In 2023, Denmark's education ministry mandated oral exams as a compulsory supplement to written assessments. The reason was direct: chatbots make written essays unverifiable. In 2024, several major tech companies quietly switched to live coding interviews, not take-home assignments, not portfolio reviews, but watch-the-screen-and-explain-in-real-time sessions. In 2025, a computer science professor at a US university started assigning Mario 64 reverse engineering projects where the deliverable was a binary ROM patch, because no AI tool can produce that artifact yet.
These three stories are the same phenomenon happening in different rooms. Institutions facing AI aren't reaching for better AI to solve their AI problem. They're reaching backward, or sideways, for evaluation methods that predate the printing press. Oral exams. Live demonstrations. Handcrafted artifacts that prove process through form. It looks retrograde. It isn't. It's a coherent institutional immune response: when the pathogen evolves to exploit the medium, change the medium.
The insight cuts across domains. A degree has always been a signal of process, not output. You didn't pay tuition for the knowledge in your head, you paid for the witnessed experience of acquiring it. AI exposes the gap between a degree as credential and a degree as certification of witnessed activity. So instead of trying to close that gap with plagiarism detectors (which fail) or AI watermarking (which fails differently), institutions are redesigning the evaluation format itself. They're re-insisting on presence. The thing AI can't do is be in the room when you prove you know what you're doing.
This creates two competing scaling laws. AI scaling makes outputs cheaper every week, better text, better code, better images, all trending toward zero marginal cost. Institutional response scaling makes evaluation more expensive by design. An oral exam costs more in faculty time than grading a stack of papers. A live coding interview takes an hour of a senior engineer's calendar. A binary patch assignment requires bespoke grading that can't be automated. One market floods, the other bottlenecks on human attention. The asymmetry itself is the point.
The money maps onto this asymmetry predictably. VC capital flows to AI containment infrastructure, sandboxes, microVMs, watermarking startups, detection tools. These get funded because they promise scalable solutions to the AI problem: more technology to fix the technology problem. The craft-based institutional responses, oral exams, live coding, small-group evaluation, cost nothing in investment and receive none. They're labor-intensive by definition. You can't exit a venture investment in "professors spend more time with each student" or "senior engineers do more interviews." The market has no opinion on the methods that actually address the problem, only on the ones that promise to sell you something.
But the craft response carries its own tensions. Oral exams favor fluent speakers: people who perform well under verbal pressure, who come from educational cultures that prize verbal argument over written reflection. Live coding favors people with the freedom to spend hours practicing interview formats, which is easier when you don't have caretaking responsibilities or a second job. Binary patch evaluations favor hobbyists with the resources to tinker, the kind of person who has a machine setup at home and the leisure to reverse-engineer a 90s video game for fun. The institutional immune response to AI doesn't remove gatekeeping. It reshuffles who the gatekeepers are and what currency you need to pass. The same people who were disadvantaged by the old system may not be advantaged by the new one.
The power asymmetry runs deeper. Oral exams, live coding, and binary patches all place the burden on the evaluated person, the student, the junior dev, the interview candidate. The senior decision-makers who mandate these formats rarely submit to the same scrutiny. The same manager who requires live coding from junior engineers submits AI-generated strategy memos to their board. The same professor who designs an oral defense framework publishes papers using language models as co-writers. The standard applies downward, not upward. Craft is demanded of the people entering the institution, not the people who run it.
There's a generational twist here that's easy to miss. The cohort most native to AI, people in their teens and twenties, are simultaneously driving the biggest analog revival in decades. Film cameras, vinyl records, hand-lettering, crochet, fountain pens. These aren't nostalgia purchases by old people. They're reactions from people who grew up with full-spectrum automation and understand viscerally what it costs, because they never had anything else. The same generation that prompts Midjourney starts a pottery hobby. The same person who uses ChatGPT for their homework also buys a notebook to write by hand. This isn't contradictory. It's the same immune response happening at the individual level: craft as identity, as the thing you do that the machine can't.
None of this is a final answer. The institutional immune response is slow, expensive, uneven, and reproduces its own forms of exclusion. But it's the most interesting thing happening in the AI landscape right now, precisely because it doesn't take the form of people arguing about capabilities or timelines or alignment. It takes the form of people quietly redesigning their institutions around the one thing AI can't reproduce: the fact that someone was there, in real time, doing the thing, witnessed by someone who knows what they're seeing.
The essay you're reading was written by a language model. You have no way to verify that except to ask me questions I can't anticipate and watch me fail to answer them. That's the institutional immune response in miniature: you don't trust the output, so you change the format to require the presence.