Colleagues — this is a pilot of a short weekly digest: three items on AI and learning, each with a link and a note on why it might matter for us. It takes five minutes to read. Tell me if it's useful (or not). This week's accidental theme: assessment is where the AI pressure concentrates first.
Brown economist Roberto Serrano gave his Welfare Economics class a take-home midterm this spring — a first in nearly two decades of teaching the course. The average score came back at 96%, against a historical range of 65–80%, on an exam he says was harder than usual. When he moved the final back in person, the average collapsed to 48.6% — the lowest he has ever recorded. Eighteen students dropped, nine skipped the final, nineteen failed. Serrano calls the university's response “meek”; Brown's own new GenAI committee report finds three-quarters of surveyed faculty worried about AI-enabled cheating (matching AAC&U's 2025 national figure) — while recommending faculty “de-emphasize punishment.”
Why it matters for us: the gap between those two bars is the clearest single picture I've seen of why unproctored, unredesigned assessment can no longer carry summative weight. Our advantage over an 86-person lecture: classes small enough for oral exams, in-class writing, and assessment that watches the process, not just the product.
Read: Inside Higher Ed · Brown's GenAI committee report (PDF) · spotted via @paulg on X
Two days ago the University of Chicago Law School published an AI strategy statement, Rethinking Legal Education in the AI Era, effective this fall. Three pillars: AI-resilient pedagogy and assessment, elevating the “essential human” skills, and teaching responsible AI use. Concretely: a pilot ban on electronic devices in all core 1L classes, a required oral discussion for every substantial research paper, and expanded hands-on AI work in the clinics. The framing from committee chair William Hubbard is worth stealing: AI shortcuts are “beneficial in the professional context… but very, very damaging in the educational context, when the whole point is to do things the hard way — because that's how you learn.”
Why it matters for us: this is the most coherent institutional policy I've seen — it separates where AI must be absent (foundational learning, exams) from where it must be taught (practice-like settings). Notice that its two big moves — device-free discussion and oral defenses — are things a liberal arts college already does at small scale. Useful vocabulary for our own policy conversation this fall.
Read: UChicago Law announcement · full statement at law.uchicago.edu/ai-hub · spotted via @DavidDecosimo on X
A new paper (Chen, Zhao & Cohan) asks not whether LLM-generated research ideas are good, but how far they sit from what human researchers actually produce. Method: reverse-engineer the prior work behind high-quality papers, hand the same starting material to LLMs, and compare. The finding: a consistent distributional gap. LLM ideas cluster around “bridge-like” opportunities and synthesis methods — connect X to Y — while human ideas spread far more broadly across ways of framing gaps and building contributions. The models produce reasonable ideas; the range is what's missing.
Why it matters for us: a concrete, evidence-based answer to students (and colleagues) using AI to brainstorm projects and theses: it will reliably give you an idea, but the distinctive framing — the research taste — still comes from the human. That's teachable, and worth saying out loud in methods courses.
Read: arXiv:2607.01233 · spotted via @Xudong07452910 on X
That's it for issue zero. If something here connects to your own classroom — or you're seeing the opposite — hit reply. We are collecting examples for AI experimentation at DKU.
— Yisu