- Dartmouth Survey Reveals What Students Think About AI — Dartmouth's inaugural multi-year AI Survey, released at the "AI at 70" conference marking 70 years since the 1956 Dartmouth workshop that founded the field, surveyed more than 1,500 Dartmouth students and 1,000 U.S. college students. The headline finding is a perception gap: students largely say they use AI to learn, but assume their peers are cutting corners — 11% report using AI on problem sets while believing 49% of peers do, and 8% admit pasting text with little editing while estimating 50% of peers do. 70% of U.S. students say take-home work has become unfair due to AI, 79% taught themselves their AI skills with little high-school guidance, and 84% believe AI writing use can hurt writing-skill development.
📌 Key takeaways:
- The perception gap matters for policy: students moderate their own use but overestimate everyone else's, which fuels pressure to cheat (39% feel pushed to use AI because peers do). Institutions should publish real usage data alongside their AI policies — visible norms shape behavior more than rules do.
- With 79% of students self-teaching AI skills before or during college, the biggest institutional gap is structured guidance, not access. Course-level AI expectations and AI-fluency programming will land on an audience that already has habits to correct.
- The survey is designed to repeat annually, making it a rare longitudinal benchmark campuses can use to track whether their own interventions move student attitudes.
🏛️ UCSD angle: The campus AI program tracks real adoption across 73K+ users behind its gateway — the kind of institution-level usage ground truth that this survey suggests students rarely see, and which can anchor local norms instead of peer-guessing.
- Purdue is first U.S. university to deploy Google TPU AI infrastructure — Purdue University is deploying Google Tensor Processing Units for faculty and student research in collaboration with Google Public Sector, Cirrascale, and DDN, becoming the first U.S. academic institution to offer the custom AI accelerators as part of a campus computing program. The addition represents a 30% increase in Purdue's AI capacity, integrated into its long-running community cluster high-performance computing program, with initial researchers on the system in January 2027. The university maintains institutional control over data, security, and governance while running the hardware in its own Mathematical Sciences Building Data Center.
📌 Key takeaways:
- The GPU shortage is pushing research universities toward accelerator diversity — TPUs, not just NVIDIA, inside campus data centers. Institutions planning AI capacity expansions should evaluate non-GPU accelerators as both a capacity and a negotiating lever.
- Purdue's model pairs vendor hardware with institutional operation: the university runs the system inside its own community cluster business model rather than renting capacity in a vendor cloud, keeping data sovereignty and governance on campus.
- The deal was announced alongside a joint Purdue-Google AI summit on physical AI, signaling that vendor-university research partnerships increasingly come bundled with summits, curriculum ties, and workforce narratives — worth factoring into procurement strategy.
🏛️ UCSD angle: UC San Diego's AI platform similarly pairs institutional control with infrastructure choice — TritonGPT runs on SDSC-hosted on-prem compute behind a model-agnostic gateway, the same sovereignty-plus-flexibility pattern Purdue is extending to TPUs.
- Thousands of "ghost students" use AI to steal millions in student loans, DOJ says — The Department of Justice announced its first broad federal crackdown on "ghost students" — fraudsters who buy the identities of Americans (mostly high school graduates who never attended college) to enroll in universities and community colleges and collect federal student aid. DOJ charged five individuals and seized over $750,000 in cryptocurrency from a network associated with Kenyan nationals, and the Department of Education has identified over 830,000 potentially fraudulent federal student aid accounts. Officials said AI is used throughout the scheme: filling out FAFSA applications, completing coursework, and even faking attendance in virtual classrooms to keep aid flowing semester after semester.
📌 Key takeaways:
- Campus security and registrar teams should brace for operational fallout: 830,000 flagged aid accounts means enrollment-verification, identity-proofing, and online-attendance checks are now fraud battlegrounds, and legitimate applicants may face friction as screening ramps up.
- AI-generated coursework and fake virtual attendance defeat the legacy signals (logins, submissions, participation) that institutions use to verify a real student exists — distance programs need stronger identity assurance than LMS telemetry.
- The cross-border dimension (an international seizure in Kenya) means this is not a perimeter problem institutions can solve alone; plan for coordination with federal investigators and expect reporting obligations to grow.
- UC San Diego Launches Center for Principled Agentic Intelligence — UC San Diego is launching the Center for Principled Agentic Intelligence (CPAi), an interdisciplinary partnership between the Halıcıoğlu School of Data Science and Computing and the School of Arts and Humanities to study AI alignment, human oversight, and independent evaluation of agentic AI systems. The center will develop an open-source testbed for assessing agentic AI, and launches with $1 million in support over three years plus $250,000 in shared computing resources.
📌 Key takeaways:
- Agentic AI is arriving on campuses faster than the governance to evaluate it; a dedicated center focused on alignment, oversight, and independent evaluation gives institutions a template for treating agent assessment as a research discipline rather than a vendor assurance question.
- The open-source agentic testbed is the piece other institutions should watch — a shared, independent evaluation harness would let any campus benchmark agent behavior without building the capability alone.
- The structure pairs computing with arts and humanities, a signal that agent governance is being framed as a socio-technical problem, not just a technical one.
🏛️ UCSD angle: This is the campus — CPAi complements the existing TritonAI program, extending the campus AI platform's operational governance work with a research arm for principled agentic evaluation.
- What a Talking Haunted House Teaches About Real-Time Voice Agents — An editorial analysis of NVIDIA's MonsterHouse project — a talking, grumpy garage door built for Halloween — argues the real value of the stunt is its architecture: a complete, local, real-time voice agent combining Nemotron 3 ASR, the 4GB Nemotron 3.5 Bolt E4B model, Magpie TTS Zeroshot, and Audio2Face-3D, all coordinated on a single workstation. The durable lessons for anyone building voice agents: separate model responsibilities, keep durable state in application logic, cache predictable moments, and treat fallbacks and field testing as core design work.
📌 Key takeaways:
- Production-grade real-time voice agents are now feasible on a single workstation with small open models — relevant for campuses piloting kiosks, accessibility tools, or multilingual student-support lines without shipping audio to a commercial API.
- The hardest problems are architectural, not model quality: latency budgeting, state management across turns, and graceful fallbacks when a component is slow or down are what separate demos from deployable services.
- The pipeline pattern (listen → reason → speak → animate, with explicit encounter state) is the same shape campus AI teams will need for any always-on conversational service.