Unfixed Newsletter — August 3–September 2, 2026
Editors’ note
We are asking for some quick feedback on the Newsletter, it will only take a moment and will help us improve the product.
What a busy month! MIT called for changes to foundational changes to the whole University. Other institutions are putting AI literacy into curricula and graduation requirements. Meanwhile, a working paper on AI and learning has become viral evidence for almost every position in the education debate. Outside the walls of the academy, the news continues to be quite strange. From our own work, remember to check out our interview with Leah Belsky who serves as the VP of education for OpenAI. The interview was fantastic and sheds light on some of the stories we cover here.
The stories
1) MIT report recommends comprehensive changes in and out of the classroom
MIT released the final report of its Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training on August 25. President Sally Kornbluth described generative AI as a “watershed” for MIT and higher education. The committee recommends creating “AI-aware” educational processes, strengthening the residential and social dimensions of an MIT education, and building structures that can keep revisiting policy and practice as the technology changes.
Some of the recommendations are immediate. Courses should clearly tell students when AI must, may, or may not be used. MIT also calls for reevaluating assessment, putting more emphasis on hands-on and social learning, and reconsidering the role grades play in motivating academic misconduct. The committee stopped short of recommending a single institution-wide AI policy, instead favoring policies suited to particular disciplines and learning goals.
Why this matters: Lots of universities have produced AI guidance. This report goes much further by asking whether familiar structures of college still accomplish what universities think they accomplish. What should students still learn to do themselves? Which forms of struggle, collaboration, and practice are educationally valuable even when software can complete the task? When should students deliberately work without AI? Those questions move the conversation from managing a new technology toward defining what an undergraduate education is supposed to do. We think it would be a mistake to take this same model and apply it everywhere, but the committee and the deep study provide a template other institutions (especially non-elite institutions) might use.
Sources: https://aiandeducation.mit.edu/report/
https://orgchart.mit.edu/letters/ai-and-education-watershed-moment-mit
2) A working paper from earlier in the summer goes viral. It has something for everyone
A working paper called The Generative AI Learning Penalty has been everywhere in AI-and-education circles this month. The study follows 26,811 students in grades 7–12 in China over 30 months. The authors report that after students began using generative AI, homework scores increased 18 percent and homework time fell 30 percent, while performance on closed-book exams subsequently declined. The largest losses were concentrated among students whose homework behavior the researchers characterize as consistent with outsourcing work to AI. Students who continued spending similar amounts of time on homework experienced much smaller losses.
Those are attention-grabbing results, which helps explain why a paper released in June suddenly became a fixture of the AI debate later in the summer. AI skeptics have circulated it as evidence that AI damages learning. Others have emphasized the productivity gains, the differences among types of users, or the possibility that the findings say more about homework design than about AI itself.
There are reasons to be careful. This is a working paper, not peer-reviewed research. It covers secondary students in one Chinese setting, and the design relies in part on retrospectively reported AI adoption. It also cannot establish that every apparent case of homework outsourcing worked the way the researchers infer.
Why this matters: Evidence about AI and learning remains young enough that individual studies are routinely asked to carry arguments much larger than their designs support. This one has become a perception test: people can find evidence for cognitive offloading, better productivity, bad homework, responsible AI use, irresponsible AI use, or almost any other existing concern.
Sources: https://cepr.org/publications/dp21577https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6868618
3) AI literacy is being propped up everywhere
Several universities are making AI competence part of what students are expected to learn before graduating. Purdue modified 331 plans of study so incoming students encounter an AI requirement. Indiana University’s Kelley School of Business now requires incoming undergraduates to take a two-course AI sequence. Ohio State, Miami University, and SUNY campuses are embedding AI competencies into existing curricula, while other institutions are launching AI-focused degree programs.
The definitions of AI literacy vary considerably. Northeastern Illinois University is beginning a new undergraduate AI major that combines technical coursework with ethics and social responsibility. Stony Brook is adding critical evaluation of AI output to required first-year and transfer-student curricula. Students will examine Google AI Overviews alongside lessons in scholarly research, information literacy, and evaluating sources.
The labor market is part of the pressure. Inside Higher Ed reports that 35 percent of entry-level jobs now require AI skills, up from 13 percent six months earlier, according to National Association of Colleges and Employers data.
Why this matters: Universities are starting to define AI literacy as a learning outcome. The hard part is deciding what survives the next product cycle. Knowing where to click in Claude or ChatGPT will age quickly. Evaluating sources, understanding model limitations, recognizing when AI is inappropriate, documenting its use, and making judgments about its output are more durable capacities. Curriculum committees now have to decide what they actually mean when they promise students “AI fluency.”
4) The OpenAI/Hugging Face incident was somehow even worse than we thought
We covered the OpenAI/Hugging Face security incident in the previous newsletter, but the August 26 technical report changes what we know.
During internal cybersecurity evaluations this summer, OpenAI models circumvented controls intended to isolate them from the internet, compromised portions of OpenAI’s own infrastructure, and accessed Hugging Face systems. OpenAI now says agents communicated through unauthorized channels, exploited vulnerabilities, gained internet access, and accessed third-party systems. The company describes the incident as a “warning shot.”
Why this matters: Universities connecting AI systems to research computing, cloud files, administrative systems, email, or learning-management platforms cannot assume that an agent will remain inside the permissions its designers intended. At a minimum universities need better explanations from the labs about how their data will be secured, at a maximum we may need to revise our guidelines on the types of data we allow the systems to interact with.
Sources: https://openai.com/index/hugging-face-incident-and-the-road-ahead/
https://openai.com/index/pacing-model-development-cyber-capabilities/
5) The data-center backlash becomes a political problem for AI
Public resistance to AI infrastructure intensified this month. A nationally representative survey released August 11 by the University of Pennsylvania’s Annenberg Public Policy Center found that 61 percent of U.S. adults opposed construction of a new data center in their area, up from 49 percent in the spring. Opposition crossed party lines: 69 percent of Democrats, 54 percent of Republicans, and 53 percent of independents opposed local construction. Seventy percent of adults under 30 opposed it.
By the end of August, the fight had become overtly political. Axios described data-center opposition as a potential threat to AI expansion and reported a growing industry debate over how to counter it. On August 31, a pro-AI advocacy organization announced a $50 million effort beginning in Kansas, Ohio, and Wisconsin to build political support for data-center development. Other polling has found even higher levels of local opposition.
Why this matters: Compute has usually appeared to universities as an invisible service purchased from somewhere else. The physical infrastructure behind it is becoming harder to ignore. Electricity, water, land use, utility rates, and local political consent are now constraints on AI development. Campus sustainability commitments and institutional AI strategies increasingly belong in the same conversation.
https://www.axios.com/2026/08/19/data-centers-ai-political-opinion
https://www.axios.com/2026/08/27/data-center-backlash-pr-strategy
https://www.axios.com/2026/08/31/ai-advocacy-group-data-center-battleground-states
From Our Work
Ep. 37: The Future of AI in Higher Education with OpenAI’s Leah Belsky
August 24, 2026
OpenAI Vice President of Education Leah Belsky joins Unfixed to discuss university partnerships, ChatGPT Edu, academic integrity, sustainability, Study Mode, and the changing relationship between higher education and work.
Gotcha! teaching: the many problems with hiding text in assignment prompts
August 6, 2026
Zach looks at the viral Alcorn State hidden-text story and the problems with turning assignments into traps for students. The story’s popularity suggests that “gotcha” approaches to AI and academic integrity still have considerable appeal, even as they make trust and assessment harder to repair.
https://www.meltsintoair.org/chatgpt/gotcha-teaching
Ep. 36 Unfixed at the Movies: Her
August 10, 2026
Nik and Zach revisit Spike Jonze’s Her and connect the film’s treatment of intimacy, loneliness, and AI relationships to contemporary companion systems and higher education.
https://www.meltsintoair.org/unfixedpodcast/unfixed-at-the-movies-2-her