AI-written submissions — what departments can realistically do
Since late 2022, every department has faced the same question and most have answered it badly — either by banning AI outright in a policy nobody can enforce, or by ignoring it until a viva goes wrong.
Neither works. Here is a position that does.
Start with what detection can and cannot do
AI-detection tools return a probability, not a verdict. They work by measuring statistical properties of text — how predictable each word is given the preceding ones, how uniform the sentence structure is — and comparing that to what human writing usually looks like.
That produces two well-documented failure modes:
- False positives on non-native English writers. Writers working in a second language often produce more conventional sentence structures and vocabulary, which is exactly what these tools score as machine-like. For a Bangladeshi cohort writing in English, this is not an edge case — it is the norm.
- False negatives on lightly edited output. Paraphrasing AI text, or asking the model to write less formally, reliably drops detection scores.
The practical conclusion: a detector score is a reason to look closer, never a reason to accuse. Any policy that treats a percentage as evidence will eventually punish an honest student, and one such incident does more damage to a department's credibility than a dozen undetected cases.
Write a policy that distinguishes uses
"No AI" is unenforceable and, increasingly, not what anyone actually wants. A workable policy separates categories:
| Use | Typical position |
|---|---|
| Grammar and spelling correction | Allowed, no disclosure needed |
| Rephrasing the student's own sentences | Allowed with disclosure |
| Generating code the student then understands and modifies | Allowed with disclosure |
| Generating prose submitted as the student's own analysis | Not allowed |
| Fabricating results, data, or citations | Academic misconduct, always |
The specific lines are yours to draw. What matters is that the policy names categories rather than technologies, so it survives the next tool.
Add a disclosure statement to the submission process — one field, asking what tools were used and how. Most students will fill it in honestly. That single field converts a detection problem into a documentation one.
Fix the process, not just the checking
Detection is the weakest control available. These are stronger, and none depend on a tool:
Require intermediate versions. A thesis that appears as a single complete document has no history. One that arrives as six versions across a semester, each building on the last, is very hard to fabricate after the fact. Milestone-based submission gives you this for free.
Ask about the work, not the text. A ten-minute conversation where a student explains a methodology choice, or why a particular result surprised them, is more diagnostic than any detector. Schedule it as a gate, not as an accusation.
Check citations. Fabricated references are still the most reliable signal of generated text, and they are trivially verifiable. A single unresolvable DOI is worth more than a 90% detector score.
Compare against the cohort. Duplicate-file detection across the submission pool catches straightforward copying between students, which remains far more common than AI generation and is much easier to prove.
If you do use a detector, use it correctly
Detection tools have a place — as a triage filter, not a judge. Rules that keep them useful:
- Never show the score to the student as an accusation. Use it to decide where to spend review time.
- Never act on a score alone. Pair it with a viva question, citation check, or version-history review.
- Record what you did. If a case escalates, the defensible position is "we reviewed this submission and here is what we found", not "the tool said 87%".
- Apply it to everyone or no one. Selective screening based on suspicion is how bias enters the process.
What to tell students
Say it explicitly, early, and in writing:
We check submissions for similarity and AI-generated text. A flag is not an accusation — it means a person will read your work more closely. If you used AI tools, disclose it in the submission form; disclosed use is assessed on its merits, undisclosed use is treated as misconduct.
That framing does most of the work. It removes the incentive to hide, and it makes the consequence of concealment clearly worse than the consequence of honesty.
The uncomfortable summary
There is no tool that reliably tells you whether a paragraph was written by a person. There probably will not be one. What there is, and what holds up:
- A version history that shows work developing over time
- A conversation with the student about their own methodology
- Citations that resolve
- A disclosure field that most students fill in truthfully
- A policy that distinguishes assistance from substitution
Detection scores sit on top of that as a triage hint. Build the process so it would still work if every detector disappeared tomorrow.
Supervizo runs similarity and AI-detection checks inside the submission workflow, alongside immutable version history for every milestone. See the AI tooling.