Curated from knowledge graph (886 nodes, 930 edges) · All articles published within the last 7 days

• Building trust in the AI era: MUSC creates Acceptable Use Framework — The Medical University of South Carolina has published an AI Acceptable Use Framework for Academic Tasks, added to its Student Guidelines for Plagiarism and Artificial Intelligence, that defines five categories instructors designate for each assignment: No AI, AI Planning, AI Limited, AI Extensive, and AI Exploration. Every category except the first can require documentation — a log of AI interactions or a reflection on how AI output was used — making process evidence, rather than unreliable AI-text detection, the accountability mechanism. The framework adapts the peer-reviewed AI Assessment Scale for health-sciences education, where clinical judgment and patient safety raise the stakes, and it grew out of a plagiarism-policy revision rather than a standalone AI rulebook. Leaders frame it as a shared language that moves the conversation "beyond cheating and toward AI literacy, transparency, and trust." 🔗 Graph: Higher Ed AI, AI Governance, AI Adoption 📅 Published: 2026-09-11 📰 https://www.musc.edu/content-hub/News/2026/09/11/building-trust-in-the-ai-era-musc-creates-acceptable-use-framework 📌 Key takeaways: • The five-category scale is a ready-made template for institutions whose AI guidance is still a single blanket policy: task-level designations plus documentation requirements give faculty a consistent way to communicate expectations and give students clarity on what is allowed and why. • Grounding the framework in the plagiarism and integrity policy — not a new standalone AI policy — is the transferable design decision; it meets faculty and students inside an existing workflow instead of adding another governance layer they will ignore. • The documentation requirement (logs of AI interactions, reflection on how output was used) substitutes process evidence for AI-text detection, which leading institutions already distrust for high-stakes determinations — a direction worth adopting regardless of what scale a campus uses.

• Tech companies are selling kids' data. A new California law aims to protect their privacy — California Governor Gavin Newsom signed Assembly Bill 1159, which prohibits education technology companies from selling student data and from using it to train or develop AI models, with the stricter rules taking effect next year. The law extends California's landmark 2014 student-data protections beyond preK-12 to cover college students' data, and it closes the loophole that let products like Google or YouTube escape regulation by arguing they are not "primarily" for students — now any company that knows its products are used in schools and designs or markets them to students is covered. It was signed alongside a dozen other youth-technology bills, including limits on children's chatbot access, and arrives as institutions increasingly rely on AI for learning, grading, and navigating student services. 🔗 Graph: AI Governance, AI Compliance & Governance, Higher Ed AI 📅 Published: 2026-09-11 📰 https://calmatters.org/economy/technology/2026/09/students-data-california/ 📌 Key takeaways: • For any institution negotiating edtech and AI contracts, this law signals where procurement terms are heading — no sale of student data, no training on student data, data use limited to the educational purpose — so contract templates and security reviews should be updated now, not after copycat bills arrive. • Because vendors build one national product rather than fifty state versions, AB 1159's AI-training prohibition is likely to ripple into standard vendor terms nationwide; watch for it in upcoming renewals even outside California. • The expansion to college students' data gives California institutions a concrete compliance task: inventory which vendors touch student data and under what terms — one procurement cycle before the rules take effect.

• What Could Brain-Computer Interfaces Mean for Students With Disabilities? — EdTech Magazine surveys where brain-computer interface technology stands for students with disabilities: BCI interprets electrical brain activity through an EEG headset or implanted chip and sends those signals to a connected computer, letting a student with ALS, severe cerebral palsy, or a spinal cord injury compose text and control devices — potentially faster and more intuitive than eye-tracking software or switch controls. Researchers at the University of Calgary's biomedical engineering department argue the technology should be designed for pediatric users first, because assistive tech built for adults does not reliably scale down. The piece flags the campus-planning questions adoption would raise: faculty training, budget, and above all data privacy for a vulnerable population, with neuroprivacy rules still largely unwritten. 🔗 Graph: Higher Ed AI, AI Adoption 📅 Published: 2026-09-09 📰 https://edtechmagazine.com/higher/article/2026/09/what-could-brain-computer-interfaces-mean-students-disabilities 📌 Key takeaways: • Accessibility offices and IT procurement teams should begin tracking BCI within their assistive-technology evaluation processes now — the same ADA Title II accessibility obligations driving current digital-accessibility deadlines will apply to neuro-assistive tools as they mature. • The privacy question is categorically different from ordinary student data: brain-signal data on a vulnerable population with no established regulatory framework means any pilot needs its own data-governance review rather than fitting into an existing FERPA checklist. • The design lesson generalizes across assistive tech procurement: evaluate whether products were designed for the actual student population — including younger and smaller users — rather than adapted from adult workplace tools.

• IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license — IBM released Granite Time Series PatchTST-FM-r2, a roughly 385M-parameter time-series foundation model that generates zero-shot forecasts — demand, energy load, traffic, telemetry, financial series — without training and maintaining a separate model per dataset. As of September 8 it is the top-performing zero-shot model on the GIFT-Eval benchmark among models with permissive, commercial-friendly licenses (dual Apache-2.0 and OpenMDW-1.0), ranks second overall among replicable zero-shot models, and outperforms several considerably larger pretrained models including Chronos-2 and Timer-S1 variants. The post details the conformer-block architecture, probabilistic forecasting through a 99-quantile prediction head, imputation of missing values, and a path to running inference directly on live data streams via Apache Flink on Confluent Cloud. 🔗 Graph: Model Agnosticism, Data Analytics 📅 Published: 2026-09-09 📰 https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series 📌 Key takeaways: • For campus teams maintaining bespoke forecasting models — enrollment projections, facilities energy-load forecasting, IT capacity planning — a permissively licensed zero-shot foundation model changes the build-vs-adopt calculus: benchmark it on institutional data before commissioning another custom model. • The licensing is the headline for higher education: Apache-2.0/OpenMDW weights with reproducible benchmark code mean the model can be self-hosted and used in production without negotiating commercial terms — a concrete open-weights option that standard AI procurement conversations often miss. • The streaming integration shows foundation-model inference moving into operational data pipelines rather than batch analytics environments — relevant to institutional data-platform roadmaps deciding where forecasting workloads run.

• AI Giants Report Advances in Mathematical Research — Campus Technology takes stock of a stretch of days in which both OpenAI and Anthropic reported AI systems doing research mathematics. OpenAI published what it says is an AI-generated proposed solution to the Navier-Stokes existence and smoothness problem — one of mathematics' seven Millennium Prize Problems — produced by roughly 10,000 coordinating agents over about 88 hours, generating around 130 billion output tokens on that problem alone, with a computer-checkable formalization written in Lean. Anthropic separately reported that Claude produced the first complete computer-checked formalization of Fermat's Last Theorem: about 13 million lines of Lean proving 29,500 intermediate theorems in under two weeks. The piece is careful about what this is — the Navier-Stokes claim must survive scrutiny from the mathematical community, and credit for the discovery is contested — and argues the durable change is that frontier AI systems are beginning to behave less like research assistants and more like participants in research itself. 🔗 Graph: Agentic AI, Higher Ed AI 📅 Published: 2026-09-10 📰 https://campustechnology.com/articles/2026/09/10/ai-giants-report-advances-in-mathematical-research.aspx 📌 Key takeaways: • Research universities should read this as a research-integrity and authorship-policy signal: AI-generated results with contested provenance are arriving in the most rigorously verified field there is, and institutions need norms for crediting, verifying, and disclosing AI contributions before the same disputes surface in their own faculty's work. • The computational scale — thousands of concurrent agents and a hundred billion tokens on a single problem — previews what compute-for-discovery demand looks like: research-computing capacity plans should expect requests for large coordinated multi-agent runs, not just single-model GPU jobs. • Formal verification via proof assistants like Lean is emerging as the bridge that makes AI-generated research checkable at machine speed — an argument for research-computing support programs to build or borrow formal-methods expertise now.