Universities are embracing AI: will students get smarter or stop thinking? (www.nature.com)

Universities worldwide are rapidly embedding generative-AI into campus life — from Tsinghua’s AI admission assistant and its three-layer architecture that routes queries across ~30 models and curated “knowledge engines,” to Ohio State’s compulsory AI literacy courses and the University of Sydney’s insistence on in-person tests to curb outsourcing. Adoption is explosive (one survey reported 86% of students regularly using AI in 2024; Anthropic’s analysis of one million student chats found heavy STEM use), and big vendors are pushing campus deals (OpenAI’s ChatGPT Edu, Google’s student tools). That speed is forcing fragmented policy responses: many faculties are experimenting locally, leaving students exposed to inconsistent rules and prompting debates about data protection, vendor lock-in, and the ethics/environmental costs of model deployment. For the AI/ML community the moment is both opportunity and challenge. Practical efforts like Tsinghua’s model-agnostic middleware and Sydney’s Cogniti show how ensembles, knowledge retrieval and domain-specific agents can mitigate hallucination and tailor feedback. Early trials are mixed: a Harvard RCT found custom AI tutors sped learning, but other studies (and Tsinghua’s preliminary work) suggest short-term gains may not translate to long-term retention and could encourage cognitive offloading. The takeaways: build robust evaluation (longitudinal studies), improve model grounding and interpretability, design pedagogy-aware tools, and prioritize privacy and auditability when deploying LLMs in education.