The devil's in the details with this one. Stories about AI exam cheating get passed around a lot, and the specifics tend to get fuzzy - especially when you're talking about a particular country's licensing exam. South Korean tax exams are notoriously competitive, sure, but I'd want to see actual reporting from a credible source before accepting the exact narrative you're describing. It's the kind of thing that sounds plausible enough that people share it without digging into whether it actually happened that way.
That said, the broader concern about exam integrity and AI is totally real, even if this specific case might be exaggerated or misremembered. Detection is genuinely hard because AI output can look polished and coherent - it won't have the typos or obvious red flags of a rushed answer. What actually catches people is usually pattern-matching: an answer that's too perfectly written compared to a student's other work, writing style that doesn't match their previous submissions, or sometimes just statistical oddities when an institution is actively looking. A few schools have started using detection tools, though they're imperfect.
The pitfall nobody talks about enough: institutions often can't actually prove AI use in a reliable way. You can *suspect* it, but legally and ethically, catching someone requires either admission (like your example claims happened) or they're caught red-handed with prompts open. That gap between suspicion and proof is huge for policy makers trying to write rules that won't get schools sued. So even if detection improves, enforcement remains messy.