AI × Learning
This week's theme is the skill question: who still builds skills when AI can do the work? A Trends in Cognitive Sciences review names the mechanism — offload the practice and you forfeit the skill, though not inevitably. A new Chinese graduate tells Sanlian Lifeweek what four years of exactly that felt like from the inside, and Nature collects the assessment redesigns professors are trying in response. The labor market sends a double signal: Google will now interview engineers on how well they validate AI output, while The Economist counts a net AI jobs boom even as the routine entry rungs thin. In Sydney, two thousand university staff strike over who decides how AI enters their work. And PISA 2025, released Tuesday, measures what happens when the oldest academic skill of all — sustained reading — stops being practiced: the 15-year-olds it tested reach our campus in three years.
Research
1. “Is AI making us stupid?” A review sorts panic from mechanism
A Science & Society review in Trends in Cognitive Sciences by Trent Cash, Megan Kelly, Brooke Macnamara and Evan Risko asks its title question and answers with a useful distinction: specific skills, which are built and maintained through practice, versus basic cognitive abilities like working memory and attention. Offloading the practice to AI “almost certainly” compromises the skill. The sharpest evidence is a high-school math experiment: co-author of that study Alp Sungu explains that “with AI access, students scored 48% higher on practice problems,” but afterward “the same students scored 17% lower on exams than those who never had AI at all” — while a restricted, tutor-style chatbot erased the penalty. Basic abilities, by contrast, look “stubbornly resistant” to change. The erosion is real, conditional, and largely a design problem: staying in the “cognitive loop” preserves, and may even boost, learning.
This gives us cleaner vocabulary than “AI rots your brain.” The question for any DKU assignment becomes: is this the practice through which a skill forms, or the product of a skill already formed? Practice must be protected from offloading (or de-staked, see item 3); products can tolerate AI. And the tutor-mode finding is actionable now — course-configured assistants that hint rather than solve behaved like tutoring, not outsourcing.
Sources: “Is AI making us stupid?” — Trends in Cognitive Sciences · PsyPost's summary of the review · spotted via @jayvanbavel on X
Students
2. Four years inside the machine: a graduate's testimony
Sanlian Lifeweek runs an oral account (by reporter Duan Nongyu) of “Xiao Sa,” who entered one of Shanghai's top universities as ChatGPT appeared and graduated this June, based on his viral Zhihu essay on the “sense of tearing” AI brought to his undergraduate years. The details are unsparing: the online exam where AI use was rampant gave him his only A of three years; by senior year his method was slides-plus-past-papers into the chatbot the night before. For an American leadership prerequisite, he had Gemini summarize the lectures, added his experiences, then had ChatGPT scrub the AI traces — and was publicly praised for independent thinking while flagged classmates got zeros. At an internship, a chip manual he should have understood on day one took him five full days. His verdict: AI didn't ruin his degree, it exposed how much of it was already formulaic labor. His question: when the tool can bypass the learning process, why would a student choose the slower, more painful path — and how does a university make that path necessary again?
This is the student's-eye view of the outsourcing statistics we cited in Issues #6 and #8, written by exactly the kind of student our first-years study alongside — and it belongs in orientation and advising conversations, not just syllabus policy. Note what he says works: the classmate who outscored him also used AI, but fed it her own experiences and thinking. His proposed fix — assess whether the core ideas are the student's and whether they can explain their choices — is precisely the CMU-style interview from Issue #8, endorsed this time from the other side of the desk.
Source: 《挤进最热门赛道的名校生，毕业后为何主动离开？》 — 三联生活周刊 (Sanlian Lifeweek), oral account via Duan Nongyu
Practice
3. Nature's tour of the redesigned assessment
Nature's careers team asked educators worldwide how they now assess. The scale of the problem first: in a 2026 Higher Education Policy Institute survey of 1,054 UK undergraduates, roughly 94% said they use generative AI to help with assessed work and 12% inserted AI text directly; a May study of more than 95,000 students at 20 US universities estimated 9% used AI knowing it broke the rules. The responses are a menu. Nicholas Mattei has students critique three LLMs' summaries of sources they found. Daniel Silver replaced the Adam Smith essay with AI agents trading in a simulated marketplace, students analyzing the transcripts. Gianmarc Grazioli awards 70% of points for working code and 30% for a one-to-one oral exam on specific lines. Thea Goldring caps grades at C if authorship can't be demonstrated in conversation. Natalia Sidarova made homework pass/fail and moved the stakes to a calculators-only exam: “we've never had as many students achieving full-score grades in the exams as we did this year.” Mattei's honest caveat: “I don't think anybody has a ‘right’ answer right now.”
Almost everything here gets easier, not harder, at seminar scale: oral defenses of code, conversation-based grading, revision-history submissions. Sidarova's move is the one to study — it operationalizes item 1 by removing grades from practice work so outsourcing it stops paying, then letting the exam measure what formed. That directly answers the mechanism behind Issue #8's “50 percent problem.”
Source: “From Trojan horses to AI-proof exams: how professors are tackling students' AI use” — Nature
Hiring
4. Google will interview engineers with Gemini at their side
A Reddit screenshot declaring “LeetCode grind is dead” went viral this week; the verified story underneath (an internal document reported by Business Insider) is nearly as notable. Google is piloting interviews in which early- and mid-career software-engineering candidates use the company's Gemini assistant during the “code comprehension” round — analyzing an existing codebase, finding bugs, improving performance. Per the document: “Interviewers will evaluate AI fluency, including prompt engineering, output validation, and debugging skills.” The pilot starts in the second half of the year with select US teams; the behavioral “Googleyness” round gains a technical-design conversation about the candidate's own past work, and one junior technical round becomes open-ended engineering problems. Recruiting VP Brian Ong says the aim is interviews “more reflective of how our teams are operating in the AI era” — at a company where, it says, 75% of new code is AI-generated. Meta began piloting AI-enabled coding interviews late last year.
Hiring is assessment, and Google's redesign converges on the same moves as items 2 and 3: work with the AI on a real artifact, then demonstrate understanding in conversation. A course that asks students to review, debug, and justify AI-produced work is no longer a pedagogical experiment — it is interview preparation. Worth sharing with our CS and data-science students before they over-invest in the old prep.
Sources: “Google Is Testing a New Rule That Could Transform Job Interviews” — Entrepreneur, on Business Insider's reporting · spotted via @coolcoder56 on X
Economy
5. The Economist: an AI jobs boom, for now
Against the displacement narrative, The Economist estimates AI has created about 1 million US jobs, against roughly 200,000 layoffs attributed to AI since mid-2023 — in a labor market where employers shed about 1.7 million workers in a typical month. August payrolls rose 162,000, “far above expectations”; unemployment sits at 4.1%, “lower than in almost 90% of months over the past half-century,” and the unemployment gap for 20–24-year-olds is near a multi-decade low. The boom is physical as much as digital: data-centre construction spending rose 60% in a year, and Indeed finds installation and maintenance jobs at data centres advertise wages about 40% above comparable work. The disruption is real but specific: since January 2023, employment has fallen roughly 10% for customer-service workers and 15% for administrative assistants, and the chart's May 2023–25 window shows the same divide, data scientists at the top and data-entry keyers at the bottom.
Figure: “Differential calculus” — US employment, % change May 2023–25, selected sectors · Source: Bureau of Labor Statistics · chart 3 in The Economist, “The jobs apocalypse is postponed. An AI jobs boom is here”
Issue #8's FT item showed teenagers choosing majors by a labor market several years stale; this is the corrective in the other direction — the apocalypse framing is also stale. For advising, both halves matter: net creation is real but skewed toward people who can work with and around AI, and the thinning is concentrated in exactly the routine entry rungs new graduates once used to climb. Neither panic nor complacency; portfolios of skill, not bets on titles.
Sources: “The jobs apocalypse is postponed. An AI jobs boom is here” — The Economist, 2026-09-04 · spotted via @rdomenechv on X
Labor
6. Sydney: two thousand staff strike over AI protections
About 2,000 members of Australia's largest National Tertiary Education Union branch walked off the job at the University of Sydney on September 2, with about 500 staffing picket lines from 7am; “almost all” classes moved online. AI protections sit at the center of enterprise bargaining, alongside workload and job security — and trust: in one school within arts and social sciences, 0% of surveyed staff agreed with the statement “I trust the University of Sydney to act in the best interests of staff, students and the broader community.” Branch president Peter Chen: “Higher education is the canary in the coalmine… Where we go on AI … will travel through the rest of the economy.” The university calls the action “premature” and prefers implementing AI commitments “through relevant policies, instead of in the Enterprise Agreement.” The Guardian notes an AI chatbot is already helping teach online-only psychology classes at nearby Macquarie.
Our AI conversations have mostly been about students and integrity; Sydney is a reminder that the staff side — who decides how AI enters teaching work, and whether commitments live in policy or in contract — arrives eventually everywhere. A small institution's advantage is that consultation can be genuine and early rather than remedial. Geography professor Kurt Iveson's formulation travels well as a design principle: protections that don't outlaw AI but ensure “the tech is serving our purpose, rather than us serving the tech.”
Data
7. PISA 2025: the reading decline, measured
The OECD released PISA 2025 on Tuesday: 760,000 15-year-olds in 91 countries and economies, and the lowest OECD-average scores ever recorded in reading and mathematics. Reading fell 28 points across OECD countries between 2015 and 2025, “about one and a half years of learning,” 14 of them in the last cycle; the peak was 2012, well before the pandemic or ChatGPT. The diagnosis is attention as much as skill: text length is the strongest driver of the decline, and the share of “hasty readers” who answer fast and wrong almost doubled since 2018. Schleicher's foreword: “we do not become fit by watching sports but by doing sports.” East Asia still owns the top of the table but is not exempt: every top system with a 2022 baseline (mainland China did not participate that year) scored lower in reading than in 2022, from 7 points in Chinese Taipei to 20 in Hong Kong; over the decade only Chinese Taipei's trend is positive, with Singapore and Japan near flat. Japan topped the OECD in all three subjects for the first time, with the lowest AI-for-schoolwork use of any member. An Ottawa teacher gives CBC the classroom version: on a math test, “students were asking me things that were directly written in the question.”
Figure: “PISA 2025 Reveals Steep Decline in Math & Reading Skills” — OECD average PISA scores of 15-year-olds, 2000–2025 · Source: OECD, PISA 2025 · Chart: Statista (CC BY-ND) · the OECD's own version is Figure I.2.6 (p. 95) in Volume I
These are the students who arrive in 2028. The AI data echo this issue's through-line: students who use chatbots to summarise texts or draft assignments score lower; those who use AI to learn and are taught to assess its output score slightly higher, and that teaching reaches advantaged students first. PISA director Andreas Schleicher's phrase for tools that spare us thinking: they are “hollowing out the capabilities that we need.” Two implications for a reading-intensive college. First, Qiu Luming's point from the Dushu roundtable (Issue #8) now has numbers: the habit of sustained reading is eroding before students reach us, including in the systems that top the table, so first-year courses have to build it, not assume it. Second, the “hasty reader” is the profile item 1 warned about, formed years before any DKU assignment; protecting time for slow reading is not nostalgia but remediation.
Sources: PISA 2025 Results (Volume I): Future-Ready Students — OECD, 2026-09-08 · OECD press release · “Canadian teens score well on OECD tests, but sharp drops in reading, math raise concerns” — Jessica Wong, CBC News · “Japan Tops OECD in Reading, Science and Math, But Average Reading Score Declines” — The Japan News (Yomiuri) · Schleicher interview — Euronews
If you would like to discuss ideas related to AI and teaching, please reach out to me.
— Yisu