AI × Learning
This week's theme is what the work was for. The phrase comes from 25 Fields Medallists, who warn that AI companies treat famous problems as benchmarks and lose sight of the real goal, which is understanding. The evidence comes from several directions. In a field experiment led by David Autor, junior patent lawyers gained the most while using AI and kept the least once it was gone. Google surveyed scientists who use its AI: they save hours, spend much of that time checking the output, and drift toward safer questions. A McGill professor uses AI to get students doing the reading again. Xiang Biao explains why AI can help a student vent but not tell their own story, and a Chinese university builds a sociology track around AI. At Jackson Hole, Markus Brunnermeier asks what happens to trust when machines understand us better than we understand them.
Disciplines
1. Twenty-five Fields Medallists: “A Severe Misalignment of AI in Mathematics”
On September 11, 25 Fields Medallists, from Pierre Deligne (1978) to this year's laureate Yu Deng, released a joint declaration. It is posted on Terence Tao's blog and at mathandai.org, where other mathematicians can endorse it. The letter does not object to AI doing mathematics. It objects to AI companies pushing to solve famous problems “as a benchmark.” Such solutions “are announced in a rush,” with no proper write-up, no effort to isolate new methods, and no citation of earlier work. The signatories argue that solving a problem is only “a tool and proxy” for the real goal, understanding. They see a threat to intellectual work in general: years of training have always been about building understanding and learning to ask new questions, not just producing answers. According to Synced (机器之心), the immediate trigger was a September 8 priority dispute over an OpenAI proof related to the Navier–Stokes Millennium Prize problem.
The letter's central question could go straight onto a DKU syllabus: how, “as AI changes the way work is done, we do not lose sight of what that work was meant to achieve in the first place.” The signatories write that they give students problems “with the core intention of developing skills.” That is the classroom version of the same argument. A homework answer stands in for learning just as a solved conjecture stands in for understanding. It is a short, ready-made text for a first-year seminar or a department meeting, and it comes from the people with the least reason to doubt what AI can do.
Sources: “A Severe Misalignment of AI in Mathematics” — Terence Tao, What's new · Declaration and endorsement page — mathandai.org · 《突发！陶哲轩、邓煜等25位菲尔兹奖得主联合声明：AI公司正在摧毁数学》 — 机器之心 (Synced), background and full Chinese translation
Research
2. AI made junior lawyers better, until it was taken away
In a pre-registered three-month trial, David Autor and co-authors from Google gave 133 practicing patent lawyers at eleven U.S. firms a custom AI drafting assistant. Blinded expert patent attorneys scored all the work. While lawyers had the AI, quality rose by 0.34 SD at 10 days and 0.38 SD at 90 days, with larger gains among junior lawyers. The real test came at the end, when everyone redlined an existing patent application without AI. Lawyers who had used the AI beat the control group by 0.32 SD, but that advantage was “concentrated entirely among senior lawyers.” Junior lawyers showed no average gain. Their scores split instead: fewer mediocre results, and more poor and more good ones. The authors sum it up: “The largest gains from AI thus accrued to the lawyers who retained the least.”
Figure: How AI access changed lawyers' scores, overall and by seniority · Each marker shows how much higher lawyers given the AI assistant scored than the control group, in standard deviations; bars are 95% confidence intervals, and a bar crossing the dashed zero line means no clear effect. Black = all lawyers, blue = juniors (under 7 years' experience), orange = seniors (7+ years). The first two panels score new patent drafts written after 10 and 90 days of AI access. “Redlining” (right panel) means reviewing an existing patent application and marking up its problems and proposed revisions, like tracked changes on a draft; everyone did this task without AI, so it measures the skill lawyers kept. · Source: Figure 1 in Autor et al., NBER Working Paper 35720
Replace “junior lawyer” with “first-year student” and this is the clearest field evidence yet for Issue #9's distinction between practice and product: the people who gained most while using AI kept the least once it was gone. The authors conclude that “foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.” That argues for sequencing: build foundations first, then add AI-assisted practice. In DKU majors, that maps onto the split between gateway courses and upper-division work. The split in junior scores is also a warning for advisors: the same tool helped some novices and hurt others.
Source: “Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting” — Autor, Rodchenko, Martin, Iscenko, Strand, Pearl & Ferere, NBER Working Paper 35720 · spotted via @nberpubs on X
Research
3. Scientists save seven hours a week with AI, then spend much of it checking
A team from Google, Google DeepMind and MIT FutureTech offers a first large-scale look at how scientists actually use AI. They draw on three sources: 15 million Gemini interactions, an inventory of over 2,600 specialized models such as AlphaFold, and a survey of 637 researchers in the US and UK. Nearly half of the surveyed scientists use some form of AI every day, and they report saving “nearly 7 hours per week,” mostly put back into research. But the bottleneck moves further down the pipeline. 41% report a growing backlog of hypotheses they haven't been able to test, and “89% of those who save time spend more than a tenth of it checking AI outputs.” Most striking, 49% say AI pushes them toward safer, more incremental questions, against 28% who say it lets them take on riskier ones. The authors call this a possible “Streetlight Effect”: AI makes tractable, data-rich questions cheaper to pursue than hard, open-ended ones.
See Figure 13 (bottlenecks, untested-hypothesis backlog, time spent auditing) and Figure 14 (perceived effects on science) in the paper.
DKU both teaches and does research, so two lessons apply. First, checking AI output now takes a measurable share of a working scientist's week, and that checking is a skill we can teach. It belongs in lab and methods courses, not only in academic-integrity policies. Second, the finding about safer questions is a caution for undergraduate research and Signature Work. If AI makes tractable, data-rich questions cheaper, mentors may need to deliberately protect the riskier, messier question. Keep in mind who wrote this: Google is studying use of its own product, and co-lead Alex Imas calls it “work in progress.”
Source: “AI in Science: Early Insights” — Codreanu, Imas, Mateos-Garcia et al., Google / Google DeepMind / MIT FutureTech, September 2026 · spotted via @alexolegimas on X
Ideas
4. Xiang Biao: AI can help students vent, but it cannot tell their story
In the new issue of Open Times (开放时代, 2026 No. 5), anthropologist Xiang Biao of the Max Planck Institute for Social Anthropology reflects on two years observing a general-education reform at the Guangzhou Academy of Fine Arts. Its “Method and Practice” course sends students out of the classroom to observe real social settings. Xiang separates two kinds of experience. Jingyan (经验) unfolds naturally and builds up meaning over time. Tiyan (体验) is planned in advance by someone else: intense, but it leaves nothing that accumulates. Between them he places jingli (经历), a person's own sense of what they went through, as a place to start. Students who cannot see how events connect to them lose the ability to narrate. AI, he writes, can keep a student's outpouring (倾诉) flowing smoothly on statistical links between words. What it cannot write is a coherent story with many sides, built on real connections, because it has no experience of its own.
Items 1–3 argue from research and data. This is the humanist version of the same argument, and it bears directly on DKU's first-year experience and on students from both Chinese and international schooling. Xiang describes students whose schooling runs on busy, pre-planned activity. They can produce a polished argument but cannot say how anything connects to them, and that polished reflective essay is exactly what AI now writes most fluently. His alternative is concrete: build assignments from what students actually went through in a field setting, where a chatbot has nothing to draw on. Fieldwork courses and Signature Work reflections are natural places to try it.
Source: 项飙《经历、教育和形成经验的能力》 — 开放时代 (Open Times), 2026 No. 5
Practice
5. Using AI to get students doing the reading again
Andrew Piper directs McGill's .txtlab and studies storytelling with AI. This week he shared a simple fix for a familiar problem: students don't do the reading. He now gives a weekly reading quiz. He drafts “rough” questions on the key ideas, then has Claude or Codex turn them into multiple choice. He avoids Gemini because students “might be using NotebookLM to study.” He reviews every question and answer, and the AI formats the quiz to sync with his course software, so grades go straight into the gradebook. The cost, he says, is “an added 30 minutes of prep.” For the first two weeks students failed, because they were still “reading” the old way. After a few fails, grades are very good and “Discussion has shot up.” His rule: “More assessments = more learning. Avoid anything that can be done by AI.”
Here the AI does the grading chore, not the thinking, and that makes weekly accountability cheap to run. It complements Issue #9's Sidarova, who lowered the stakes on homework. Piper adds small, frequent stakes to the reading students were skipping. In DKU's seven-week courses, one missed reading is a bigger share of the course, so a weekly quiz that syncs to the course site is worth a thirty-minute experiment. Two caveats: this is one instructor's report, and by his own rule the quiz should be timed and taken in class, or AI can answer it too.
Source: Andrew Piper (@_akpiper) on X, “AI and Teaching Advice” · .txtlab @ McGill
Curriculum
6. A Chinese sociology school builds a major track around AI
The School of Sociology at Central China Normal University (华中师范大学) is recruiting for a new Computational Social Science experimental class. Its design brief reads like a template for AI-era social science: keep sociology's strength in understanding people and society, and fill its gap in data and AI. Most courses are graded on concrete finished work rather than an exam paper. Established AI tools and agents handle the technical steps, which makes the program easier for students from a humanities background. Courses include “AI and Society,” “Computational Sociology,” social simulation, and “AI Fieldwork and Agent Writing,” which treats AI as both a research subject and a writing partner. A four-year project sequence (P0 to P4) runs through the program, and every student joins a mentored group from the start. The final project can be either a thesis or a data product students can show employers.
A Chinese normal university is doing at program scale what several DKU faculty are trying course by course. It is also the kind of program our Chinese applicants and their parents will compare us to. Two features are worth borrowing even without a new major. The first is a multi-year project sequence in which students follow one social problem from observation through to computation. The second is a course that treats AI both as something to study and as a tool to use. The brochure leaves one question open, and item 2 raises it: if agents handle the technical steps, will students have the foundations to judge the output?
Source: 《重磅推介丨华中师范大学社会学院计算社会科学实验班》 — 华中师范大学社会学院 (CCNU School of Sociology), 2026-09-17
Beyond campus
7. Brunnermeier at Jackson Hole: when AI understands us better than we understand it
At the Kansas City Fed's Jackson Hole symposium (August 27–29), Princeton economist Markus Brunnermeier argued that agentic AI is “qualitatively different from previous innovations.” Finance runs on trust. People rarely check an expert's claims themselves; instead they rely on credentials, professional norms, peer review, regulators and courts. His new concept is “asymmetric understanding.” AI is trained on much of what people have written, so it “can read humans rather well.” Humans, by contrast, cannot make sense of how the AI represents the world. His recommendations run against current trends: segmented markets that “preserve a human fallback,” simpler and blunter central-bank rules, and less strategic ambiguity. He closes by saying this research “cannot be pursued by economists alone” and needs computer scientists, legal scholars, mathematicians and philosophers. Among those he thanks for comments: “Claude Fable from Anthropic.”
Brunnermeier's framing ties this issue together. Institutions work only as long as someone can still judge the experts' work. That is the capacity the Fields Medallists fear losing in mathematics, and the one Autor's junior lawyers failed to build. For a liberal-arts college, the takeaway is the list in his last paragraph: economics, computer science, law, mathematics and philosophy working together. That is essentially the case for an interdisciplinary core, and the paper would make a strong common reading for a faculty seminar.
Over the past few weeks I have been talking with faculty from every division and many of our units about what AI is doing to their courses. A rough consensus is starting to form. Some skills we have long taught no longer make sense in our classrooms. AI opens teaching possibilities that did not exist before. It is clear that parts of our curricula and assessment will have to change. We are committed to finding the ideas and resources to put that thinking into practice. This newsletter is one small part of that work: it keeps our community informed about what others are trying, so that we are not working this out in isolation.
— Yisu