The article says Danish high school students will increasingly have to verbally defend written work, a move aimed at making it harder to outsource assignments to AI and easier to check whether a student actually understands what they turned in. For many readers this sounded new. For people familiar with Denmark and much of Europe, it is mostly a return to older exam formats rather than a radical invention. Oral exams, thesis defenses, and mixed written plus spoken assessment are already normal in many universities and in some school systems long before university. Several Danes pushed back on the article’s framing and said Danish practice varies a lot by subject, level, and class size. Thesis defenses are common, but not every university course is oral, and many existing oral exams are used because they test understanding well, not just because of AI.
The strongest consensus was that live questioning reveals whether someone can actually explain, connect, and apply ideas. People kept coming back to the same practical test. If you can teach it, answer follow-ups, or walk someone through your reasoning, you probably know it. If you only produced clean prose, that signal is now badly degraded. A lot of commenters said this was true even before LLMs. Teaching assistants, code reviewers, and instructors all described the same pattern. Real comprehension becomes obvious fast when you ask someone to explain what they wrote or built.
But the comments were just as clear that oral defense is not a universal fix. It scales badly. A professor can grade piles of written work asynchronously, delegate some of it, or blind-grade it. A live defense is one student at a time. It also introduces examiner bias, inconsistency, and accessibility issues for students with severe anxiety, speech or hearing issues, or weaker fluency in the language of instruction. The useful conclusion was not "make everything oral." It was that institutions are going to split assessments by what they are trying to certify. High-trust, high-stakes work will move back toward in-person exams, handwritten work, short live defenses, or spot checks. Lower-stakes assignments will be treated more like practice.
A second theme ran underneath the whole conversation. AI has not just made cheating easier. It has weakened trust in digital output itself. Several commenters argued that once any take-home text can be plausibly machine-generated, schools have to verify understanding in person or accept that the credential says less. Others argued the opposite direction and said education should adapt to AI instead of retreating into older formats. That view got less support, but it sharpened the real divide. The question is no longer whether students will use AI. It is whether a school is certifying polished outputs or certifying a person’s ability to think, explain, and perform when the scaffolding is removed.
If you run education, training, or hiring, stop treating polished written output as proof of understanding. Add some form of live explanation or follow-up questioning where stakes justify it, but design around the cost, bias, and accessibility problems that come with oral evaluation.
Mostly positive about oral defense as a practical way to test real understanding in the age of AI, with a strong undercurrent of "this is not new." Support came from people who had gone through oral exams and felt they were fair and revealing. The main reservations were about scalability, examiner bias, and the risk of unfairly punishing students with serious anxiety or other disabilities.
Key insights
01
The article overstates how novel this is
Danish assessment already mixes oral and written formats across schools and universities, and the deciding factors are usually subject area and logistics, not some clean academic-level rule. That matters because the policy is better understood as expanding an established verification method into AI-vulnerable coursework, not inventing a new model from scratch.
Do not read this as a wholesale redesign of Danish education. If you borrow the idea, start by identifying where you already have live assessment practices and extend them selectively.
Having to teach the material, answer tangents, and survive rephrased questions is what breaks shallow memorization and exposes AI dependency. Several examples from TA work, homework help, and exam prep made the same point. Preparing an explanation can itself become the learning process, which means AI-generated notes are not fatal if the student has to metabolize them into real understanding.
Use live follow-up as the validation layer, not as a ban on AI tools. You can let people draft with AI and still require them to show they own the reasoning.
They catch ghostwritten or AI-written work well, but they also add subjectivity that written exams can reduce through blind grading and consistent rubrics. Examiners drift over a long day. Personal style, ethnicity, sex, fluency, and confidence can leak into the grade. Oral assessment is not more objective. It is objective about different things.
If you add oral defenses, narrow the rubric and keep them short. Use them to verify authorship and depth, then anchor final grading with artifacts that can be reviewed consistently.
The workable version is not a medieval-style defense for every piece of work. It is selective oral verification layered onto take-home assignments, especially in smaller courses or online programs. One short challenge per student or a quiz based on take-home work captures much of the benefit without forcing every class into a one-by-one bottleneck.
For large cohorts, use random or rotating defenses instead of universal ones. You will preserve deterrence and get a clearer signal without blowing up staffing costs.
Asking students to submit their AI chats sounds principled, but it also invites a second layer of curation where the most conscientious students do extra theater for the grader. That shifts effort from learning to managing appearances. The critique landed because it points at a broader trap in AI policy. Documentation requirements can become easier to game than the original assignment.
Be careful with authenticity paperwork. If your goal is to verify understanding, a short live check is often cleaner than asking students to prove innocence through process logs.
A lot of people who started with severe public-speaking fear said routine oral exams made the experience survivable and eventually normal, especially when examiners were calm and supportive. The key nuance is not that anxiety is fake. It is that structured repetition, good examiner behavior, and low-drama questioning can turn oral defense into training as well as measurement.
If speaking under pressure is part of what you want graduates to do, introduce it early and often in low-stakes settings. Do not make the first exposure the final gate.
The sharpest dissent said schools are clinging to old assessment rituals instead of redesigning for a world where LLMs are cheap, capable collaborators. On this view, the valuable skill is not manually producing every artifact but using AI to reason, critique, and extend your own work. Forcing handwritten or oral-only verification can preserve credential integrity while missing the bigger opportunity to train students for the tools they will actually use.
Do not let anti-cheating policy become anti-tool policy. Separate foundational courses where unaided reasoning matters from applied work where AI fluency is part of the job.
The hardline "just get over it" stance ran into commenters who pointed out that stage fright and test anxiety can be serious medical problems, not character flaws. Exposure helps some people, but overwhelming exposure can backfire. Medication, extra time, and carefully designed accommodations were presented as legitimate ways to measure knowledge without turning the exam into a stress endurance contest.
If you use oral defenses, build an accommodation path in from the start. Otherwise you risk measuring stress response more than subject mastery.
A smaller but credible line of argument said oral defense does not need to exclude AI at all. If a student uses an LLM to prepare talking points, challenge their thinking, or generate a defense plan, and can then explain the material under questioning, that is still learning. The problem is not tool use. It is unowned output.
Write policies that target unverifiable authorship, not AI use in general. In many cases the right rule is "use any tool you want, but be ready to defend every claim live."
Massachusetts curriculum frameworks
Linked in a side argument that education can be standardized at state scale, so scale alone is not a decisive objection.
mezha.net