AI tools and remote exams: writing a policy candidates can actually follow
Start from what the exam measures
An exam certifies something specific: unaided recall, applied judgement, practical skill, or the ability to work with real tools. AI rules make sense only against that purpose. A licensing exam testing emergency decisions without references should ban assistance outright. A professional exam mirroring a job where AI tools are standard equipment may allow defined uses and test the judgement around them instead.
Writing the purpose down first turns borderline cases from arguments into readings of the policy. If the exam measures unaided knowledge, AI explanation during the exam is assistance and out. If it measures output quality with tools, the line sits elsewhere, around undisclosed substitution of the candidate's own work.
Name the tools and the uses
Workable policies list categories with examples: spellcheck and grammar tools, calculators, translation tools, generative assistants, code completion, and full answer generation, each marked allowed, allowed with declaration, or banned for this exam. Candidates follow specifics. They cannot follow a principle they have to interpret under exam pressure with their qualification at stake.
Declaration is the underused middle option. Requiring candidates to declare AI use, and what it was used for, mirrors academic and professional practice, produces an audit trail, and separates honest users from concealment. Concealing a declared category then becomes the clear offence, which is far easier to enforce fairly than policing the tool itself.
Enforcement must match the paper
Detection claims for AI generated work are unreliable, and building sanctions on a detector score alone invites wrongful accusations that damage the exam body's credibility more than cheating does. Stronger enforcement combines question design that resists generic AI answers, process signals from proctoring, and human review of flagged work against the declared policy categories.
Question design deserves the investment. Prompts tied to provided case material, local data, or the candidate's own earlier answers produce work an AI cannot fabricate convincingly without the context, while still testing exactly the skills the exam exists to measure. Policy is the fence; design is the terrain, and terrain works even when nobody reads the fence.
Communicate before the exam, not after
The policy belongs in registration materials, preparation guides, and the exam day instructions, in the same words each time, with worked examples of borderline cases and the answers. A short pre exam attestation where candidates confirm the specific rules, not a generic honesty pledge, proves the message was received and understood.
Finally, review the policy every cycle. Tool capabilities move faster than exam board calendars, and a rule written against last year's assistant will misclassify this year's. The aim is a policy a careful candidate can follow with confidence and a dishonest one cannot hide behind, and that target needs regular maintenance.
Keep reading: Second devices and screen sharing: closing the quiet cheating channels, Impersonation in remote exams: identity checks that actually hold up, and How long should remote exam recordings and proctoring data be kept? cover the nearby cases in this series.
Frequently asked questions
Why is a simple no AI rule not enough for remote exams?
Because AI is embedded in ordinary tools candidates use innocently, like grammar and spelling assistance, and the same tool can help trivially or substitute for the candidate entirely. A blanket ban leaves honest candidates guessing where the line is.
What should an AI exam policy actually list?
Tool and use categories with a status for each: allowed, allowed with declaration, or banned. Spellcheck, translation, generative drafting, and code completion are different acts and a workable policy treats them differently, with examples.
Can AI detectors enforce the policy?
Not on their own. Detector scores are unreliable evidence and produce false accusations. Enforcement works better through question design that resists generic answers, proctoring signals, declaration records, and human review.
What is an AI use declaration in an exam?
A statement by the candidate of what AI tools they used and for what, where the rules allow declared use. It creates an audit trail and makes concealment, rather than use, the clear enforceable offence.