Our September collection of articles and announcements to share this month.
Education
Why access to AI privacy must be part of the Digital Inclusion Conversation
Our colleague, Tom Moule, argues that digital inclusion should go beyond access to devices, connectivity, and basic skills: it must also include privacy protections in AI tools. Students from disadvantaged backgrounds are more likely to use free tools with fewer privacy safeguards, which inhibits their ability to engage fully, experiment, and explore complex or sensitive topics. Privacy isn’t only an ethical issue—it affects how confidently students interact with AI. When they suspect interactions might be logged or shared, they hold back, simplifying prompts, avoiding risk or creativity. This can stunt skills development over time.
This has downstream effects. Those who can’t afford or don’t trust private tools may end up with weaker learning, less creativity, and fewer opportunities in a job market where AI proficiency increasingly matters.
Oxford University gives students access to AI platform – BBC News
Oxford University has become the first UK institution to provide campus-wide access to ChatGPT Edu, OpenAI’s tailored version of the AI chatbot designed for educational use. The rollout follows a successful pilot involving hundreds of staff and students and is part of a broader five-year partnership with OpenAI. The initiative is intended to support more personalised learning, spark innovation, and enhance research, all while safeguarding data privacy. Accompanying the launch, Oxford will also offer training focused on ethical and effective use of generative AI tools in academic settings.
Universities are increasingly shifting away from traditional exams toward more varied assessment types—presentations, vivas, quizzes. Our colleague, Sue Attewell, says this change comes from a need to “AI‑proof” assessments, favouring methods like verbal discussions.
When choosing a university, it’s important to look not just at course content but how you’ll be assessed. Historic assessment methods aren’t always fit for today’s world, so many institutions are experimenting with authentic assessments beyond just exams. The article argues that students who dislike high‑stakes exams should check the assessment style of prospective courses.
Assessment, and Academic Integrity
We cannot address the AI challenge by acting as though assessment is a standalone activity | Wonkhe
As AI reshapes higher education, Duna Sabri argues that focusing only on assessments is too narrow. The piece urges universities to shift from just policing academic integrity towards rethinking how teaching, learning and assessment interrelate; subject experts must build curricula that address AI’s impacts ethically and practically.
Rather than seeing assessment as a separate challenge to be “fixed,” Sabri suggests a programme‑level, subject‑oriented redesign that supports intellectual, creative and ethical student development in an AI world.
Research
Why language models hallucinate | OpenAI
OpenAI’s paper on “Why language models hallucinate” investigates why even the most advanced language models continue to produce confidently wrong statements (“hallucinations”), and how the way we train and evaluate them plays a major role.
Current benchmarks focus on accuracy without giving much reward for admitting uncertainty or saying “I don’t know.” Since guessing when unsure gives some chance of being right, but admitting uncertainty scores nothing, models are biased toward guessing. This leads them to produce incorrect statements instead of qualifying their responses.
Proposed fixes:
- Evaluate differently: Adjust scoring so that models are rewarded not just for being right, but penalised more for being confidently wrong. Also, give partial credit for expressing uncertainty.
- Encourage abstention: In situations where the model is uncertain, it should be better to withhold an answer or ask for clarification rather than guess.
- Calibration: Build models that better understand their own confidence (or lack thereof), rather than those that inflate confidence because it leads to better benchmark scores.
How People Use ChatGPT – by David Deming – Forked Lightning
OpenAI researchers have released a detailed study charting ChatGPT’s rise from its 2022 launch to 750 million weekly users by mid-2025. The data shows not only rapid adoption but deepening engagement, with users sending over 2.6 billion messages per day. Initial demographic gaps—such as those tied to gender and income—have closed rapidly, with usage now widespread across income levels and countries.
Privacy
LinkedIn set to start to train its AI on member profiles | TechRadar
LinkedIn is rolling out a policy, starting 3 November 2025, to use member profiles, resumes, public posts and other activity data to train its AI models. By default, accounts will be opted in to this data‑use for AI training. LinkedIn says users can opt out, but that this only prevents future data from being used—not past data already collected. There’s also a formal objection process via LinkedIn’s Data Processing Objection form.
How to disable training on your profile
If you want to stop LinkedIn using your profile/data for AI training, here’s what to do:
- Go to Settings on LinkedIn.
- Navigate to “Data privacy” section, then to the “How LinkedIn uses your data” subsection. Within that you’ll find a setting called “Data for Generative AI Improvement.”
- Disable that setting. That opts you out of future data being used.
Vendor news
Gemini for Students — Get Google AI Pro Free for a Year for Study Help
Students at eligible UK universities can get Google Gemini + AI Pro free for 1 year. The offer includes tools to turn lecture notes into quizzes and audio overviews, help with homework and writing, and research support with sources. To claim it, students need to verify their enrolled status, have a Google account, be aged 18+, and sign up by 3 November 2025.
Microsoft to lessen reliance on OpenAI by buying AI from rival Anthropic | TechCrunch
Microsoft will begin incorporating AI from Anthropic into its Office 365 applications n a bid to reduce its reliance on OpenAI. This strategy comes while Microsoft renegotiates its agreement with OpenAI and as both companies move toward greater self‑sufficiency: Microsoft has rolled out its own AI models and OpenAI is investing in its own inference and training infrastructure (including producing AI chips with Broadcom and launching a jobs platform to compete with LinkedIn).
Microsoft insists that it remains committed to its long‑term partnership with OpenAI even while diversifying its AI sources.
Building towards age prediction | OpenAI
OpenAI has published two related pieces: “Building Towards Age Prediction” (which explains how they plan to detect whether a user is a minor) and “Teen Safety, Freedom, and Privacy” (which lays out the principles guiding how they’ll treat minors vs adults). Together, they show how the company is trying to strike a difficult balance between protecting under‑18 users, preserving individual rights, and navigating the privacy trade‑offs that entails.
OpenAI will use behaviour‑based age prediction to tell if someone is under 18. If uncertain, they’ll default to treating them as teens. While OpenAI values privacy, if there’s a conflict, teen safety has priority. Examples include limiting certain content (e.g. flirtatious language, self‑harm discussions) and possibly contacting parents or authorities in urgent harm situations.
For users aged 13 and above, parents will be able to link into the teen’s account, configure how the system responds, disable features like memory or chat history and set blackout times. In urgent cases (e.g. suicidal ideation), OpenAI may contact parents or authorities.
OpenAI acknowledges that some of these measures reduce privacy. They commit to limiting access (even OpenAI employees won’t routinely access private conversations) but retain automated monitoring for misuse or serious risk.
Government
Switzerland launches transparent ChatGPT alternative – SWI swissinfo.ch
Switzerland has rolled out a new open-source large language model called Apertus, developed by Swiss universities. It aims to be more transparent, secure with data, and better suited for scientific and industrial users rather than to compete feature‑for‑feature with big commercial models. Apertus gives users full access—not just to the outputs, but to the model’s design, training recipe, and the ability to run it locally.
Trusted third-party AI assurance roadmap – GOV.UK
The UK government has doubled down on its dual strategy for AI: fuelling investment while building trust. A new roadmap outlines how third-party assurance—independent checks on AI safety, fairness, and reliability—will become a cornerstone of responsible AI deployment, with funding and professional frameworks to support it.
At the same time, private investment in UK AI firms has hit record highs (£2.9 billion invested last year), signalling strong growth but also underlining the need for skilled talent and robust governance—areas where universities have a big role to play.
Head of UK’s beleaguered Alan Turing Institute resigns | Artificial intelligence (AI) | The Guardian
Jean Innes, who became chief executive of the Alan Turing Institute in 2023, has announced she will resign later this year. Her departure comes after heightened internal discord at the institute, including a staff revolt, a whistleblower complaint, and growing pressure from its key funder, the UK government. A letter from the technology secretary in July urged ATI to shift focus towards defence and national security, and urged consideration of new leadership. Innes said that with the current strategy concluding, it was time for a “new chapter.”
The institute has been undergoing a strategic overhauling under a programme dubbed Turing 2.0, which re-prioritises its work around three main areas: health, the environment, and defence & security. This transformation has led to controversy—staff have expressed concern over ATI’s credibility, some projects have been dropped (including ones on online safety, housing, and health inequality), and about 50 roles are at risk through redundancy. While the board has pledged to maintain work in non‑security areas, some employees fear that the shift will dilute the institute’s broader mission to tackle societal challenges.
Between November 2024 and February 2025, the UK government ran a trial where over 1,000 tech workers across 50 departments used AI coding assistants from Microsoft, GitHub (Copilot), and Google (Gemini Code Assist).
The trial found AI coding assistants can boost productivity, with ~65% of users saying they do tasks more quickly; but concerns remain about code quality, security, and the need for downstream checks. The UK is aiming to save £45 billion across public services via AI rollout, early trials show real benefits in the coding domain.
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