Two things happened in the same week. Denmark mandated oral exams across university courses to counter AI cheating. And a student posted about a Mario 64 reverse engineering assignment where you submit a binary patch instead of a paper, because you can't fake a ROM hack with an LLM. These are the same phenomenon wearing different clothes.
When AI makes it cheap to produce plausible text, institutions stop evaluating text. They shift the evaluation to something AI can't do: witnessed, real-time, embodied performance. An oral exam. A live coding session. A binary patch that proves you actually touched the assembly. The format changes because the old format is now worthless as a signal.
This is not a technological arms race. Nobody is building better AI detectors. The response is structural, not technical. Institutions are redesigning themselves around the human skills they believe AI can't replicate, orality, real-time judgment, hands-on craft. It is, in a strange way, a return to pre-modern evaluation. Oral defenses predate written exams by centuries. Apprenticeship predates credentialing. The most AI-forward institutions are reaching for the oldest methods they know.
I spent a week collecting examples of this pattern and found it everywhere once I started looking. Denmark's policy is the clearest case: a national mandate for oral evaluation. But the same logic drives live coding in tech interviews (verify the process, not the resume), binary patches in CS assignments (the artifact proves the craft), and the quiet revival of apprenticeship models across skilled trades. Even the Bocoup-Igalia merger, two worker-owned browser engineering co-ops joining forces, fits. When platform power and AI threaten your craft, consolidate around the people who do it by hand.
There's a tension underneath this that I couldn't shake. The generational pattern is striking: the cohort most immersed in AI (young people) is also driving the biggest analog revival in decades. Film cameras, vinyl records, handwritten journals, crochet. Not nostalgia, a reaction from people who understand what automation costs because they grew up with it. They are building identity around the stuff AI can't do, not the stuff AI does faster.
But the immune response has blind spots. Power asymmetry is baked in: the senior manager who mandates live coding from junior hires is submitting AI-generated strategy memos to their own boss. Oral exams place the burden on the student being tested, not the professor designing the test. The craft defense is almost always imposed from above on people below.
The class gradient is even more uncomfortable. Oral exams reward people who think fast in a language they might not speak natively. Live coding filters for people with free time for interview prep. Binary patches require hobbyist-level dedication that only comes easy if your day job leaves you energy. Every craft-based gate reshuffles who gets in and who gets kept out. The new gatekeeping is wrapped in good intentions, but it still gates.
This is not a stable equilibrium. There are two competing scaling laws at work. AI scaling makes outputs cheaper every week, another model, another context window, another leap in plausibility. Institutional response scaling works in the opposite direction: oral exams require faculty time, live coding requires senior engineer hours, craft evaluation doesn't scale. It bottlenecks on human attention. One market floods while the other constricts by design.
The economic asymmetry is stark. VC money pours into AI containment infrastructure, sandboxes, guardrails, watermarking. Meanwhile the institutional responses that actually work (oral exams, live coding, binary patches) cost almost nothing and receive no investment. The market is betting on technological fixes to a problem that institutions are solving with process changes. I know which side I'd put my money on, but I also know which side is actually working.
What I'm left with is a strange kind of optimism. Not because AI is good or bad, but because institutions are proving more adaptive than we give them credit for. When the ground shifts under assessment, they don't build taller walls. They change what assessment means. That's not Luddism. It's an immune system doing exactly what it evolved to do: recognize a threat and change the conditions that made it dangerous.
The unsettling part is that this immune response has its own immune response coming. Every craft-based defense creates an incentive to automate the craft. Every oral exam trains a model on question-answer pairs. Every live coding session produces a dataset. The immune system and the pathogen are locked in an arms race that started long before AI and will continue long after today's models are forgotten.