What a 98% AI Score Actually Means: How Educators Should Interpret Detection Probabilities

A student’s essay comes back at 98% AI, and your stomach drops. It can feel like a verdict before you’ve read a word of the actual paper. But that number isn’t a confession, it isn’t a measurement, and it isn’t proof of anything a student did. It’s a probability, and probabilities trip up people who treat them as facts.

If you teach, you’ve probably had to decide what to do with a score like that. This piece is about reading it correctly before you act on it.

Key takeaways

• A 98% AI score is a probability estimate, not a measurement of how much of the paper a machine wrote, and not 98% certainty that cheating happened.
• Detectors guess from patterns like low perplexity. They can’t see who typed the words or whether any tool was used.
• Even a tiny false-positive rate produces real wrongful flags at classroom scale, and the risk falls hardest on non-native English writers and some neurodivergent students.
• A high score should open a conversation, not close a case. Treat it as one weak signal next to drafts, version history, and the student’s own voice.
• Document your process. The score is evidence about text. It’s never proof about a person.

What the percentage is actually reporting

Here’s the most common misread. A 98% score does not mean “98% of this essay was written by AI.” What it usually means is that the model estimates a 98% chance the document, taken as a whole, matches the statistical fingerprint of AI-generated writing.

Those two claims are miles apart. One is about how much of the paper a machine wrote. The other is just the tool’s confidence in a yes-or-no hunch.

And confidence isn’t correctness. A weather app can call a 98% chance of rain on the morning you walk into a dry afternoon. The model is telling you how strongly the text resembles its idea of “AI,” based on the examples it trained on. It can’t tell you what happened in the student’s room at 1 a.m.

Detectors land on these numbers by measuring things like perplexity (how predictable each word is) and burstiness (how much sentence length and rhythm vary). AI tends to write smooth, low-perplexity, evenly paced prose. So does a careful student following a five-paragraph template. If you want the mechanics, we walk through the statistics behind how detectors actually score text in a separate piece. The short version: the model rewards unpredictability and flags polish, and plenty of honest writing is polished.

Probability is not proof

Tools in this space are usually careful to call their output a “likelihood” or “probability,” not a determination. That wording isn’t a legal dodge. It’s accurate. The system saw patterns and made a bet.

When you turn that bet into a disciplinary outcome, you’re adding certainty the tool never claimed to have.

Why a 98% can still be wrong

Two things explain most wrongful flags: base rates and biased inputs.

The base-rate trap

Picture a detector that’s “99% accurate.” Sounds airtight. Now run a class of 30 students through it, and say every one of them wrote their own work. A 1% false-positive rate works out to roughly a 1-in-4 chance that at least one honest student gets flagged on that single assignment. Stretch that across five sections and a full semester and false accusations stop being a freak event. They become a scheduled one.

This is the part the percentage hides. A high score on one paper feels rare. Across hundreds of submissions, rare things happen all the time. (That 1-in-4 figure is an illustration of how base rates work, not a published accuracy claim about any specific product.)

Some students get flagged more

The errors aren’t random, either. A 2023 Stanford study led by James Zou found that several popular detectors disproportionately misclassified writing by non-native English speakers as AI-generated; in one test, more than half of TOEFL essays written by non-native speakers were wrongly flagged. Why? Second-language writers often lean on simpler vocabulary and steadier sentence structures, which reads as low perplexity, which reads as “AI” to the model.

The same logic can snag students who write in plain, even prose and students drilled on rigid essay formats. It may also catch some neurodivergent writers whose natural style runs more uniform. None of them did anything wrong. The tool just doesn’t like how regular their writing looks.

A mini-scenario: Maya’s lab report

Maya turns in a chemistry lab report. It comes back at 96% AI.

The bad workflow: you see the number, mark the report as misconduct, and email her parents. Maya, who wrote every word, now has to prove a negative against a tool that won’t show its reasoning. The burden lands on the person with the least power in the room.

The better workflow: you notice the flag and get curious. Lab reports are formulaic by design, with procedure, observations, and results in a fixed structure. That’s exactly the kind of low-burstiness writing detectors over-flag. So you open her document’s version history and see 40 minutes of edits, typos fixed, a sentence reordered. You ask Maya to walk you through her method, and she does, fluently, including a mistake she made with a measurement. The score hasn’t budged. Your reading of it has.

Same number, two very different outcomes. The difference was treating the percentage as a first question instead of a final answer.

How to interpret a score responsibly

You don’t have to throw the tool out. You just have to put it in its place: one input among several.

Treat it as one signal, not the signal

Strong evidence about authorship is rarely a single number. It’s a pattern. The writing matches the student’s known voice from earlier in-class work. The drafts show real evolution. The document metadata is consistent. The student can talk through their own argument without notes. A detector score is one thin slice of that picture.

When the score and everything else point the same direction, you have something. When the score stands alone against a student’s track record, the score is probably the weak link.

Look at the writing process, not just the product

Process evidence is harder to fake and far more telling than a probability:

• Version history in Google Docs or Word shows whether a paper grew over time or appeared fully formed.
• In-class writing samples give you a baseline for each student’s natural voice.
• A short conversation reveals whether the student actually owns the ideas.

These cost a few minutes and tell you more than any percentage will.

Separate the tool from the verdict

Set a rule for yourself: a detector flag triggers a review, never an automatic penalty. The flag says “look here.” Your judgment, informed by the evidence above, decides what it means. And whatever you decide, write down what you looked at and why. A documented, repeatable process protects students from arbitrary calls and protects you from claims of unfairness. If your school is weighing which checker to lean on, our pricing and plans page lays out what’s included so you can compare honestly.

The bigger pattern worth remembering

Detection always trails generation. As language models shift, detectors scramble to catch up, which is why accuracy wobbles instead of climbing in a straight line. OpenAI made that concrete in July 2023 when it shut down its own AI Text Classifier, citing a low rate of accuracy. The company that builds the generators couldn’t reliably detect them.

That’s not a reason to ignore detection. It’s a reason to hold every score loosely. A 98% from any tool is a hypothesis with a real error rate baked in, not a measurement you can take to a hearing on its own.

Frequently asked questions

Does a 98% AI score mean 98% of the paper was written by AI?

No. Most detectors report the probability that the whole document resembles AI-generated text, not the fraction a machine produced. A 98% reading means the model is highly confident the writing matches AI-like patterns. It says nothing reliable about how much was AI, and a confident model can still be confidently wrong on short, formal, or formulaic writing.

Can I fail a student based on an AI detector score alone?

You shouldn’t. A score is one statistical signal, not proof of misconduct. Most integrity policies require a preponderance of evidence, and a probability from a tool that admits its own error rate rarely clears that bar by itself. Use the flag as a reason to look closer, not as a verdict.

Why do AI detectors flag honest students?

Detectors penalize writing that’s too smooth or too predictable. Students who write in plain prose, learned English as a second language, or follow rigid templates can produce low-perplexity text that looks AI-like. The 2023 Stanford study led by James Zou found detectors disproportionately misclassified non-native English writers as AI.

Are AI detectors getting more accurate over time?

They improve in some spots and slip in others as models change, and accuracy is never absolute. OpenAI retired its own AI Text Classifier in July 2023 for low accuracy. Treat every score as provisional, because detection keeps lagging behind generation.

What’s a fair way to respond to a high AI score?

Start with curiosity. Tell the student their work was flagged, ask them to walk you through how they wrote it, and review drafts or version history. If the writing matches their known voice and process, the number doesn’t override that. Document what you reviewed and how you decided.

Reading the number, not fearing it

A 98% AI score is loud, but it’s still just a probability with a known habit of being wrong about certain students. Read it as a prompt to look closer, weigh it against process and voice, and keep the final call human. For more on detection, false positives, and academic integrity, browse our other posts on detection and academic integrity.

Want to see what trips these tools before you judge a student’s work? Run a draft through PaperBleach and watch how clean, human writing actually scores.