If AI can ace the exam, how do you prove what you know? A Meta engineer says tests should watch how you solve problems with AI, not the answers
Nov 22, 2026Original 실밸개발자(메타 시니어 엔지니어) · 커리어해커 알렉스 · 티타임즈TV · Canva · 404 Media ↗
In 3 lines
- A Meta engineer says the AI cheating scandal shows exams can no longer confirm anything
- Canva requires AI in its interviews, and Meta lets some coding candidates use it
- Skill now shows in good questions, checking AI's answers, and writing line by line to understand

"The exam can no longer confirm anything." That is how 'Silval Developer,' a YouTuber who works as a senior engineer at Meta, read the confession of a top-scoring tax accountant candidate who cheated with AI, in a video posted November 20. Without defending cheating, the engineer predicted that exams taken with AI at your side from the start will become the norm. We added coding interviews at companies that already allow AI, an education scholar's diagnosis, and a warning from a Silicon Valley practitioner about handing your thinking to AI.
The top scorer missed one of 160 questions, and AI supplied the answers
It started with Korea's 63rd certified tax accountant exam. The top scorer missed just one of 160 questions in the first round in April, scoring 397.5 out of 400, and got a perfect score in Introduction to Tax Law, where the average among test takers was 30.18. As suspicions grew, the candidate turned themself in to police on November 15, confessing to photographing the questions with a micro camera, sending them to generative AI and reading the answers on a smartphone.
세계일보 · Segye Ilbo ↗"The exam can no longer confirm anything"
Silval Developer sees an exam as a way to certify that 'I know at least this much.' If anyone can score perfectly with AI, the exam neither separates people nor confirms their level. AI cheating has been fairly common since around 2023, the engineer said, and remote interviews during the pandemic saw plenty of cheating with a second screen open. A year or two ago, one applicant even publicly claimed to have passed big tech coding interviews with AI. In this view, the confession simply brought to the surface something that was already happening.
Secret use is cheating, but exams with AI at your side are coming
Using AI secretly against the rules is plainly cheating, the engineer said. Yet for a generation that will live with AI all their lives, exams taken with AI at hand from the start will become normal, and should. Borrowing an acquaintance's analogy, the engineer said AI is less a calculator to today's children than a teddy bear they carry everywhere. Questions that AI can push above a certain score are no longer worth asking, the argument goes, so exams should be rebuilt around watching how people work through a problem.
Meta brought AI into coding interviews
Corporate hiring has already moved that way. In July 2025, 404 Media, citing internal Meta communications, reported that Meta would allow some coding job candidates to use an AI assistant during interviews and asked employees to volunteer for a 'mock AI-enabled interview.' Mark Zuckerberg has said vibe coding will soon be a major part of Meta's engineering work. Silval Developer also said Meta has introduced coding interviews in which some candidates use AI, with the focus on how they solve problems with it.
404 Media · 404 Media ↗Canva: "Use AI in our interviews"
Canva went further. In a June 2025 engineering blog post, it said it now expects backend, machine learning and frontend candidates to use AI tools like Copilot, Cursor and Claude during technical interviews. AI solved its old algorithm questions in seconds, so Canva switched to complex, ambiguous tasks such as 'Build a control system for managing aircraft takeoffs and landings at a busy airport.' In the pilot, the strongest candidates asked clarifying questions, critically reviewed and improved AI-generated code, and debugged issues themselves. Candidates with minimal AI experience often struggled, Canva wrote, because they lacked the judgment to guide AI effectively or spot weak suggestions.
Canva Engineering · Canva Engineering Blog ↗"The agent's limit is my limit"
What does it take to learn well with AI? Silval Developer named two things: the ability to quickly understand what an agent produces, and the ability to give it good input, meaning context, so it produces good results. Agents have already gone far beyond human limits, the engineer said: "People are the bottleneck. I am the bottleneck." Since a person sets the goal, the bigger that person's ambition, the bigger the result the agent delivers.
Four parts of a good question, and 'Socratic study'
To the engineer, good input means a good question. A question AI can answer right away without asking back includes four things: why you want to know (the purpose), what you already know and have tried, exactly where you are stuck, and the shape of the answer you want, such as length, format and difficulty. Deeper domain knowledge leads to deeper questions, so the engineer recommended Socratic study, asking the agent to question you. Even when making a presentation, the engineer first attaches a question to every slide and then looks for answers with AI.
A warning about people becoming 'AI routers'
Career Hacker Alex, who has done AI talks with Silval Developer, described a phrase going around Silicon Valley: 'meat proxy.' You ask a colleague something, the colleague pastes it into their own AI, and copies the answer back to you. Alex admitted to working that way and feeling that "information goes back and forth, but nothing sticks." Now, even after having AI gather and verify the research, Alex writes the proposal line by line with AI from a blank document, asking about anything unclear, and understands about 95% by the end. The video, posted November 15, has passed 31,000 views.
커리어해커 알렉스 · YouTube ↗An education scholar: "AI gives information; understanding has to pass through a person"
On the same day, Kim Jae-in, a professor at Kyung Hee University, said on TTimes TV that the cheating cases exposed how today's exams fail to select real experts. Kim divides knowing into data, information, knowledge and wisdom. AI reliably delivers information, but turning it into knowledge requires understanding that happens only inside a person's head. That is why Kim named language ability, centered on reading and writing, as the most important skill ahead. People who already went through that training use AI to extend their abilities, Kim said, but younger people risk keeping only the feeling of having done the work while missing the practice. Kim's conclusion: assessment has to happen through real work.
티타임즈TV · YouTube ↗Two kinds of experience
Silval Developer closed the video with words for students who had just taken Korea's college entrance exam. The first is the experience of actually handing your work to AI and pushing it to the limit to see what happens. The second is meeting new people and unfamiliar places outside AI; the engineer said time with family on weekends or a trip often sparks new ideas. The old formula in which a good school and a good job guaranteed success and happiness is fading, but new opportunities have multiplied, the engineer said. The video's last request was to stay positive.
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.



