AI in Professional Licensure Exams: Every Published Sanction Number (2026)

1 What Has Anyone Actually Been Penalised For AI Use On An Exam?

The detail that makes this case useful is how ordinary it was. The partner uploaded the training materials to an AI platform and had it generate his answers. There was no exotic bypass tool and no humanizer involved, which is worth holding onto if your mental model of AI exam cheating is built on the humanizer market and its claims. He used a chatbot the way most people use a chatbot.

It is also worth being precise about what this penalty is not. It was imposed by an employer on its own internal training exam, not by a licensing body on a licensure exam. No regulator issued an order. The figure reached the public record through Financial Times and Australian Financial Review reporting and an Australian senate inquiry, where a senator called the result a toothless system. As a data point about how professional bodies sanction AI use in the exams that grant credentials, it is the best we have and it is thin.

There is one more reason this case is the right place to start. Internal training exams are where the credential system does its quiet maintenance work, the annual ethics module and the compliance refresher that nobody talks about until somebody games one. Those exams are almost always delivered online, almost never invigilated, and sit outside the scrutiny that a licensure sitting attracts. If more than two dozen people at one firm in one country were caught, the base rate across every firm running the same kind of unsupervised annual module is not something anyone has measured.

2 Are The Huge Exam Cheating Fines Actually About AI?

Put next to the AU$10k (commonly cited; no primary source traceable) in the previous section, this table shows a real asymmetry rather than a trend. When a firm is found to have industrialised answer sharing, the penalty runs to tens of millions and an individual leader can be barred for life. When a partner is found to have used AI, the penalty is four figures. Same category of conduct, wildly different consequence, and the only variable that changed is the tool.

The other thing this table does is puncture a rhetorical move you will see often, where a writer opens with the EY figure and pivots to AI in the next sentence. Nothing in the SEC order against EY or the PCAOB order against KPMG Netherlands concerns AI. One detail from the 2019 KPMG order is worth keeping for its own sake: some staff achieved passing scores while answering fewer than a quarter of the questions correctly. Exam integrity was breaking long before anyone needed a language model to break it, which is the same pattern visible in school-level malpractice data.

3 Can AI Really Pass These Exams, Or Was The Percentile Wrong?

This matters for the enforcement question, not just for accuracy. The 90th percentile (commonly cited; no primary source traceable) number is what convinced a lot of committees that their exam was trivially beatable, and committee panic is what produces the policies candidates then live under. The corrected figures describe a system that is mediocre at the multiple choice and weak at the writing, which is a different threat model and calls for a different response.

We are reporting both figures rather than picking one, because both are in the literature and a reader deciding what to believe needs to see the pair. The pattern is familiar from detector accuracy claims, where the headline number and the number measured under realistic conditions routinely diverge, as the disagreement between detectors on the same document shows.

4 Which Bodies Have Pulled Remote Exams Over AI?

Read the decision for what it concedes. ACCA is not claiming better detection, it is removing the channel where detection failed. Its chief executive described the sophistication of cheating systems outpacing the safeguards that can be put around them. That is an exam body saying, in public, that it cannot police AI use in a remotely invigilated sitting, which is a more candid position than most institutions have reached and closer to what universities stepping away from AI detection have concluded.

The direction of travel is backwards, toward supervised rooms and paper, and it is not confined to professional bodies. The same logic is driving the return of the handwritten exam in higher education. If you are a candidate, the practical consequence is that the convenience of sitting at home is being withdrawn because of what other people did with it.

For anyone currently studying, the planning consequence is concrete. A qualification route that assumed you could sit papers from home now assumes you can reach a test centre, which changes cost, travel and scheduling for candidates in places where centres are sparse. That burden falls hardest on exactly the candidates remote delivery was expanding access for, and it is being imposed without any published evidence about how common the underlying problem actually is.

5 What Do The Exam Rules Actually Say, And What Can They Do To You?

We extracted the full text of the April 2026 rules and searched it for artificial intelligence, AI, generative, ChatGPT, machine learning and large language. Every term returned zero. What the document does contain is a prohibition on seeking help from any other party, which is the clause AI use would have to be read into, and a provision that results may be cancelled where there is a basis to question their validity.

The evidence standard is the thing to notice, and it is the reason this page exists. Invalidation does not require a detector finding or an admission. Unusual answer patterns or an unusual score increase from one sitting to the next is enough to open the question, and the sanctions available run all the way to a revoked license. If you have ever wondered whether a score jump could be held against you, the answer in the rule text is yes, and the same evidentiary weakness that makes detector output a poor standalone proof in academic cases applies here with far more at stake. Building a record of your own preparation, in the way an authorship packet does for written work, is the only lever a candidate actually controls.

Two practical notes follow from the rule text rather than from any advice column. First, the clause that would catch AI use is a general prohibition on outside help, which means a body does not need a new AI policy to act, and the absence of the words in the document tells you nothing about your exposure. Second, because the trigger for questioning a result can be a score pattern rather than an observation, a candidate who legitimately improves a great deal between sittings is in the same evidentiary position as one who did not. That is an uncomfortable place to stand with no contemporaneous record of how you prepared.

6 How Many Candidates Have Been Sanctioned? Nobody Publishes It

We checked the obvious places. That article is the credentialing sector’s own publication writing directly about AI and exam security, and it is sponsored by a proctoring vendor. It asks whether a credential still proves anything if AI helped the candidate pass, and it answers with rhetoric rather than data, because the data has not been collected or has not been released. The pattern of vendors supplying the analysis for a problem they sell the remedy to is one we have documented before in the detection industry numbers.

So the honest state of the record is this. One employer-imposed four-figure fine, more than two dozen cases at that single firm, one large body withdrawing remote delivery, a set of rules that can void your result without naming the tool, and no published counts from anyone. Anybody telling you what percentage of licensure candidates are using AI is guessing. We will update this page the moment a body publishes an actual number, and for the school-level equivalent the counts that do exist for students are a useful contrast, alongside the documented consequences students face and the measured effect of all this on the people being tested.

It is worth saying plainly why this absence is not innocent. Bodies collect this information as a matter of course, because every invalidated result generates a file. Publishing counts would expose both how rarely AI use is actually proven and how often results are voided on inference, and neither number flatters the institution. Until one of them publishes, candidates are being asked to trust a process whose error rate nobody outside the organisation can see.