) found that several popular detectors misclassified more than half of non-native English writing samples — TOEFL essays written by humans — as AI-generated, while staying nearly perfect on essays by native speakers.
• Vendors themselves generally state in their own documentation that scores are probabilistic and shouldn’t be the sole basis for an accusation. The people who make the tool are telling you not to use it the way a screenshot-email uses it.

None of this means detectors are useless. A score can be a reasonable prompt to look more closely. It just can’t be the thing you hand a student as proof.

Why two tools disagree on the same paragraph

If you’ve ever run a passage through two detectors and gotten 90% from one and 20% from another, that’s normal. Each tool is trained on different data and tuned to a different decision threshold. They weigh signals like word predictability and sentence variety differently, so they land in different places on the same text.

That disagreement is the whole argument against the screenshot. If Tool A and Tool B can’t agree on one paragraph, a single image of Tool A’s number isn’t a fact about the student. It’s a snapshot of one model’s opinion at one moment. Run a few samples yourself and watch the scores move — seeing the uncertainty firsthand tends to change how educators treat that number.

What to do instead

Replace the screenshot reflex with a short, repeatable process:

1. Treat the score as a prompt, not a proof. A high score means “look closer,” not “send the email.” Keep the number to yourself for now.
2. Gather process evidence. Ask for drafts, outlines, and version history. In Google Docs or Word, version history shows how the document grew. Steady, organic editing looks very different from a single paste.
3. Have the conversation first. Meet privately. Ask the student to walk you through how they wrote it — sources, choices, what a tricky paragraph means. Most genuine authors can do this easily, and it’s a lot more humane than an accusatory inbox.
4. Compare to known work. Does this match their past writing and their in-class voice? That comparison is more reliable than any single percentage.
5. Follow your institution’s process. If concern remains, use the formal academic-integrity channel, which is built for fairness and documentation. Don’t improvise a verdict over email.

None of these steps need a screenshot, and all of them produce evidence that actually holds up.

Frequently asked questions

Can I use an AI detection score as proof in an academic integrity case? No. Most detectors return a probability or likelihood, not a determination, and their vendors generally say the output shouldn’t be the sole basis for an accusation. Treat a score as one weak signal that might prompt a conversation. Process evidence — draft history, timestamps, the student explaining their reasoning — carries far more weight in a fair review.

What’s wrong with just emailing the student the screenshot so they can respond? It frames the conversation around a number instead of the work, and it puts a conclusion in writing before you’ve heard from the student. If the score is a false positive, that email becomes part of a record showing the process started from a flawed premise. A neutral request to meet and discuss the assignment is fairer and harder to misread.

Aren’t AI detectors accurate enough to trust now? They’ve improved, but they still make mistakes in both directions, and accuracy drops on short, edited, or newer-model text. The 2023 Stanford study found several detectors flagged non-native English writers far more often than native speakers, and OpenAI retired its own classifier in 2023 over a low rate of accuracy. No detector outputs certainty.

Why do two detectors give different scores for the same essay? Each is trained on different data and tuned to a different threshold, so they weigh signals differently. One tool saying 90% and another saying 15% on the same paragraph is common — and it’s exactly why a single screenshot can’t stand in for a careful review.

What should I do if a student’s work genuinely concerns me? Document your specific observations, request drafts and version history, and have a calm private conversation where the student walks you through their work. Compare it to their known writing, and follow your institution’s integrity process. The goal is to understand what happened, not to win an argument with a chart.

The bottom line

A detection score can start a conversation. It can’t end one, and it definitely shouldn’t be the first thing in a student’s inbox. Lead with the work, the drafts, and a real talk — and keep the screenshot out of the email entirely.

If you want a feel for what these tools measure before you ever act on a number, try a passage yourself and watch how the score shifts as the text changes. For more on AI writing and detection, the blog goes deeper, and if you’re weighing it for a team, here’s how PaperBleach is priced.