Colleagues — thank you for the replies to issue zero. A fuller issue this week, six items, because the material earned it. The through-line: judgment is becoming the curriculum. Basic skills are wobbling, institutions are writing their AI rules — and from Seoul to Zhihu, the sharpest thinkers are converging on the same answer about what stays human: evaluating answers, and asking better questions.
Over 1,800 math and science lecturers across the University of California signed an open letter about first-years arriving with “severe preparation deficits” — at UC San Diego, students entering with below-high-school math skills grew nearly thirtyfold in five years. The Economist pairs the anecdotes with OECD adult-skills data: across rich countries, roughly 8% of tertiary students score no better in literacy than a ten-year-old, a share that has more than doubled in a decade; in America it's one in seven. Layered on top: grade inflation (79% of Yale grades were A or A− in 2022–23) and AI — a Berkeley researcher's analysis of 500,000 grades at a Texas university found the share of A's in AI-friendly courses up 13 points since ChatGPT launched. His line is the quote of the week: “You can't ban a tool that you are also expected to teach.”
Why it matters for us: our students mostly arrive through a different (and mathematically rigorous) pipeline, so the preparation story isn't automatically ours — but the grade-inflation and AI-assessment dynamics absolutely are. The article's quiet lesson: institutions that kept externally anchored, invigilated assessment saw the least drift. Small classes let us anchor grades in work we actually watch happen — worth protecting deliberately, not by inertia.
Read: The Economist (June 25)
The Digital Education Council's 2026 global survey — 27,284 students and 18,114 faculty in 35 countries — finds 88% of students now use AI in their learning and 77% of faculty in their teaching, up 16 points in a year. What hasn't kept pace is guidance: 57% of students say their assessments come with inadequate AI instructions, and only 29% think their instructors are equipped to guide them — though 64% of faculty have done AI literacy training. Early dependence signals: 22% find it harder to work without AI; 19% say they're retaining less. Most striking is the regional split: faculty intent to use AI runs 89–94% across Latin America, APAC, and EMEA, while in the US and Canada it fell from 76% to 67% — and 43% of students there would support an institution-wide ban.
Why it matters for us: that chart runs straight through our hallways — DKU sits geographically in the most AI-optimistic region while many of us trained in the most skeptical one. Worth noticing which instinct we each default to. And the students' actual complaint isn't too much or too little AI; it's that 57% get no clear guidance per assessment. A one-paragraph AI statement on every fall syllabus is the cheapest fix in this entire survey.
Read: DEC survey report · ETIH summary
Last issue covered UChicago Law's AI strategy; Inside Higher Ed now shows it's a pattern among top law schools. Berkeley Law's policy, effective this summer, prohibits students by default from using AI in “conceptualizing, outlining, drafting, revising, translating, or editing any work submitted for credit” — while instructors teaching AI fluency can opt out. Texas Law's dean wants class time spent in “sustained and rigorous dialogue,” screens closed. And UChicago's design is worth a second look: first-year students in legal writing must write without AI, then use it for research, revision, and oral-argument prep — with professor feedback on both. All this while the ABA reports 58% of practicing lawyers already use generative AI in daily tasks.
Why it matters for us: “default off, deliberately on” is a policy shape a liberal arts college can actually administer — clearer than course-by-course improvisation. And UChicago's two-track writing assignment (unaided draft, then supervised AI pass, feedback on each) is directly copyable in any writing-intensive seminar we run, no committee required.
Read: Inside Higher Ed · Berkeley Law policy (PDF) · UChicago Law memo
Arvind Narayanan (Princeton, co-author of AI as Normal Technology) gave the ICML keynote in Seoul and posted annotated slides — the most readable version of his framework yet. The argument: AI keeps getting better at verifiable tasks, so “purely technical skills get devalued” and human effort shifts from building to evaluating. His software example: the “execute” layer (coding) is compressing, but the “decide” layer (requirements, specification, judgment) and the “deliver” layer (understanding work deeply enough to be accountable for it) are not. His metaphor: when engines replaced rowing, the work became steering and navigation. He's skeptical of imminent recursive self-improvement, and invokes the ATM story — automation grew bank employment — against the lump-of-labor fallacy.
Why it matters for us: if the durable human work is deciding, evaluating, and being accountable, then a liberal arts education is oddly well-positioned — a seminar is evaluation training. The practical move: make students grade AI output in our disciplines — critique its history essay, audit its statistics — because that's the skill his framework says the market will keep paying for.
Read: annotated slides + transcript · spotted via @random_walker on X
The same debate, on the Chinese internet. A Zhihu roundtable question — “In the AI era, which is scarcer: asking good questions or judging good answers?” (18,000+ views, 172 answers) — drew a sharp answer from 赵泠, an answerer with 890,000+ followers: asking was always scarcer, AI or no AI. Many good answers verify themselves in practice; a good question “turns vague confusion into a testable conflict, draws a boundary around the unknown, and points toward verification.” Both skills, she notes, rest on the same foundation — knowing many true answers, plus systematic doubt, especially of your own conclusions. And one under-appreciated AI effect: chatbots remove the social cost of asking “dumb” questions that keeps students silent in classrooms.
Why it matters for us: read alongside Narayanan (item 4): evaluation and question-formulation are two names for the same scarce judgment — and both are teachable. This conversation happening in Chinese matters for our campus: it's material for bilingual classroom discussion, and a reminder that our students' intellectual world spans both internets. The “no-embarrassment questions” point is also an argument for sanctioned AI use in office-hours prep.
Read: 赵泠's answer on Zhihu (Chinese)
A widely shared engineering essay from Catalin Voss's team, building a real-time AI tutor that teaches math and reading to children aged 4–9. The premise: “for AI to actually teach a five-year-old, pedagogy must be baked into the engineering.” A two-second pause loses a child's attention, so they abandoned the standard agent loop, stream teaching actions as they're generated, and run a planner asynchronously in the gaps while the child thinks — the same gaps where human teachers make their judgment calls (challenge, or let them succeed?). When a question is on screen, the model is handed scaffolding moves instead of answers. And because “a child can't unhear anything a model gets wrong,” a safety classifier gates every action before it plays.
Why it matters for us: we're not building tutors, but we're buying them — and this piece is a checklist of questions to ask any vendor: what pedagogy is baked in, does it scaffold or answer, what happens when the model errs? Note the design's core: the tutor is engineered to withhold answers. Most tools our students use are engineered to do the opposite.
Read: @CatalinVoss's essay on X
Six items is a lot — tell me if this length works or if three is the right ceiling. If your syllabus is getting an AI statement this fall, or your students are already grading chatbot output, I'd like to hear about it. We are collecting examples for AI experimentation at DKU.
— Yisu