Copilot and enterprise AI in universities: a reading list

SURF, the Dutch sector IT cooperative, closed its Data Protection Impact Assessment of Microsoft 365 Copilot for Education in May 2026 with two medium risks still unresolved, after advising institutions against the paid product in December 2024 and moving to cautious, case-by-case use in September 2025. Scripted tests by SURF and Privacy Company produced fabricated dates, non-existent sources and gender-biased salary advice, and only users aged 18 and over were assessed. The reading list also cites NTNU's 2024 sandbox pilot, which found Copilot hallucinated and tended to confirm whatever the user asserted, and Virginia Tech's January–March 2025 pilot, in which 82% of the 85 survey respondents used Copilot daily and 92% would recommend it.

It was tricky to find this scattered literature so I’m sharing here in case it’s useful for other people. These are Opus 5.5 summaries in the first instance, I’ll update any that inaccurate as I continue to work my way through this list: Microsoft Copilot in universities SURF and Privacy Company, DPIA Microsoft 365 Copilot for Education 2024–2026 https://vendorcompliance.surf.nl/wp-content/uploads/2026/05/20260527-SURF-2nd-Update-DPIA-Microsoft-365-Copilot-public-version.pdf SURF, the Dutch sector IT cooperative, advised institutions not to use the paid Microsoft 365 Copilot in December 2024. It moved to recommending cautious, case-by-case use in September 2025, and closed the assessment in May 2026 with two medium risks still unresolved. Scripted tests produced fabricated dates, non-existent sources and gender-biased salary advice. Only users aged 18 and over were assessed. Experts raise privacy concerns about Microsoft 365’s Copilot Cursor, TU/e, 2025 https://www.cursor.tue.nl/en/news/2025/januari/week-1/experts-raise-privacy-concerns-about-microsoft-365s-copilot After SURF’s advice, Eindhoven University of Technology kept its pilot of the paid product to about fifty employees, fewer than one per cent of staff. Its privacy officer drew a line between the paid product, which is connected to the university’s documents, and the generic chat assistant. Datatilsynet and NTNU, Copilot through the lens of data protection 2024 https://www.datatilsynet.no/en/regulations-and-tools/reports-on-specific-subjects/reports/ntnu-copilot-through-the-lens-of-data-protection/ NTNU piloted Microsoft 365 Copilot in the Norwegian data protection authority’s regulatory sandbox. The regulator concluded the tool could be adopted, though not by everyone and not for everything, and that each use needs its own defined purpose and legal basis. NTNU judged that its information management was not ready for organisation-wide use. NTNU, Pilotere Copilot for Microsoft 365: findings report 2024, in Norwegian https://www.ntnu.no/documents/10507/1350391340/2024-06-17+NTNU+Funnrapport+Pilotere+Copilot+for+Microsoft+365.pdf NTNU’s own report on the sandbox pilot has eight findings. Copilot worked well when users already understood the task, but it hallucinated and tended to confirm whatever the user asserted. Checking its outputs sometimes took longer than doing the task by hand. The report also identifies risks around governance and employee monitoring. NTNU, DPIA of Microsoft 365 Copilot Chat 2025, in Norwegian https://i.ntnu.no/documents/1305837853/1306916738/2025-10-31+DPIA+Microsoft+Copilot+Chat+with+enterprise+data+protection.pdf/d1504e00-eeab-8a72-0f8b-15a5daf1d2a4?t=1762329605759 Copilot Chat has been open to NTNU staff since 2023 and to students since 2024, and it logged more than 250,000 interactions a month by autumn 2025. The assessment concedes that users’ purposes are not fully described. After the student parliament objected, the measures prohibit using chat logs to investigate cheating. Virginia Tech, Pilot for Microsoft 365 Copilot: Outcomes 2025 https://ai.vt.edu/content/dam/ai vt edu/Virginia-Tech-Pilot-for-Microsoft-Copilot-Outcome-Report-04-2025.pdf 177 volunteer faculty and staff used Microsoft 365 Copilot from January to March 2025. Of the 85 who answered the closing survey, 82% used it daily and 92% would recommend it. Expectations fell over the pilot for automation, translation and data analysis. The university then approved it as an option for departments to buy, rather than licensing everyone. University of South Carolina, Microsoft customer story 2025 https://www.microsoft.com/en/customers/story/25399-university-of-south-carolina-microsoft-365-copilot Microsoft’s account of a Microsoft 365 Copilot pilot that included students. It reports that 84% of users saved one to five hours a week, but gives no sample size, method or costs. University of South Carolina Faculty Senate minutes April 2025 https://sc.edu/about/offices and divisions/faculty senate/about/news/2025/april2 minutes.pdf The minutes record 160 pilot participants and 350 licences bought by individual units at US$30 per user per month, with no discount. Storage use rose roughly threefold, and extra storage would cost US$300,000 a year. Ohio University, Microsoft 365 Copilot pilot 2025 https://www.ohio.edu/news/2025/08/ohio-conducts-pilot-program-assessing-artificial-intelligence-tool-microsoft-365 A year-long pilot with about 130 staff, designed to test whether the add-on licence was worth what it cost. Only some users reported consistent savings, typically one to three hours a week. Administrative and research-support roles benefited more than teaching roles. University of Manchester, Microsoft customer story 2024 https://www.microsoft.com/en/customers/story/19241-university-of-manchester-microsoft-copilot Microsoft’s case study of Manchester’s first Microsoft 365 Copilot cohorts, which grew from 120 to 300 licences. Reported daily time savings range from 15 minutes to more than three hours. Two thirds of the first cohort said they needed ongoing learning. Indiana University Kelley School of Business, Microsoft customer story 2025 https://www.microsoft.com/en/customers/story/24613-indiana-universitys-kelley-school-of-business-microsoft-365-copilot Microsoft’s case study cites a randomised trial in which students using Copilot performed 10% better and took about 40% less time. No paper reporting the trial could be found. Conde Ruano, Satisfaction with the Use of Microsoft 365 Copilot in Technical Translation Training 2026 https://link.springer.com/chapter/10.57088/978-3-7329-8690-3 16 29 translation students at the University of the Basque Country used Copilot across five tasks. The share who would recommend it ranged from 100% for software localisation to 60% for sentence translation, where literal output was the main complaint. Klawa, Rajpal and Thomas, Gen AI in Proof-based Math Courses: A Pilot Study 2025 https://arxiv.org/abs/2509.13570 Copilot was the only generative AI tool Indiana University East approved for use with university data. Of 19 student survey responses, 8 had not used generative AI and 7 regarded using it as cheating. Students who did use it found its proofs often inaccurate. Martín-Moncunill and Alonso Martínez, Students’ Trust in AI and Their Verification Strategies 2025 https://doi.org/10.3390/educsci15101307 A Madrid university licensed Copilot for all its students, trained them and promoted the tool. Among 132 students surveyed, ChatGPT remained the main tool for 82.6%, and 56.1% had never used Copilot. Most students’ verification practices were weak. Mellado and Cubillos, Can Generative Artificial Intelligence Outperform Self-Instructional Learning in Computer Programming? 2025 https://doi.org/10.3390/app15115867 71 engineering students in Chile learned PHP either with Copilot, accessed through university accounts, or with instructional videos. After adjusting for prior knowledge, the video group scored significantly higher on the post-test. Hasanein, Responses to the AI Revolution in Hospitality and Tourism Higher Education 2025 https://doi.org/10.3390/ejihpe15030035 A survey of 760 students at nine Egyptian public universities who were already using Copilot. Ease of use and institutional support were the best predictors of intention to use it. The study measures acceptance, not learning. Osadcha, Szynkiewicz and Chishti, Using Microsoft Copilot Chat in the Work of IT Educators 2024 https://www.ntnu.no/ojs/index.php/nikt/article/view/6202 NTNU computer science educators attended a one-day workshop and received a guide to Copilot Chat. Self-reported knowledge and use shifted afterwards, but the sample was small: 18 respondents before and 14 after. Osadcha et al., Studying IT Educators’ Satisfaction with Using Microsoft Copilot Chat 2025 https://journal.iitta.gov.ua/index.php/itlt/article/view/6184 Eighteen NTNU educators rated Copilot Chat’s responses to 22 professional tasks, with an average of about 3.8 out of 5. The authors advise against using it for rubric-based essay grading. Özdemir, Getting Support from Microsoft Copilot in Lesson Plan Preparation 2024 https://journal.iitta.gov.ua/index.php/itlt/article/view/5859 24 pre-service teachers in Turkey used Copilot to draft lesson plans. They valued its suggestions for sources and activities, but judged its subject knowledge and its assessment questions unreliable. Siiman, AI in Teacher Education: An Introductory Training Session Involving Microsoft Copilot 2024 https://link.springer.com/chapter/10.1007/978-3-031-65884-6 24 A 45-minute session introduced Copilot to 43 pre-service teachers. Participants saw most value in lesson planning, writing assessment questions and generating resources. Pellesi, Trust in AI Feedback: Challenges in Pharmacy Education 2025 https://link.springer.com/article/10.1007/s40670-025-02440-6 36 master’s pharmacy students could ask Copilot for feedback alongside peer feedback. 60% found the AI feedback clear, but only 40% found it useful, and they rated peer feedback as more relevant. Enterprise AI in universities Virginia Tech, ChatGPT Edu Pilot Findings Report 2025 https://ai.vt.edu/content/dam/ai vt edu/ChatGPT-Edu-Pilot-Findings-Report.pdf 480 staff and graduate assistants used ChatGPT Edu from January to May 2025. Usage logs showed 78% were active in a typical week, and 77% of survey respondents saved at least an hour a week. Time spent verifying hallucinations was the most frequently cited drawback. Virginia Tech, ChatGPT Edu Beneficial Use Cases 2025 https://ai.vt.edu/content/dam/ai vt edu/ChatGPT-Edu-Beneficial-Use-Cases-Report.pdf A companion report collecting uses described by 405 pilot participants across administration, research and teaching. It is illustrative rather than quantitative. University of Oxford, ChatGPT Edu for all staff and students 2025 https://www.ox.ac.uk/news/2025-09-19-oxford-becomes-first-uk-university-offer-chatgpt-edu-all-staff-and-students After a year-long pilot with around 750 staff, Oxford gave ChatGPT Edu to all staff and students. It sits alongside Copilot Chat, Microsoft 365 Copilot licences bought by departments, Gemini and NotebookLM. No pilot data or costs have been published. Google and Oxford Gemini and NotebookLM collaboration 2026 https://blog.google/products-and-platforms/products/education/oxford-gemini-notebooklm/ Google reports an Oxford pilot in which 85% of survey respondents said its tools increased their productivity. No sample size or method is given. Cal State struck a deal with OpenAI. Some students and faculty refuse to use it CalMatters, 2026 https://calmatters.org/education/2026/05/california-state-university-open-ai-chatgpt-contract/ California State University signed an 18-month ChatGPT Edu contract worth US$17 million. By April 2026, only 0.7% of students and 16% of faculty had completed the voluntary training. He et al., Embracing Generative AI Tools in Higher Education: A Survey Study at HKUST 2025 https://doi.org/10.1080/17516234.2024.2447195 A survey of 680 students at one of the first universities to offer an officially protected ChatGPT service. Students reported substantial use and perceived benefits for both learning and careers. How and Why the University of Michigan Built Its Own Closed Generative AI Tools EDUCAUSE Review, 2024 https://er.educause.edu/articles/2024/2/how-and-why-the-university-of-michigan-built-its-own-closed-generative-ai-tools Michigan built U-M GPT in 2023 and reports about 15,000 daily users. It gives privacy, accessibility for screen-reader users and equal access as its reasons for building rather than buying. Ushering in a New Era of AI-Driven Data Insights at UC San Diego EDUCAUSE Review, 2025 https://er.educause.edu/articles/2025/2/ushering-in-a-new-era-of-ai-driven-data-insights-at-uc-san-diego TritonGPT runs on UC San Diego’s own servers and is available to 38,000 employees. Adoption follows a Pareto distribution, with about a quarter of users active daily. How Three Universities Developed Their Chatbots EdTech Magazine, 2025 https://edtechmagazine.com/higher/article/2025/05/how-three-universities-developed-their-chatbots A profile of the in-house platforms at Michigan, UC Irvine ZotGPT, with 15,000 unique users and Arizona State CreateAI . Letting Educators Take Control of Generative AI Teaching@Sydney, 2023 https://educational-innovation.sydney.edu.au/teaching@sydney/letting-educators-take-control-of-generative-ai-to-improve-learning-teaching-and-assessment/ Cogniti lets Sydney educators build their own AI agents inside Canvas, steered by their own instructions and rubrics. Teachers can see the conversations students have with them. How Generative AI Can Make Personalised Feedback at Scale More Consistent and Efficient Teaching@Sydney, 2024 https://educational-innovation.sydney.edu.au/teaching@sydney/how-generative-ai-can-make-personalised-feedback-at-scale-more-consistent-and-efficient/ Markers on courses of 220 to 1,400 students used Cogniti agents to expand their brief comments into feedback linked to the marking rubric. New DPIA on EduGenAI Privacy Company, 2025 https://www.privacycompany.eu/blog/nieuwe-dpia-op-edugenai-nederlandse-generatieve-ai-oplossing-voor-het-onderwijs EduGenAI is a platform run by the Dutch sector that gives pseudonymised access to open-weight and commercial models. The assessment identified 12 risks, which it judges can be reduced only if the planned safeguards are delivered. Frequently Asked Questions About SURF AI SURF, 2026 https://www.surf.nl/en/article/frequently-asked-questions-about-surf-ai SURF AI leaves its pilot phase at the end of 2026 and becomes a sector-wide service in January 2027. Open-weight models are included in the sector package, and commercial models cost extra. The Advantages of Context Specific Language Models: The Case of the Erasmian Language Model 2024 https://arxiv.org/abs/2408.06931 Erasmus University Rotterdam trained a small model on its own research output for under €4,000. In feedback from 23 students it was praised for its academic register, but its benchmark scores were close to chance. Testing ELM’s Ability to Return Useful Results with Prompts About the Editorial Style Guide University of Edinburgh, 2025 https://blogs.ed.ac.uk/website-communications/testing-elms-ability-return-useful-results-prompts-about-editorial-style-guide/ An informal test of ELM, Edinburgh’s in-house AI platform. On longer passages it gave irrelevant or contradictory advice and drifted away from the style guide it had been given. AI: Friend or Foe? Reflections from UCISA DIG 2025 AppsAnywhere, 2025 https://www.appsanywhere.com/blog/ai-friend-or-foe-reflections-from-ucisa-dig-2025 Summarises UWE’s student-support chatbot, which was built in seven weeks after the university closed its on-campus advice points. It reportedly led to 38% fewer student calls. The Digital Divide: Student Generative AI Access Inside Higher Ed, 2025 https://www.insidehighered.com/news/tech-innovation/artificial-intelligence/2025/04/21/half-colleges-dont-grant-students-access Half of 108 US chief technology officers surveyed said their institution gave students no institutional access to generative AI. Cost was the most frequently cited barrier. AI Procurement in Higher Education: Benefits and Risks of Emerging Tools EDUCAUSE Review, 2025 https://er.educause.edu/articles/2025/3/ai-procurement-in-higher-education-benefits-and-risks-of-emerging-tools Flags AI features added to existing products without notice, opaque pricing and underestimated costs. It recommends shorter contracts and spreading risk across vendors. EDUCAUSE QuickPoll Results: AI-Related Procurement EDUCAUSE Review, 2025 https://er.educause.edu/articles/2025/5/educause-quickpoll-results-ai-related-procurement In a poll of 270 members, teaching and learning units were involved in AI procurement at only 58% of institutions. Only 36% of institutions re-review a product when new AI functions are added.