The Burden of Proof in AI Cheating Cases: Why a Score Alone Isn’t Enough

A detector flashes “98% AI,” and for a second it feels like the case is closed. It isn’t. A score is a starting point, and if you treat it as the finish line, you can end up accusing a student you can’t actually convict, on evidence that wouldn’t survive five minutes of scrutiny in an appeal.

This is a piece about the part of an AI cheating case that gets skipped in the heat of the moment: who has to prove what, and what it actually takes to prove it.

Key takeaways

• In a misconduct case, the educator or institution carries the burden of proof — a detector percentage is one data point, not the finding itself.
• Detectors return probabilities, not facts; OpenAI retired its own AI text classifier in July 2023 because it wasn’t accurate enough.
• Most panels use a “preponderance of the evidence” standard, and a bare score rarely clears that bar alone.
• A defensible case corroborates the score with drafts, version history, a conversation with the student, and assignment-specific knowledge.
• Treating a number as a verdict invites false accusations, appeals, and bias problems, especially against ESL writers.

What “burden of proof” means in your classroom

Strip away the legal-sounding language and it comes down to one question: who has to convince whom? In nearly every academic integrity system in the country, the answer is that the accuser carries the burden. If you bring the charge, you have to support it. The student doesn’t walk in owing you proof of innocence. You walk in owing the panel a case.

That framing matters because the detector quietly flips it in people’s heads. A high score *feels* like it shifts the burden onto the student — the tool says you cheated, so explain yourself. But a probability score isn’t an accusation that’s already been proven. It’s the reason you’re looking, not the looking itself.

Most institutions ask you to meet a preponderance of the evidence standard. That’s the “more likely than not” bar, often described as just past 50 percent confidence. A few schools use the stricter “clear and convincing” standard. Almost none use the criminal “beyond a reasonable doubt” bar. Either way, the honest question is whether your *total* case clears the line, not whether a single number looks alarming. Wording varies by school, so check your own policy before you lean on any of this.

Why a score can’t carry the case by itself

Here’s the uncomfortable engineering reality. An AI detector doesn’t read for meaning or check authorship. It measures how *statistically predictable* a passage is and compares that to patterns it has seen in machine-generated text. When it returns “98% AI,” it’s reporting that the writing looks like the kind of text its model associates with AI — not that it caught anyone in the act. If you want the mechanics, detectors report a probability, not proof, and that distinction is the whole ballgame.

Three facts should keep any percentage in its place.

The vendors themselves admit the limits. OpenAI launched an AI text classifier and then pulled it in July 2023, citing a low rate of accuracy. The company that builds the models couldn’t reliably detect the models’ output.

False positives are documented, not theoretical. Clean, human-written work gets flagged. A polished, formulaic essay from a careful student can read as “predictable” to a detector for the same reasons good editing makes prose smoother.

The error isn’t evenly distributed. A 2023 Stanford study found detectors disproportionately flagged writing by non-native English speakers. An ESL student who writes in plainer, more regular sentences can trip the same statistical wires as a chatbot.

Put those together and a lone score looks less like evidence and more like a smoke alarm. A smoke alarm is useful — it tells you to go check the kitchen. Nobody calls the fire report complete because the alarm beeped.

A mini-scenario: same score, two outcomes

Two students in the same section both come back at 92% AI on the same assignment.

Student A can’t talk about her own paper. Her draft history in the LMS shows the entire essay pasted in as a single 1,400-word block at 11:52 p.m., with no earlier versions. The argument cites a framework you never taught and that doesn’t appear in any assigned reading. Asked to explain her central claim, she restates the prompt.

Student B has six saved revisions stretching over four days, with comments in the margins and a messy outline that became the final structure. He walks you through his argument, disagrees with a point in the reading, and tells you why he cut a paragraph. English is his second language, and his sentences tend to be even and clean.

Same number. Completely different cases. For Student A, the score lines up with a stack of corroborating evidence that points at authorship. For Student B, the score is almost certainly a false positive, and pressing it would be a mistake an appeals committee, a parent, or a journalist would take apart. The percentage didn’t decide either case. The *other* evidence did.

What actually builds a defensible case

If the score is the alarm, this is the investigation. Strong cases lean on evidence that answers the real question — did this student do this work — rather than on a probability about style.

Process evidence

Drafts, version history, timestamps, research notes, and outlines show the work taking shape over time, or reveal that it didn’t. Document version history is especially hard to fake after the fact. If your assignment platform keeps revision data, that record is worth more than any detector reading.

A conversation

Sit down with the student and ask them to discuss their own argument. Someone who genuinely wrote a paper can usually explain why they made a choice, what they cut, and where they got stuck. This isn’t an interrogation. It’s the most direct test of authorship you have, and it worked long before AI existed.

Assignment-specific knowledge

You know what you taught. Content that contradicts your lectures, invents sources, or solves a problem in a way the class never covered is a concrete, explainable signal. “This cites a 2019 study that doesn’t exist” is evidence. “The software says it’s predictable” is not.

A second human read

Pull up the student’s earlier work. A sudden, unexplained leap in style, vocabulary, or argument quality is a real signal, and one you can put into plain words for a panel, which a black-box percentage never is.

Notice what these share: a person could stand in front of a hearing and *explain* them. That’s the test. If your only answer to “what’s your evidence?” is “the tool said so,” you don’t have a case yet. You have a reason to keep looking.

How to keep the burden where it belongs

A few habits keep the score in its proper role.

Treat detection as a screen, not a verdict. A flag triggers a closer human look. Nothing more is decided by the number alone.

Write it into your policy. State clearly that a detector score won’t be used as sole evidence, and tell students what *will* be considered. Predictability protects everyone.

Design assignments that generate process. Required outlines, in-class drafting, annotated bibliographies, and short oral check-ins create authorship evidence as a byproduct. They also make cheating harder and accusations rarer.

Document your reasoning. When you do bring a case, write down the corroborating evidence, not just the percentage. That record is what survives an appeal.

If you want to see how a detector reads student writing before you ever weigh one in a case, run a passage through PaperBleach yourself and watch how the same text scores differently with small changes. For programs weighing how these tools fit into day-to-day grading, it helps to see how the checker fits into a classroom workflow as one input among several.

Frequently asked questions

Is an AI detector score enough to prove a student cheated?

No, not on its own. A detector outputs a probability that text resembles AI patterns, not a finding of fact about who wrote it. The tools have documented false positives, and OpenAI discontinued its own AI text classifier in July 2023 because it wasn’t accurate enough. A fair process treats the score as a prompt to look closer, then corroborates it before reaching any conclusion.

Who has the burden of proof in an academic AI cheating case?

In the overwhelming majority of integrity systems, the institution carries it. As the instructor or accuser, you have to show a violation more likely than not occurred. The student isn’t required to prove innocence, though in practice their process evidence often becomes the deciding factor. Check your specific institution’s policy, since wording varies.

What standard of proof do most integrity panels use?

Most use “preponderance of the evidence” — more likely than not, often described as just over 50 percent confidence. Some schools apply a higher “clear and convincing” standard. Almost none use the criminal “beyond a reasonable doubt” bar. Even under the lower standard, a single detector percentage rarely carries enough weight by itself.

What evidence actually strengthens a case?

Corroborating evidence that speaks to authorship: a sudden, unexplained style shift, version history showing text pasted in whole, an inability to discuss the submitted work, or content that contradicts what was taught. These point at the real question in a way no percentage can.

What’s the risk of relying on the score alone?

False accusations, overturned cases on appeal, and harm to a student you may have gotten wrong. The 2023 Stanford finding that detectors disproportionately flag non-native English writers also means a score-only approach carries a bias problem that can put your institution on the wrong side of a fairness complaint.

The bottom line

A detector score is the moment you start asking questions, not the moment you stop. The burden of proof in an AI cheating case sits with you, and percentages don’t carry it — drafts, conversations, and your own knowledge of the assignment do. Build the case on what you can explain to a panel, and you’ll be right more often and fair to everyone. For more on the topic, browse our other guides on AI detection and integrity.