Last updated 9 min read
Key takeaways
Honestly, no unsupervised written task is. If students can take an assignment home and type the prompt into a chatbot, a capable model can produce something that meets a typical rubric for an essay, a summary or a set of short answers. Supervised exams remove that option, but they test a narrow slice of skills under time pressure and cannot be your only tool.
AI detectors do not solve the problem either. They can flag human writing as machine-generated and miss text that has been lightly edited, and concerns have been raised that writing by non-native English speakers is flagged more often. Treat a detector score as, at most, a reason to talk with a student about their work, never as evidence on its own.
So the useful question is not how to make faking impossible but how to make it pointless. An AI-proof assessment, in practice, is one where AI help is visible, where it cannot supply what the task requires, or where using it well is itself the skill being graded.
Most effective redesigns combine two or three of these principles. You do not need all six on every task, and stacking too many makes an assignment confusing for students and slow for you to grade.
Each format below limits AI help in a different way and costs a different amount of your time. The right mix depends on class size and on what the course is meant to teach, which is why it helps to start from clear learning objectives before choosing a format.
| Format | Why AI help is limited | Teacher time | Watch out for |
|---|---|---|---|
| In-class handwritten essay | Supervised, with no device access | Low to set, moderate to grade | Rewards speed; some students need accommodations |
| Oral exam or short conference | Student explains and extends in real time | High in large classes | Anxiety; use set questions and a rubric |
| Process portfolio | Drafts, notes and responses to feedback show development | Moderate | Invented drafts; pair with a conversation |
| Local data project | A general model has none of the data | Moderate | Privacy and permission when collecting data |
| Critique of an AI answer | The AI output is the starting point, not the submission | Low to moderate | Pick an answer with real, findable flaws |
| Practical or performance task | The skill is demonstrated directly | High | Equipment, space and scheduling |
Start from an assignment you already use rather than inventing a new one. Take a common example: a 1,500-word take-home essay comparing the causes of the American and French Revolutions. As written, it is easy to outsource. Here is how to rebuild it without throwing it away.
The essay still exists and still matters, but it is no longer the only evidence. A student who outsourced it will struggle in the conference and the reflection; a student who did the work will find both straightforward, because they are simply talking about decisions they already made.
A short conversation is one of the strongest checks available, because explaining and extending your own work in real time is hard to fake. It does not need to be a formal viva; five minutes per student is enough if the questions are prepared. Do the arithmetic before you commit: thirty students at five minutes each is two and a half hours, which you can spread across lessons while the rest of the class works independently.
Use three question types. Explain: “Walk me through why you chose this source over the others.” Extend: “How would your conclusion change if this event happened ten years later?” Defend: “Someone argues the opposite; what is your best reply?” Score each on a simple three-level scale and note one phrase from the answer, which keeps grading consistent and gives you something concrete if a grade is questioned.
Tell students in advance that the questions are coming and what kind they will be. The aim is not to catch anyone out. It is to make understanding the cheapest route to a good grade.
Another approach turns AI output into the material being assessed. Give students an AI-generated answer to a question in your subject and ask them to grade it against your rubric, find the errors, check every claim against a source and write an improved version with tracked changes. The skill being tested, judging a fluent but flawed text, is one they will need anyway.
Where AI use is allowed, ask for a short log: what the student asked, what they kept, what they changed and why. The log turns invisible use into visible decisions you can grade. Be explicit about which tasks allow it; the student guide on where AI help becomes cheating helps set expectations, and the guide to using AI tutors in the classroom covers routines that make attempt-first use the norm.
Some of the most AI-resistant formats carry their own costs. Timed handwritten work disadvantages students with certain disabilities and those who write slowly; oral exams can be hard for anxious students or those working in a second language. Plan accommodations in advance rather than case by case, and offer an equivalent route wherever a format would stop a student from showing what they know.
Watch your workload too. Swapping every essay for an oral exam is unsustainable in a large class. Mix formats across a term, put the heaviest checks on the assessments that count most, and reuse the same conference questions across a unit so your preparation pays off more than once.
Cavua courses are built around testing, not just reading. Every chapter and lesson has its own test with saved results, and passing the final test earns a certificate that is not accredited by an outside body. If you teach from your own material, you can upload a PDF of it as an author: Cavua builds a syllabus, lessons and tests in four visible phases, and before you approve the course you can rewrite any question that only checks recall into the explain, extend or defend kind described above. Cavua's team then checks the course before it is published.
Each course's 24/7 AI tutor answers from that course's own syllabus, lessons and tests, by video call, voice call, chat or audio summary. That makes it a practice partner for the kind of oral explanation described above: a student can talk their reasoning through aloud before doing it with you. There is more on how it works on the AI tutor page.
AI in the classroom, made practical: what AI tutors do well, the ground rules to set, lesson routines that keep students thinking, and how to tell it helps.
Not if students complete it unsupervised, because a capable AI model can produce text that meets many typical rubrics. What you can do is make AI help visible, make it unable to supply what the task needs, such as class-only sources or local data, or add a live component where students explain their work. Together those make outsourcing pointless rather than impossible.
Not reliable enough to act on alone. Detectors can flag genuine student writing and miss AI text that has been edited, and concerns have been raised that non-native English speakers are flagged more often. Use a score only as a prompt for a conversation about the work, and rely on assessment design for your evidence.
Tasks that need something the AI does not have or cannot do on the student's behalf: supervised in-class work, oral explanations, projects built on local or personally collected data, process portfolios with drafts and feedback, and practical performances. Critiquing an AI-generated answer is also hard to outsource well, because the critique itself is the skill being assessed.
Keep them short and structured. Five minutes per student with three prepared question types (explain, extend, defend) and a simple scoring scale is manageable when spread across lessons while others work independently. You can also check a different group of students on each assignment, provided everyone knows in advance that anyone might be asked.
It depends on what the assessment is meant to measure. If the skill is writing an argument unaided, AI use undermines the evidence. If the skill is research, evaluation or revision, structured AI use with a log of what was asked and changed can be part of the task. State clearly, for each assessment, which applies.
An assessment that asks students to do something resembling real work in the subject, such as analyzing local data, advising a client, designing an experiment or presenting to an audience, rather than reproducing information. Authentic tasks tend to resist AI because they depend on specific context, real decisions and the student's own observations.
Every Cavua course comes with a 24/7 tutor you can video call, speak to or chat with — on the material you are actually studying.
