AI × Learning
This week's theme is the learning penalty. The strongest evidence yet that AI-completed homework quietly erodes what students actually learn arrives just as institutions place their bets in opposite directions: Harvard dismantles its writing center while Purdue makes AI competency a graduation requirement. In between stand the students themselves — some of whom, given the chance, decline to use AI at all — and the employers and economists reminding us that in an AI-saturated economy, the scarce skill is still human judgment.
Research
1. Better homework, worse exams: hard numbers on the learning penalty
The Economist's data team analyzes what may be the first large causal study of generative AI in schools. David Stromberg of Stockholm University, with Victor Lei and Wu Yanhui of the University of Hong Kong, tracked 27,000 pupils aged 12–18 in China, where about 80% had taken up models such as Doubao and DeepSeek; the remaining 20% formed the control group. After six months, AI users' average homework score had risen by 18% across all subjects, and time per assignment had fallen from 64 minutes to 45. But come exam time, the same students scored 20% below classmates who hadn't called on AI. Homework scores once predicted exam performance; the paper — titled "The Generative AI Learning Penalty" — shows the shortcut severing that link, even as a Chegg survey finds 80% of rich-world undergraduates using AI, with recent polls at 94% in Britain and 93% in Germany.
Chart: The Economist · original article (gift link)
Chart: The Economist
The mechanism here is formative assessment breaking down: when graded homework can be outsourced, it stops telling either the student or the instructor what has been learned. For our small classes the fix is within reach — shift credit from take-home polish to in-class writing, oral checks, and drafts discussed face-to-face — but it needs deciding before syllabi lock, not after the first exam shock.
Sources: "Does AI stop children from learning?" — The Economist (gift link) · "The Generative AI Learning Penalty" — Stromberg, Lei & Wu, SSRN
News
2. Harvard closes its writing center — and hands everyone else a permission slip
Tyler Austin Harper reports in The Atlantic on the backlash to Harvard's dismantling of its writing center and the layoff of its director, Jane Rosenzweig — a champion of human writing in the age of AI, who wrote in 2024 that "the hard work of reading, writing, and thinking is not a problem to be solved." An open letter from current and former tutors drew more than 100 signatories. The context stings: Harvard gives undergraduates premium access to Gemini, ChatGPT Edu, and Claude, while a Harvard Crimson survey found the average student used AI to complete 34.5 percent of their homework. The dean of undergraduate education calls the closure "largely a problem of terminology"; Harper's worry is mimetic — if Harvard doesn't need a writing center to be elite, trustees everywhere will ask why their college does.
Harper's "permission structure" argument is aimed squarely at institutions like ours: budget logic that starts at the apex arrives at small liberal-arts colleges with extra force. It is worth saying now, on the record, that one-on-one writing support is core academic infrastructure at DKU — the human counterweight to exactly the outsourcing the Economist study measures — and not administrative overhead waiting for a spreadsheet.
Source: "Why the Closing of Harvard's Writing Center Matters" — The Atlantic
Teaching
3. When students are the conscientious objectors
Philosophers Carla Fehr (Waterloo) and Jennifer Saul flip the usual integrity story: what happens when the syllabus requires AI and students refuse? Fehr designed a research-methods assignment in which groups had to use generative AI and reflect on the experience. The first question from the floor was "What if we don't want to use GenAI?" She added a conscientious-objector option — do the work traditionally and defend the choice in writing — and seven of eight groups took it; the eighth tried AI and preferred working without it. Students' essays cited cognitive dependence, environmental costs, bias, and academic integrity. The authors note Pew finds 61 percent of Americans under 30 worry AI will erode creative thinking, and urge instructors who require AI to at least recognize principled refusal exists.
As DKU courses start building AI use into assignments, this is the missing design question: is the learning objective the AI itself, or the thinking it's meant to support? Where it's the latter, an opt-out with a written rationale costs little and turns refusal into analysis. Seminar-scale classes like ours can run both tracks side by side — and the objectors' reflection papers may be the best AI-literacy artifacts of all.
Source: "Students as Conscientious Objectors to Generative AI" — Inside Higher Ed
Policy
4. Purdue's AI graduation requirement meets its first cohort
The opposite bet from Harvard's: this fall's incoming class at Purdue is the first covered by the "AI working competency" graduation requirement its trustees approved in December — billed as first in the nation, applying to all undergraduates at West Lafayette and Indianapolis starting with new students in fall 2026. The design is more interesting than the headline: rather than a single AI course, the provost and the deans of every college are to create and continually update discipline-specific criteria, with each college convening a standing industry advisory board on employers' AI-competency needs to feed an annual curriculum refresh. President Mung Chiang's framing — Purdue "must lean in and lean forward" — makes it the clearest institutional counter-position to the caution running through the rest of this issue.
If AI expectations reach our curriculum, Purdue's mechanism is the part worth stealing: discipline-specific standards, owned by faculty, revised annually — not a bolt-on tool-training course. A liberal-arts version would define competency as judgment: knowing what AI does to one's own thinking (item 1) and when not to use it (item 3). That's a standard our seminars can actually teach, and one worth articulating before an external template arrives.
Employers
5. Gillian Tett: in finance, the scarce skill is still human
In the Financial Times, Gillian Tett argues that even in finance — the industry racing hardest to automate — humans still matter more than AI, and that "recruiting digital natives with critical thinking skills is going to be crucial." The pattern she describes: firms are augmenting staff rather than replacing them. AI excels at processing market data, flagging anomalies, and surfacing patterns, freeing professionals for the strategic thinking clients actually pay for; but consequential decisions run through shifting markets, regulation, and geopolitics that fall outside any training set — the 2008 crisis and the COVID shock both caught models flat-footed. The professionals who thrive treat AI as a tool, not a substitute for thinking, and know which decisions can be partially automated and which cannot.
This is the answer to the student (or parent) who asks why a liberal-arts degree in an AI economy: the hiring premium Tett describes — critical thinking layered on digital fluency — is precisely what happens when our graduates add AI competence to seminar habits of skeptical reading and judgment under uncertainty. Career services and admissions should be making this argument with employer voices like this one, not just our own.
Source: "Humans still matter more than AI in finance" — Gillian Tett, Financial Times (gift link)
Context
6. How big is the AI economy our students are graduating into?
Exponential View's "The State of the AI Economy" (Azeem Azhar and colleagues) attempts the first bottom-up census of AI demand — "capturing every real dollar of customer demand, no double-counting." Their headline: $110 billion of generative-AI revenue over the past twelve months, a $175 billion annualized run rate on the latest month, growing roughly three times more rapidly than the mobile or internet waves — yet, as they put it, "still small enough to be early." Two details deserve faculty attention: demand is strikingly price-elastic ("every 10% price cut leads to 12-18% more tokens in use, so the total spend still rises"), and revenues so far "just about clear the depreciation expense" on the infrastructure — a real economy, but not yet a comfortable one.
Chart: Exponential View, “The State of the AI Economy”
The report is an interactive chart deck with a downloadable PDF — browse the original figures.
Numbers for the classroom and the curriculum committee alike: large enough that no field our students enter will be untouched, early enough that this cohort will shape the norms rather than inherit them. It also arms us against both hype and dismissal — the growth is real, and so is the strain of merely covering the infrastructure bill. A useful shared reference when divisions debate what "AI in the major" should mean.
Sources: "The State of the AI Economy" — Exponential View · companion essay
We are collecting examples for AI experimentation at DKU.
— Yisu