The Burden of Proof Problem: Who Has to Prove AI Use in an Honor Hearing

Getting an email that starts with “we need to discuss the integrity of your recent submission” is the kind of thing that empties your stomach in one sentence. Most students assume the worst right away: that they now have to prove they’re innocent. Here’s the part nobody explains in the panic of that moment. In almost every academic integrity process, that’s backwards. The school has to prove you did something wrong, not the other way around.

That single distinction, who carries the burden of proof, shapes how an entire hearing plays out. So let’s unpack it plainly.

Key takeaways

• In nearly every campus integrity system, the institution carries the burden of proof, not the accused student.
• Most hearings run on a “preponderance of the evidence” standard: more likely than not, not “beyond a reasonable doubt.”
• A detector score is circumstantial evidence, not proof. Detectors output probabilities, and they get things wrong.
• The burden can quietly shift onto you in practice when a policy is vague or a panel treats a percentage as a verdict.
• Process evidence (drafts, version history, notes) is your strongest counterweight because it answers the real question.

What “burden of proof” actually means

Burden of proof is just a fancy way of asking: who has to convince whom? In a fair process, the person making the accusation has to back it up. You don’t walk in owing the panel a defense. They walk in owing you a case.

This matters because it sets the default. If the evidence is a wash, if it could go either way, the side carrying the burden loses. So when an integrity office says “we believe AI was used,” the unspoken second half of that sentence should be “and we can show it.” If they can’t, the tie goes to you.

There’s a related idea worth knowing: the standard of proof. The burden says who must prove it; the standard says how convincingly. Those are two different dials, and people mix them up constantly.

The standard is lower than you think

Television has trained all of us to expect “beyond a reasonable doubt.” That’s the criminal standard, and it’s deliberately high because the state can take away your freedom. Academic hearings almost never use it.

The common standard is preponderance of the evidence. In plain terms: is it more likely than not? People often describe it as anything past the 50 percent line, even 50.1 percent. A handful of schools use a middle standard called clear and convincing evidence, which is tougher but still short of the criminal bar.

Why does this matter to you? Because a lower standard means a thinner case can still stick. If the only thing on the table is a detector reading of 70 percent, and the panel treats “more likely than not” loosely, you can see how a flimsy accusation gains weight it hasn’t earned. Knowing the standard your school uses (it’s in the student handbook, usually under “academic integrity procedures”) tells you exactly how high the wall they have to climb really is.

Why a detector score isn’t proof

This is the crux of the whole problem. An AI detector doesn’t announce “this was written by a machine.” It produces a probability. The output is a guess dressed up as a number, and a confident-looking percentage hides a lot of uncertainty underneath.

A few facts worth keeping in your back pocket, because they’re verifiable and they shift the conversation:

• OpenAI retired its own AI text classifier in July 2023 because of its low rate of accuracy. The company that builds the models couldn’t reliably detect their output.
• A 2023 Stanford study (Liang et al.) found that detectors were biased against non-native English writers, flagging their genuine work as AI far more often.
• Detectors work by measuring statistical patterns. If you want the mechanics, here’s a deeper look at how detectors output probabilities, not proof, built on perplexity and burstiness.

So a detector score is evidence in the same way a smudged fingerprint is evidence. It might point somewhere. It is not, by itself, a conviction. A careful panel treats it as one input that needs corroboration, like a sharp and unexplained shift in writing style, or a student who genuinely can’t talk through their own argument. A bare percentage standing alone rarely clears a fair preponderance bar.

Mini-scenario: the 92 percent flag

Picture Maya, a junior who writes her midterm essay over four evenings in Google Docs. She runs it through a grammar checker at the end, accepts a batch of phrasing suggestions, and submits. Two weeks later she’s flagged: a detector scored her essay at 92 percent AI.

If Maya assumes she has to prove her innocence, she panics and over-explains. But the burden isn’t hers. The instructor has to show it’s more likely than not she used AI. Maya’s move is to put the real evidence on the table: her Google Docs version history, which shows the essay growing sentence by sentence across four days, complete with the messy deletions and rewrites no AI session produces. That history speaks directly to the question that actually matters: did she do the work? A 92 percent score is an opinion. A timestamped paper trail is a record.

(Worth flagging: the grammar checker probably nudged her perplexity down, which is part of why the score came back high. Polished doesn’t mean fake, but detectors can’t tell the difference. These percentages are illustrative, by the way, not figures from any specific case.)

When the burden quietly shifts to you

On paper the school carries the burden. In a real room, it can slide onto your shoulders without anyone admitting it. Watch for these moments:

• The score becomes the verdict. A panel opens by treating the percentage as settled fact and asks you to explain it away. Now you’re defending, which means you’re carrying the burden they were supposed to hold.
• The policy is vague. If your handbook never defines what “unauthorized AI use” even means, panels improvise, and improvisation tends to favor whoever spoke first.
• Silence reads as guilt. Some students stay quiet out of fear. An empty record gives the panel nothing to weigh against the accusation, so the one-sided story wins by default.

You push back not by getting adversarial, but by gently naming the standard. Something like: “I understand the concern. Can we talk about what evidence beyond the detector score supports this, since the score is a probability?” That sentence does real work. It reminds everyone in the room where the burden sits.

Building your side of the record

You can’t control the accusation, but you can control how much evidence exists on your behalf. The strongest material is process evidence, the trail that shows your thinking developing over time:

• Version history. Google Docs and Microsoft 365 both keep it automatically. It’s hard to fake and easy to read.
• Drafts and outlines. Even rough ones. Especially rough ones.
• Research notes, bookmarks, library checkouts, timestamps. Anything that situates the work in your actual life.
• Your own voice in the room. Being able to walk through your argument, your sources, and the choices you made is persuasive in a way no document is.

None of this is about gaming a system. It’s about being able to answer the only question that genuinely counts: did you do the work? If you did, your fingerprints are all over the process, and process evidence is exactly how you show them.

A quick word on tools, since people ask. An AI humanizer is a revision instrument, not a magic shield, and it can’t invent a writing process you didn’t follow. Where understanding detection helps is recognizing the patterns that trip false positives in the first place. If you want to see how PaperBleach reads your text, you can learn to write with that awareness instead of fear. For more on the policy side of all this, browse our other guides on AI detection and integrity.

Frequently asked questions

Who has the burden of proof in an academic AI accusation? In the overwhelming majority of integrity systems, the institution does. The accuser has to show a violation more likely than not happened; you don’t have to prove you wrote the paper, though showing it usually helps. Check your school’s specific policy, since wording varies.

Is an AI detector score enough to prove I used AI? Usually not on its own. A detector returns a probability, not a fact. False positives are documented, and OpenAI pulled its own classifier in July 2023 for poor accuracy. A fair panel treats a score as one data point needing corroboration.

What standard of proof do honor hearings use? Most use “preponderance of the evidence,” meaning more likely than not. Some use the higher “clear and convincing” standard. Almost none use the criminal “beyond a reasonable doubt” bar. The lower the standard, the more your own evidence matters.

What evidence helps if I’m accused but didn’t use AI? Process evidence: saved drafts, version history, research notes, timestamps, outlines, plus your ability to discuss the work in person. It speaks to whether you did the work, not to a probability score.

Can I use an AI humanizer to avoid a false accusation? A humanizer is a revision tool, not a shield, and it can’t manufacture a process you didn’t follow. If you wrote your own work, your drafts and version history are far stronger protection than any rewrite. Tools help most by showing you how detectors read text, so you can recognize the patterns that trip false positives.

The bottom line

The burden of proof problem isn’t really about technology. It’s about who has to convince whom, and the answer almost always favors the student more than the panic in that first email suggests. The school has to prove its case to a “more likely than not” standard, a detector score isn’t proof on its own, and your saved drafts can carry more weight than any percentage.

Understanding how detection works is the calm version of preparing for this. If you want to see exactly how a detector reads your writing before anyone else does, try PaperBleach and write with knowledge instead of dread.