Generative AI has entered Ghana’s examination halls, and it defeats the main method WAEC uses to catch cheating
During the 2026 Basic Education Certificate Examination, seven people were arrested across the Central, Bono and Ashanti Regions. Five were found with mobile phones inside examination halls. According to WAEC’s Head of Public Affairs, they had photographed question papers and circulated them in WhatsApp groups, and some had used ChatGPT to generate answers that were then dictated to candidates during the paper.
The candidates were fourteen.
Not a sudden arrival
This had been anticipated. In December 2023, releasing the WASSCE results, WAEC withheld the scripts of candidates in 235 schools for collusion and for the suspected use of artificial intelligence.
What has changed in the three years since is not the availability of the tools. It is that suspicion is no longer required.
Why this is a detection problem
The more serious difficulty is not the cheating itself but what it does to the capacity to detect cheating.
Ghana’s most effective detection method is script scrutiny during marking. An examiner notices that a bundle of scripts from one centre carries the same answers and, crucially, the same errors. Shared errors are the signature. Thirty candidates working honestly get things wrong in thirty different ways; thirty candidates working from one distributed answer get the same things wrong in the same way, and that improbability is what allows the Council to act against a whole centre.
Generative AI does not behave like a photocopied answer sheet. Asked the same question repeatedly, it produces differently worded and differently ordered responses that are largely correct. The statistical signature that collusion detection relies on is much weaker, and may be absent.
The same holds for physical evidence. Prepared notes and textbooks can be seized, which is part of why foreign materials remain the largest single sanction category. An answer delivered verbally leaves nothing on the desk.
The risk, then, is that an offence which is becoming easier to commit is also becoming harder to prove. If that is what is happening, published sanction figures could fall while the underlying conduct does not, and the fall would be read as progress.
What an examination item is for
An examination item is a deliberately constructed instrument. It is written so that candidates who possess a particular kind of thinking tend to answer correctly and those who do not tend not to. That relationship is the basis on which a score can be said to mean anything.
When answers enter the hall from outside, the item does not become somewhat less accurate. It stops measuring the candidate’s thinking and begins measuring access to an adult with a phone. Two candidates with the same mark no longer have the same thing in common.
An instrument answered by something other than the person being assessed has stopped functioning as an instrument.
What is likely to help, and what is not
Detection technology is not a solution. Metal detectors and phone sweeps are worth doing, and were already in use during the sittings that produced these arrests. AI-detection software is unreliable on adult prose and less reliable still on the writing of fourteen-year-olds. A false accusation against an honest candidate is a serious harm, and an integrity measure that produces them is not an improvement.
Computer-based testing is not automatically the answer either. Serialised, randomised delivery does disrupt the leak-and-distribute model, and WAEC has moved in that direction. But a national rollout on current infrastructure would favour urban, better-resourced schools first, which would substitute an equity problem for an integrity one. That trade-off is worth stating before procurement rather than after.
Reducing the stakes addresses the cause. Where a single morning determines a candidate’s future, and where the adults around that candidate doubt he can manage it alone, the incentive to seek help is structural rather than simply a defect of character. School-based assessment, properly moderated and accumulated over two years of visible work, is harder to fabricate than a single paper, and it changes what a teacher is for. It is slower and more expensive than any device, and it is the only measure on this list aimed at the cause.
The candidate who never finds out
There is a sequence worth naming.
A candidate passes the BECE with assistance and is placed on the strength of it. The work is beyond him, so the assistance continues, by now with teachers and parents invested in the outcome. He passes the WASSCE. He is admitted to train as a teacher, a nurse or an engineer.
At no point has he completed a sustained piece of reasoning unaided, and so he has never established what he is capable of. Later, people will depend on the answer.
The concern about generative AI in examinations is not that the technology is capable. It is that it can be used to spare candidates the experience of finding out what they know, and that experience is most of what schooling is for.
Data sources
Accounts and figures are drawn from WAEC Ghana press releases and statements by the Council’s Head of Public Affairs on the 2023 WASSCE and the 2026 BECE, and from reporting by Graphic Online, Citi Newsroom and the Ghana News Agency.
Ato Kwamina Arhin is a Senior Lecturer at the University of Skills Training and Entrepreneurial Development (USTED), Kumasi, and lead at Arc Educational Consult. He has studied, worked and researched extensively in assessment practices and mathematics instruction. He can be reached at
