# Detecting AI in student work: what a detector can and cannot prove

> **The short answer**: A detector score on its own does not prove that a student used AI. It is a statistical estimate about how a text looks, so it needs human review: ask for drafts and an outline, talk with the student about the work, then decide on the evidence and your course policy.

**Key takeaways**

- A detector score estimates how text looks. It is not evidence of who wrote it.
- Ask for drafts and an outline before deciding, with a reasonable deadline.
- A short conversation about writing choices shows understanding better than a number.
- Record your reasons in writing and give the student a right to respond.

![Detecting AI in student work: what a detector can and cannot prove](https://tiswa.app/assets/blog/detecting-ai-in-student-work/detecting-ai-in-student-work-cover-en.webp)

A report arrives with a number in bold, and the temptation is to read it as a verdict. It is not one. The score reflects how a text looks to a model trained on other texts, so it changes with length, format and register, and it says nothing about who typed the words. Treat it as a reason to look, never as the reason to act.

## How reliable is an AI writing detector?

Most detectors compare patterns in a text with patterns they associate with machine-generated writing, then return a likelihood. A Stanford team published a study in Patterns in 2023 that tested several popular detectors on essays by native and non-native English writers. The detectors labelled many non-native essays as machine-written, and classified native-speaker essays far more accurately.[^1] OpenAI, which released a public classifier, switched it off in 2023 because its accuracy was low.[^2]

Short texts and reference lists give weak signals. A student who writes in a formal, tidy register may look machine-made even if no tool touched the work. And a student who used an editing tool for phrasing produces a mixed result that no single number can settle.

## Which AI writing detector evidence settles the question?

The evidence that settles things almost always exists before the deadline: drafts, version history in the file, the outline, and a comparison with earlier work you saw the student produce in class. Do not hunt for one suspicious sentence, since that is how reviews go wrong. Look for a change that repeats across several submissions.

| What you look at | What it can show | How to request it |
|---|---|---|
| Voice across assignments | A sudden shift in vocabulary and sentence rhythm | Compare with an in-class exercise or an earlier short task |
| Drafts and version history | How the argument developed and where it was corrected | Ask for the first draft or earlier saved versions |
| Outline or plan | Whether the structure was planned before drafting | Ask for the list of main points well before the due date |
| Sources | Whether the student read what they cite | Ask what one named source adds to the argument |

![Comparison card of the guide table](https://tiswa.app/assets/blog/detecting-ai-in-student-work/detecting-ai-in-student-work-inline-2-en.webp)


## How do you review without accusing anyone?

1. **Save the evidence first**: Keep the submission, the tool output and the submission time together in one file. Write no accusation in an email or a comment until the review is finished.
2. **Ask for proof of process**: Drafts, an outline or research notes, with a clear deadline for handing them in.
3. **Talk with the student**: About the idea and the writing choices, using the questions in the section below.
4. **Apply the written course policy**: If the permitted level appears in the course specification, compare the work with it. If nothing is written down, do not penalise a rule nobody announced.
5. **Write the decision and its reasons**: Give the student the right to respond and to appeal under your faculty and university procedures.

![5-step flow: Save the evidence first / Ask for proof of process / Talk with the student / Apply the written course policy / Write the decision and its](https://tiswa.app/assets/blog/detecting-ai-in-student-work/detecting-ai-in-student-work-inline-1-en.webp)


If your faculty has an academic integrity committee, send the file there instead of deciding alone. A decision that passed through a committee is harder to overturn, and it protects you as much as the student.

> **tiswa's take**: In our view, the assignments that hold up best are built in stages and tied to a discussion held in class. If your task is a general summary of a topic that every article covers, change it before next term. Waiting for a higher score in a later report will not fix it.

## Which questions reveal understanding in ten minutes?

A short conversation measures understanding directly. The student who wrote the work usually explains the choices easily, even when they stumble at the start. The student who did not often struggles with details of the text they handed in. Keep the tone calm and ask about the work, not about the software.

Prompt for preparing conversation questions:

```
I am a lecturer in [subject]. The assignment is [describe the task], and the main points it should cover were: [list the points].
Write eight open questions for a ten-minute conversation with the student about their writing decisions, why they chose their sources, and how they would apply the main idea to a new case.
Do not write questions that can be answered with one word, and do not ask about specific tools.
```

Write the answers in a short note. Do not expect the student to recite anything word for word, because understanding shows in how they explain a choice.

## How should you design the task so learning shows?

The strongest protection against wholesale copying is the task itself. A prompt about a popular topic can be imitated in a minute. A prompt tied to a class discussion, to data collected in the room, or to the student's own placement is much harder to finish with a tool alone. A [clear rubric](https://tiswa.app/en/blog/rubric-guide/) that rewards stages rather than only the final text makes the same point to students:

- Anchor the prompt to a discussion from the previous session.
- Stage it: a proposal, then a draft, then a final version with your short comment.
- Add a two-minute oral segment where the student defends one decision.
- Ask for a brief reflection on what was hard and what changed their view.

Write down why each assessment decision was made. Your course coordinator or a quality reviewer will ask for that reasoning, and a note written at the time is far more persuasive than one reconstructed later.

tiswa for Universities: student benefits and discounts on campus and beyond, free for the student, designed with each university. Learn more at [tiswa for Universities](https://tiswa.app/en/universities/), and share the [student deals](https://tiswa.app/en/deals/) page with your class. [tiswa for Universities](https://tiswa.app/en/join/)

## What do you do if doubt remains?

If doubt remains, use your university's formal procedure rather than an accusation. The [academic integrity guide](https://tiswa.app/en/blog/academic-integrity-guide/) explains how clear procedures protect the student and the lecturer alike. If you want to review the course policy itself, the guide to [AI in university teaching](https://tiswa.app/en/blog/ai-in-university-teaching/) covers how to set the allowed level.

Two mistakes come up again and again. The first treats a strong student as a suspect because a score went up. The second punishes a rule that was never published. The advice above also does not fit a very short task, where no indicator tells you much.

## Questions

### Is a detector score enough to prove cheating?

No. A score is a likelihood. OpenAI withdrew its classifier because its accuracy was low, so the decision needs evidence from the writing process itself.[^2]

### Are AI detectors reliable for essays written by non-native speakers?

Not on their own. The 2023 Patterns study found many non-native essays were labelled machine-written, so a high score on such a text needs extra care before you read anything into it.[^1]

### What if a high-achieving student gets a high score?

Record the result, ask for drafts or an outline, and talk with the student. Being strong is not evidence of innocence, and it is not evidence of guilt either.

### How can I reduce reliance on a tool in future assignments?

Tie the task to class discussion or to data you collect, stage it, and add a short oral part. Better design works better than chasing cases afterwards.

### Should I tell students that work may be checked with a tool?

Yes. State it in the course specification, and say that decisions depend on human review rather than on the tool alone.


## Sources

[^1]: OpenAI, New AI classifier for indicating AI-written text, https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
[^2]: Liang et al., GPT detectors are biased against non-native English writers, Patterns, https://doi.org/10.1016/j.patter.2023.100779