The reliable answer to AI in student writing is assessment design, not detection. Tasks that collect process, anchor some writing in class, and draw on content specific to your group give you evidence no detector can match. They also keep working as models improve, which detection does not.
Key takeaways
- Detection is a losing position: as models produce more varied text, detectors chase a target that keeps moving.
- Collecting drafts and version history turns writing into a visible process rather than a single artefact to be judged.
- One unaided in-class writing task early in a course gives you a baseline for every later submission.
- Tasks built on class-specific content are hard to generate convincingly and better for learning anyway.
- Assessing the spoken defence of a piece of writing sidesteps the question entirely.
- These methods are fairer to language learners, who are the group detectors treat worst.
Why detection cannot be the answer
Two reasons, one technical and one ethical.
Technically, detection works by measuring how predictable text is. As language models improve they produce more varied output, so the signal detectors rely on gets weaker with each generation. Improvements against one model do not transfer to the next. It is a permanent catch-up game.
Ethically, the costs fall unevenly. Research published in Patterns found seven detectors wrongly flagged an average of 61.22% of human-written TOEFL essays by non-native English speakers, against 5.19% for native-speaker essays. In an English language teaching context, building your integrity policy on detection means building it on a tool that misjudges most of your students.
Assessment design has neither problem.
Five approaches that work
1. Assess the process, not just the product
Break a writing task into stages and collect each one: a topic choice, an outline, a first draft, a revision, the final piece. Writing that develops over time, with visible false starts, is difficult to fabricate convincingly.
Shared documents make this almost free, since version history records text appearing gradually rather than arriving in one paste. The pedagogical benefit is the real prize though: drafting and revising is how writing improves, and most courses do too little of it.
2. Anchor one piece in the room
A single timed, unaided writing task early in a course gives you a sample in that learner’s own voice. Every later submission can be read against it.
It is also the best diagnostic you will get. You learn what they can do unsupported, which tells you what to teach, entirely separately from any integrity question.
3. Build tasks on content only your class has
A model can write a competent essay on climate change. It cannot write convincingly about the specific discussion your group had on Tuesday, the text only you assigned, the mistake a classmate made that became a running joke, or a learner’s own experience of arriving in a new country.
Prompts that reference shared class context are harder to generate and produce better writing, because the student has something particular to say.
4. Make the student defend the work
Two or three minutes per student, asking them to explain a choice they made. Why this structure? What was hardest? Where did this example come from?
A writer can answer. Someone who generated the piece usually cannot discuss decisions they never made. This also protects honest students, because it gives them a way to demonstrate authorship rather than disprove a score.
5. Make AI use explicit rather than forbidden
Many institutions now ask students to declare what tools they used and how. A short statement with each submission, naming the tool and describing what it did, moves the conversation from policing to professional practice.
This is closer to how these learners will actually work. Teaching them to use AI transparently and critically is more useful than pretending they will not use it at all.
Choosing the right approach for the task
| Approach | Best for | Cost to you |
|---|---|---|
| Staged drafts and version history | Longer written assignments | Set-up time once, then minimal |
| In-class unaided baseline | Any course with written assessment | One lesson, early |
| Class-specific prompts | Essays and extended responses | Planning time, offset by better work |
| Spoken defence | Small groups and one to one | A few minutes per student |
| Declared AI use | Any level, any task | Almost none once it is routine |
What institutions are doing
The direction of travel is clear. Cambridge University Press and Assessment’s principles for AI-powered marking insist humans remain in the loop for high-stakes judgement, and UNESCO has argued that generative AI forces a rethink of what is worth measuring, pushing assessment toward authentic and process-based evaluation rather than one-off products.
Both point the same way: design assessment that tells you what a learner can do, rather than designing enforcement around a tool that cannot reliably tell you who wrote what.
Where to start
Pick one course and change one thing. The in-class baseline is the easiest, costs a single lesson, and improves your teaching regardless of the integrity question. Add staged drafts to the next long assignment after that.
You will find the detector question quietly stops mattering, because you now have evidence that actually holds.
For the wider context see our pillar guide to AI detectors for teachers, how teachers check for AI, and our guide to AI in English language teaching.
Frequently asked questions
How do I stop students using AI to write essays?
You largely cannot, and designing around it works better than trying. Collect drafts, set one unaided in-class piece as a baseline, and build prompts on content specific to your class. Those make generated work obvious without relying on detection.
What is AI-proof assessment?
Assessment that a model cannot complete convincingly on the student’s behalf. In practice that means tasks tied to class-specific discussion or texts, writing produced in staged drafts, work done in the room, and anything the student has to explain aloud.
Is it better to ban AI or allow it?
Most institutions are moving toward declared use rather than prohibition, because bans are unenforceable and push use underground. Asking students to state which tools they used and how teaches transparent practice and keeps the conversation honest.
Does staged drafting create a lot of extra marking?
Not if you do not mark every stage. Collecting an outline and a draft without grading them still gives you process evidence, and a brief comment at draft stage usually improves the final piece enough to save marking time later.
How does this help English language learners specifically?
It removes reliance on tools that misjudge them. Research found detectors wrongly flagged 61.22% of human-written TOEFL essays by non-native speakers. Process evidence and spoken defence judge what a learner can actually do, rather than how predictable their vocabulary is.
Will assessment design stop working as AI improves?
Less than detection will. Tasks grounded in a specific classroom, in staged work and in spoken explanation stay hard to automate, because the evidence is about the learner rather than about the text.
Quick favour: how do you actually use AI in your teaching?
We are running the first real study of AI use among TEFL teachers. 12 questions, about 3 minutes, fully anonymous. Leave your email at the end and you will get the report before anyone else.
