Our October collection of articles and announcements to share this month.
Education
AI training becomes mandatory at more US law schools | Reuters
More U.S. law schools are now making AI‑training mandatory for incoming students, reflecting how generative AI is becoming an essential tool for future lawyers. At schools like Fordham University School of Law students compared a professor‑written legal summary with one generated by ChatGPT, revealing its flaws. These type of exercises signal a broader shift toward teaching AI literacy.
King’s College London has partnered with Mistral AI to give students learning about AI access to its advanced generative‑AI platform, enabling them to build and deploy their own models and work on real‑world projects. As part of its teaching and research in the Department of Informatics, King’s will provide its artificial intelligence students with Mistral’s tools and models to prepare them for future roles and responsible AI use. The collaboration gives both students and researchers at King’s access not just to software but also to support from Mistral’s engineering team, webinars and knowledge‑exchange opportunities, reflecting a shared commitment to open‑source and responsible AI
AI and the Future of Universities – HEPI
This collection brings together prominent voices from higher education, industry and policy to argue that AI shouldn’t be viewed only as a threat but as a prompt to rethink how universities teach, research and operate in a fundamentally changed world. Focusing on themes such as AI‑literacy, assessment redesign, workforce development and strategic operations, the report urges UK universities to adopt a coherent AI roadmap—embedding ethics, technology choice and pedagogical leadership.
Environment
AI trial cuts sewage pollution, says South West Water – BBC News
South West Water’s AI trial is helping to prevent sewer blockages before they cause pollution. Using data from 12,000 sensors and CCTV footage to spot early warning signs, the system has already helped avoid around 200 pollution incidents. The initiative aims to improve environmental performance after the company received a poor rating from regulators. By harnessing AI to monitor sewers, South West Water has improved its ability to catch blockages early, cutting pollution incidents and freeing human operatives to cover more ground. The approach supports a wider £760m push to raise environmental standards.
Research
Abundant Intelligence – Sam Altman
Sam Altman reflects on the extraordinary growth in AI services and argues that increasing access to AI will become a fundamental driver of both the economy and human opportunity. He suggests that access to AI might even become a “fundamental human right”
He then turns to the infrastructure side: the real constraint today is compute — both inference compute (using models) and training compute (making them better). To address this, Altman lays out a bold vision: build a “factory” capable of producing a gigawatt of new AI infrastructure every week. He acknowledges this is extremely challenging and will require innovation across the full stack — from chips to power to buildings to robotics. He further emphasises doing much of this in the US, noting that other countries are advancing on chip fabs and energy production faster and that his organisation wants to help reverse that.
He explains the stakes: if we don’t scale compute, we’ll be forced to decide which big problems to prioritise (for example: curing cancer OR tutoring every student). Altman’s call to action is to not choose between them — instead, invest in scaling infrastructure so we can pursue many of these major goals in parallel.
Measuring the performance of our models on real-world tasks | OpenAI
OpenAI has introduced a new evaluation framework named GDPval to track how AI models perform on economically significant, real‑world tasks. The idea is to shift from conventional academic style benchmarks (e.g., exam questions, synthetic puzzles) to tasks that reflect actual deliverables in workplaces—such as drafting a legal brief, creating an engineering diagram, developing a nursing care plan or analysing financial spreadsheets.
The benchmark spans 44 occupations drawn from nine major industries that contribute most to the U.S. GDP and includes 1,320 specialised tasks in the full set (with a public “gold subset” of 220 tasks). These tasks were crafted and vetted by professional subject‑matter experts.
New research finds that large‑language models often oversimplify scientific findings, losing key details in the process — and the latest models appear worse than earlier ones. By analysing almost 4,900 model‑generated summaries of academic papers, researchers found that LLMs were about five times more likely than humans to over‑generalise results, sometimes turning “may be safe under certain conditions” into “is safe and effective.” The study warns that as tools like ChatGPT, Llama and Claude become more embedded in scientific workflows, the risk of large‑scale misinterpretation of evidence grows — especially because models are often trained on simplified journalistic summaries rather than full primary sources.
A new wave of ‘Shadow AI’ is worrying workplaces | TechRadar
A recent study finds that 71 % of UK workers have used unauthorised AI tools at work—and only 32 % express concern about the risks—raising major security and data‑governance concerns for universities and other organisations. The use of so‑called “shadow AI” (employees using AI outside sanctioned systems) is growing because many feel their organisation doesn’t provide appropriate tools or the training, but this creates vulnerabilities around data leaks, compliance and tool accuracy. For higher‑education institutions and similar organisations, the push is now to respond by offering approved enterprise‑grade AI tools and education—not only to enable productivity, but to manage risk, rather than simply banning individual tools.
Perceptions of AI
AI disclosure? Maybe it’s nunya. – Kelly Webb-Davies
Kelly Webb-Davies makes a provocative case against mandatory AI use disclosure in writing, arguing that bias against AI-generated text can reinforce existing linguistic discrimination. She suggests that AI tools can be empowering for marginalised groups and that requiring disclosure could do more harm than good. This article challenges the norm of AI transparency, especially for those using it to overcome systemic language barriers.
Avi Schiffmann’s AI pendant “Friend” is making waves—and enemies—in NYC subways with a $1M ad that’s been vandalized. He calls the graffiti “art,” seeing backlash as part of the product’s message about tech, loneliness, and modern friendship. Marketed as an AI confidant, the Friend necklace records interactions and replies via Google’s Gemini, but its privacy terms are raising red flags. Schiffmann admits the fine print is intentionally “extreme” and expects legal trouble—but sees that, too, as part of the journey.
Accessibility
AI as an access tool for neurodiverse and international staff | THE Campus Learn, Share, Connect
This article argues that when used thoughtfully, generative AI can act as an access tool for neurodiverse and international university staff—reducing the cognitive burden of admin, language editing and formatting, and thereby improving equity of access. Instead of seeing AI purely as a threat or shortcut, the author suggests framing it as a support mechanism: drafting emails, trouble‑shooting language registers, structuring documents and generating accessible materials—all under human review. The piece recommends that universities embed AI use into inclusive workflows by providing safe tools, clear guidance, peer support and reusable prompt libraries—so that staff from diverse backgrounds can reclaim time for substantive work.
Vendor news
Introducing ChatGPT Pulse | OpenAI
OpenAI has launched a mobile preview of Pulse, a version of ChatGPT that proactively researches and suggests personalised updates each day based on your chats, calendar and preferences. Pulse is designed to change ChatGPT from a passive‑tool you ask into an assistant that brings relevant ideas, reminders and next steps to you each morning — you control what it looks up and what you see. In this rollout, ChatGPT will scan your history, feedback and connected apps (like Gmail/Calendar) overnight, then deliver a focused card‑based digest. It’s still early stage, and OpenAI emphasises it won’t always get things perfect.
The next chapter for UK sovereign AI | OpenAI
OpenAI’s recent UK‑government collaboration gives civil servants access to ChatGPT Enterprise and introduces UK data residency for the UK API platform, reinforcing the UK’s “sovereign AI” ambition. With an agreement with the UK Ministry of Justice and the launch of UK‑based data storage, OpenAI signals its commitment to enabling AI adoption across Britain’s public and private sectors. For UK universities and institutions, the piece highlights how OpenAI is aligning with the UK AI Action Plan — offering local infrastructure (“Stargate UK”), data residency, and tailored public‑sector tools.
Microsoft is giving Copilot AI faces you can chat with | The Verge
Copilot Labs now offers “Portraits” — about 40 stylised human avatars you can pick when talking to the AI via voice, to make interactions feel more natural. The feature uses the model VASA‑1 from Microsoft Research to animate a single image into a speaking avatar, and it’s being rolled out in the US, UK and Canada in a limited experiment. Microsoft says this avatar‑approach was developed for users saying they’d be more comfortable talking to a face, and while it’s still experimental it points to future AI assistants that look and sound more human.
AI-powered teaching and learning for all Microsoft 365 education customers | Microsoft Community Hub
Microsoft is rolling out a range of AI‑powered teaching and learning tools at no extra cost for its academic Microsoft 365 licences—covering lesson planning, rubrics, flashcards, and more. Students age 13+ will have access to new learning agents and activities (flashcards, quizzes, etc.), while educators get AI support for differentiated instruction, feedback and LMS integration.
Legal
In a major legal move, Reddit has filed suit in a federal court in New York against Perplexity AI, Lithuanian web‑scraping firm Oxylabs, the botnet alias AWMProxy and query‑API provider SerpAPI, accusing them of orchestrating an “industrial‑scale, unlawful” operation to harvest Reddit comments for commercial AI‑training use.
Reddit alleges that these companies circumvent its anti‑scraping protections and legitimate licensing routes—some even allegedly scraping content indirectly via search engine results—rather than engaging with Reddit’s paid agreements for access.
Reddit further frames this as part of a broader challenge in the AI industry: many bots rely on large text‑archives (including Reddit’s vast user discussions) to learn human‑language patterns, so the question of how data is collected and used is becoming central. The case follows a previous suit against Anthropic, signalling Reddit’s intent to assert control over its community content as AI development accelerates.
Government
Italy first in EU to pass comprehensive law regulating use of AI | Italy | The Guardian
Italy has become the first EU country to pass a comprehensive AI law, introducing criminal penalties for harmful AI‑generated content and limiting minors’ access to such technology — a move aimed at promoting “human‑centric” innovation. The new Italian legislation aligns with the EU AI Act by requiring transparency and human oversight in sectors from healthcare to education, while also earmarking state‑backed VC funding to support local AI firms.
Under the law, children under 14 need parental consent for AI access, deepfakes that cause harm carry prison sentences of 1–5 years, and AI‑generated works receive copyright protection only if genuine human effort is involved — signifying a major regulatory milestone.
Find out more by visiting our Artificial Intelligence page to explore publications and resources, learn more about our communities and sign up for our AI Literacy training.
For regular updates from the team sign up to our mailing list.
Get in touch with the team directly at AI@jisc.ac.uk
One reply on “October 2025 round-up of interesting AI news and announcements”
[…] AI” (unauthorised employee use of AI tools) is becoming a real risk in organisations. Artificial intelligenceTrend insight: Employees use AI tools outside official stack, creating data governance, security, […]