01
The proxy collapsed
For decades the essay, the problem set, and the exam stood in for cognition because producing them was expensive. The cost was the evidence: if the artifact existed, someone had probably done the thinking. That was never a measurement of learning — it was a proxy that held only while production was slow.
AI removed the cost, and the artifact stopped proving much about the person who submitted it. In the faculty discussion I coded for the original essay, institutional energy pointed overwhelmingly at detection and enforcement rather than redesign: an arms race against an improving generator, run on behalf of assignments that rarely specified what cognitive work they were supposed to develop in the first place.
02
The same grammar as my research
A submitted essay is a benchmark output. A grade promotes that output to a claim about a person's capability. Promotion without adjudication — no provenance for how the artifact was produced, no named cognitive operation it was supposed to demonstrate, no verifier with contact to the real skill — is the same failure I catalogue in agent evaluation. A gameable verifier does not become trustworthy by being defended harder.
The redesigns that work follow the same checklist as honest evaluation anywhere: name the operation being measured; make the process visible as lineage rather than one polished artifact; measure in patterns across time instead of single events; and move verification to a layer with reality contact — an oral defense, an external audience, a derivation done in the room. Education is not adjacent to my research. It is a special case of it.
03
Going back anyway
The 2025 version of this note read like a verdict on school. It was a verdict on assessment. The parts of a university that AI cannot cheapen — a cohort, laboratories, structured depth, sustained access to people who know more than you — became more valuable, not less. So I enrolled: first year at XJTLU, beginning autumn 2026.
Holding the critique and enrolling is not a contradiction; the critique is an operating manual. Treat coursework as access to problems and people. Keep the real verification external and public — shipped systems, accepted papers, a notebook that carries its own provenance — and never let the credential substitute for the record. I also no longer expect exposure to force fast redesign. My dated position, not a forecast dressed as fact: incumbent Western institutions will metabolize this on something closer to a generational clock than a semester one, and waiting for them is not a plan.
04
Still open
What I watch for is an institution willing to publish its verifiers: to state, per assignment, which cognitive operation it develops and how anyone would know it happened. A syllabus that can explain its own zeros would collapse the distance between education and evaluation entirely — which is where this notebook has been pointing all along.