By Stuart Kerr, Technology Correspondent, LiveAIWire

The AI exam cheating crisis reached a new threshold in December 2025, when the Association of Chartered Certified Accountants told its members that online exams could no longer be trusted to certify their competence. From March 2026, the accounting body ended routine remote invigilation for its core professional papers, sending candidates back to physical test centres. ACCA is not a marginal player making a defensive move. It represents more than 250,000 qualified members and over 500,000 students across more than 180 countries, and its decision is the clearest signal yet that a major professional institution has concluded the current examination system cannot survive contact with the AI tools now available to candidates.

The scale of the problem is not theoretical. A University of Reading study published in the journal PLOS ONE secretly submitted ChatGPT-generated answers alongside real student work across five undergraduate psychology modules and found that 94 percent of the AI submissions went undetected by the markers grading them. The AI answers also scored higher, on average, than the real students’ work. That single result captures the exam crisis in miniature. Assessment systems built for a world without generative AI are now failing to catch the tools that have replaced the skills they were designed to test, and in some cases rewarding them for it.

This matters well beyond accountancy. Anyone who holds a professional qualification, is studying for one, or relies on graduates and credential holders being competent at their jobs has a stake in whether assessment can still tell the difference between someone who understands the material and someone who prompted a chatbot. It also matters to a group of people who are not cheating at all: international students and non-native English speakers, who the evidence shows are being wrongly flagged by the detection tools institutions have deployed to fix the problem. That equity failure, not the cheating itself, may be the more urgent part of this crisis.

The AI Exam Cheating Crisis Is a Validity Problem, Not a Detection Problem

Framing this purely as a cheating problem misses what is happening. When OpenAI launched GPT-4 in 2023, it reported a score in the top 10 percent of takers on the Uniform Bar Examination. A later re-analysis by an MIT researcher found that against first-time test-takers specifically, the more accurate comparison, the model’s performance was closer to the 48th percentile. Even on that more conservative reading, a general-purpose chatbot can produce answers competitive with a licensed lawyer’s. Anthropic’s research on AI and labour markets found that 94 percent of tasks within computer and mathematical occupations are theoretically within reach of current models, even though real-world usage for those tasks sits closer to 32 percent today.

That gap between theoretical capability and actual use matters, because it shows the crisis is not really about AI running rampant. It is about assessment formats that were never designed to distinguish a human doing the work from a machine capable of doing it convincingly. UK figures on AI-related academic misconduct show proven cases rising from roughly 1.6 per 1,000 students in 2022-23 to about 7.5 per 1,000 by 2024-25, a near-fivefold increase in three academic years. LiveAIWire’s earlier reporting on the wider cheating crisis in schools and universities found that rise running alongside a decline in traditional plagiarism, suggesting students are switching methods rather than cheating more overall.

Why the Fix Is Punishing the Wrong Students

The equity problem is the least discussed and most serious part of this story. A 2023 Stanford study tested seven widely used AI detectors on TOEFL essays written by non-native English speakers and found an average false positive rate of 61.3 percent, compared with roughly 5 percent on essays by native English-speaking students. The researchers behind the study found the detectors were not really identifying AI text. They were identifying the careful, grammatically conservative sentence structure common among people writing in a second language, and mistaking it for machine output. Non-native speakers were flagged as AI-generated writers at roughly twelve times the rate of native speakers, for reasons that had nothing to do with whether they had used AI.

This is not a hypothetical risk. Australian Catholic University flagged nearly 6,000 students for suspected AI misconduct in 2024 using a Turnitin AI-detection tool, and internal documents later showed the university knew the tool was unreliable for more than a year before quietly discontinuing it in March 2025. Roughly a quarter of those referrals were dismissed, in many cases because the algorithm’s flag was the only evidence against the student. One student had her results withheld for six months while an accusation built on nothing but a detector score worked its way through the university’s process, a case later reported in detail by technology outlet Futurism.

The institutions with the most credibility on this issue have stopped treating detection tools as reliable enough to build policy around. Princeton’s McGraw Center for Teaching and Learning and MIT Sloan’s Teaching and Learning Technologies group have both told instructors not to rely on AI detectors to make findings of academic misconduct, pointing to the tools’ high error rates. The Center for Democracy and Technology found that 62 percent of teachers now feel more distrustful of their students’ work, even though only 19 percent of students who use AI say they submitted it unedited. A 2025 study of faculty at one internationalised university found just 28 percent rated their AI-specific plagiarism guidance as effective.

The institutions handling this responsibly have converged on the same two adjustments. They require human review of every flagged case before any allegation is raised, rather than letting a detector score stand as evidence on its own. And they train staff to recognise that the formal register, avoided contractions, and careful sentence structure typical of second-language writing can trigger false positives that reflect writing quality, not AI generation. Neither step solves the underlying detection problem. Both prevent the worst outcomes while institutions do the slower work of redesigning assessment to depend on detection less.

What Assessment Looks Like When It Stops Trying to Outrun AI

The institutions generating credible responses have stopped asking how to catch AI use and started asking what their assessments are actually supposed to certify. That question points toward formats AI cannot yet fake convincingly: oral examination, where a candidate reasons out loud in response to unscripted follow-up questions; portfolio assessment that tracks how someone’s thinking develops over weeks rather than judging one submission; and supervised practical assessment that observes a candidate applying knowledge in a real setting rather than describing how they would. ACCA’s shift to physical test centres is a blunter version of the same instinct: move the moment of assessment somewhere a chatbot cannot reach it.

None of this is cheap. Oral examination requires far more examiner time than marking a written script, and portfolio review requires sustained engagement from staff who are already stretched. HEPI’s most recent survey of UK undergraduates, published in March 2026, found that nearly two-thirds now say assessment has changed significantly because of generative AI, up from 59 percent a year earlier, and that 94 percent use AI to help with assessed work. The direction of travel is not in doubt. What institutions are still working out is how fast they can afford to move, and how to keep resource-heavy formats fair to students who cannot easily travel to a test centre or take unpaid time for a placement.

The Public Interest Reason Professional Bodies Cannot Wait

What makes the professional credentialing version of this crisis more urgent than the university version is the public interest attached to it. A university degree is mostly a private signal between graduate and employer. A professional licence exists specifically to protect the public from incompetent practitioners. Helen Brand, ACCA’s chief executive, put the underlying problem plainly in comments reported by AccountingWEB: cheating technology is “outpacing what can be put in, in terms of safeguards.” If a credential can be obtained without demonstrating the competence it certifies, the harm does not stay contained to the exam hall. It shows up later, in the audit that missed something, or the advice given to a client who had no way of checking it.

The same dynamic is visible outside accountancy in the broader erosion of quality controls that AI adoption is exposing across professional and editorial work, as LiveAIWire’s reporting on the quality collapse in professional standards has documented. Regulators and professional bodies carry a specific obligation, beyond what a university owes its own students, to ensure the credentials they issue mean what they say. That is why ACCA’s decision carries weight beyond its own membership. It signals that professional credentialing has accepted the old format cannot be defended, and must now answer the harder question of what should replace it.

Exam cheating is not a new problem for the accountancy profession, but AI has changed its scale. As far back as 2022, the UK’s Financial Reporting Council flagged exam cheating as a live issue after finding cases inside major audit firms, and that same year EY paid a 100 million dollar fine to US regulators after dozens of its staff were found to have cheated on an ethics exam and then misled investigators about it. What AI adds to that history is accessibility. A safeguard that once required insider access to leaked material now requires nothing more than a subscription and a phone.

The Reform the Exam Crisis Is Forcing

The institutions most likely to come through this well are treating the exam crisis as a reason to build assessment that was overdue anyway, rather than a threat to manage with better detection software. Oral examination tests a candidate’s ability to reason under pressure and respond to challenge, closer to what professional practice actually requires than reproducing a memorised answer. Portfolio and practical assessment reward the sustained development of judgement over time, which most evidence on learning suggests is a better predictor of real competence than performance in a single high-stakes sitting.

That case is made in more depth in LiveAIWire’s analysis of how AI is forcing higher education to justify what it teaches, and our earlier look at why students are quietly replacing search engines with AI covers the adjacent shift in how students actually learn. Both pieces address the university end of this crisis in more depth than this article does; this one has focused on the professional credentialing stakes and the equity cost of the tools institutions reached for first.

None of this makes the transition comfortable. It is more expensive, harder to schedule, and less convenient for candidates than the remote exam infrastructure it is replacing. But the alternative, chasing detection tools that misclassify a large share of essays from fluent non-native English speakers while missing most genuinely AI-written work, is not a stable place to leave a credentialing system the public is relying on. ACCA concluded that race was not winnable. The institutions that reach the same conclusion early, and start building assessment around what AI genuinely cannot do, are the ones likely to still have credentials worth trusting in five years.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.