# Generative AI at the University of Roehampton: Report and Recommendations > *"Modern AI is fundamentally dependent on corporate resources and business practices, and our increasing reliance on such AI cedes inordinate power over our lives and institutions to a handful of tech firms.”* > (Whittaker, 2021) To whom it should concern, but seemingly does not enough. I am writing the present in preparation to my appearance on a student panel regarding AI at the University of Roehampton, in which I will report some serious concerns I have about UR's position towards, and adoption of, generative artificial intelligence technologies. I feel compelled to prefix this text by a disclosure of my recent, official medical diagnosis of Autism, which I hope will help contextualise my tone: we Autists put our sense of justice and honesty above social niceties. I should also specify that this was written without AI assistance, not even to revise the tone down. ## About me Some background on my own position will be helpful: I am currently studying for a PhD in UR's School of Education. My research topic is student use of AI in HE assessment - specifically faculty's attitudes towards it. My director of study is Paul Dickerson, and my supervisor Miles Berry. I first trained in Computer Science in the early '00s, then worked as a software engineer in web technologies for more than a decade, moving through development, architecture and technical management. I am therefore familiar with the manufacturing process of the proverbial sausage, and am now mostly vegetarian, as it were. I then trained to teach Computing in secondary education, here at Roehampton, under the tutorship of Miles, taught for a couple of years, before coming back for this PhD. This experience, and my critical position towards the digital technologies lobby, is what informs my technological choices, my research in general, and the present letter in particular. I am cautiously enthusiastic about the potential of the technology, in education or beyond, but severely critical of its deployment by Big Tech. Over the past two years, I have witnessed with increasing consternation the institutional moves of the University as regards AI, decisions that I find lacking in expertise, foresight, or criticality. As far as I know, there was little to no student consultation in these matters. ## Overview The present is an attempt to remedy this, and speak Truth to Power, even if Power seems more receptive to *statistically plausible text* these days. I will detail my feedback in the chronological order of institutional initiatives as they became manifest to the student body. What all these points are in common, from the point of view of a student, is *performativity*. Decisions regarding AI in this institution seems, at least to me, to be driven by a need to be *seen* to do *something*, rather than the outcome of a joined-up thinking including 1) an understanding of the technology, from a technical, but also political, and economic perspective 2) a prioritisation of the pedagogical mission of the university, and 3) a meaningful inclusion of a variety of voices in the decision making process. From the outside, it seems the University's response is more of a *reaction*: the delay in formulating one have not been in the service of considerate reflection. This delay means the response fails even as a performative endeavour as Roehampton finds itself 2-3 years behind the times, both in terms of the available options for AI integration, and of the perceived student attitudes to which it is trying to respond. To the latter point: whilst almost all students have now adopted the technology, they express serious, and growing concerns with its effects, both from a global, environmental perspective and a local, pedagogical one, and uncritical adoption from their institution seems unlikely to be seen as a plus [@stephensonStudentGenerativeAI2026]. **US survey**. To the former point, what AI initiatives the university has taken seems dominated by OpenAI's products, ignoring both the plethora of alternative options, and the increasingly manifest problematic nature of the company's practices, products, or politics. ## WriteNow The first visible AI initiative was WriteNow, Paul's GPT bot for assessment support. As far as the idea is concerned, it is commendable, editorial feedback being a valid use case for systems based on large language models (LLMs); the implementation, however, is problematic. There are many ways of designing and deploying an LLM-backed conversational agent ('chatbot'). Even in the fall of 2024, there were many alternatives to OpenAI in this domain: not just the systems of its competitors, but also many open-weight models one could deploy, locally or in rented 'cloud' compute. The choice to go for a ChatGPT agent, seemingly for lack of awareness of the alternatives, or understanding of the implication of this choice, presents a number of issues: 1) Environmental harms are climbing, year on year, in the list of most pressing concerns, as regards AI, for undergraduates in surveys from the Higher Education Policy Institute (HEPI) [@freeman2024provide;@freeman2025student;@stephensonStudentGenerativeAI2026]. The irony of signage in the Library advertising the system, next to these promoting the university's sustainability commitments, is biting. It feels absurd having cafeteria products labelled according to their carbon footprints, yet directing students to systems whose environmental harms are well known, and growing. As a use case, feedback on writing does not require the kind of very large, 'frontier' model that OpenAI or its competitors offer. At the time of writing, it is entirely feasible to run a system offering this functionality on a one's computer, assuming it has the GPU capacity. It is definitely feasible to run it on university infrastructure, local or cloud, which gives improved data protection, and a better visibility on the compute and energy expense. **UPDATE WriteAboutNow** 2) Privacy and data protection concerns, too, are shared by many students. Whilst students making their own choices as what information they choose to upload to what system is one thing, the University tacitly endorsing them uploading _specific_ information to a _spectic_ system is another. WriteNow, the assessment support bots, or any custom GPT relying on a 'knowledge base' as an attached file, cannot be accessed from an anonymous ChatGPT session anymore, a change in OpenAI's policy that happened sometimes between November 2025, and April 2026. This means that when we direct student to WriteNow, we ask them to have an OpenAI account. I do not think we should. 3) This is a broader issue with OpenAI's GPT models: the company will change the behaviour of the underlying system without notice, meaning systems built on top of GPTs will exhibit inconsistent behaviour. This means in particular that, even if these systems go through some form of testing, successful Quality Assurance (QA) at deployment does not, in fact assure quality in production. In early 2025, OpenAI changed the behaviour of GPT-4o without this update being advertised, let alone resulting in a version number change. This means even systems built on top of OpenAI's application programming interfaces (APIs), locked to a specific version of GPT, were affected by this change - to say nothing of systems built on custom GPTs. This became known as the "sycophantic" update, rolled back in May 2025. [@openaiSycophancyGPT4oWhat2025], and it has been linked to a several instances of users causing harm to themselves or others, including at least 7 suicides, in court cases against OpenAI [@bettsChatGPTAccusedActing2025]. As regards WriteNow, over this period of time, it sought to be so helpful it ended up giving users an actual mark for their work. This happened in a live demo by Miles in a School of Education PGR workshop in February 2025, just as he was trying to illustrate the guardrails of the system. This was the first such farcical moment, but many more would occur. Beyond the manner of its implementation, I have concern about the way WriteNow has been introduced. Framing WriteNow as a meaningful response to Student Voice feedback asking for more support in assessment, as for instance on signage in the David Bell building, is disingenuous. Students are perfectly able to use AI Chatbots to improve their writing, it is likely they were doing this before Paul created WriteNow, and continue do to so on their own term. As such, WriteNow is less a solution than an abdication of responsibility in providing the kind of assessment guidance students were *actually asking for*: a moral hazard, in that the mitigation of this issue by this stop-gap measure decreases incentives to meaningfully address it. To be sure, not all students may know how to use the AI system of their choice to get feedback on written work. I question whether creating a system to solve the problem, as opposed to educating students in solving the problem themselves, is aligned with the stated intention of the University as regards its role in educating students in AI literacy. The university has made the choice of *giving students a fish* - and a problematic one at that, when *teaching them how to fish* is within their grasp: WriteNow's custom prompt is 800 words of common sense prose. ## The University's AI policies In September 2025, Roehampton published student- and academic- facing policies. It is notable that this arrived quite late compared to the rest of the sector, but, as per the saying, better than never. One could have hoped that this delay would be in service of a thoughtful response; this hasn't been the case. ### Student guidelines The three-tier assessment format has provided much needed clarity to students. Much as they are concerned with the technology itself, their most reported concern is use of AI, actual or suspected, leading to them being wrongly accused of academic misconduct[@freeman2024provide;@freeman2025student;@stephensonStudentGenerativeAI2026]. It is, however, unclear whether these guidelines genuinely alleviate this concern, given how much hinges on a load-bearing, but ambiguous clause *"AI should not do the work for you"*. Students and lecturers are left to define the ethical line themselves, something that Corbin and colleagues [-@corbinWheresLineIts2025] found to generate significant anxiety amongst both groups. In fairness, it is hard to imagine this concern being resolved solely by policy, as opposed to *structural* changes to assessment practices [@corbinTalkCheapWhy2025]. What is concerning to me in these guidelines is shifting the onus for responsible use onto the students, First, _"Avoid sharing your academic work or assessments in AI tools that do not guarantee data security"_ [@universityofroehamptonStudentGuidelinesUse2025 p. 3]. It is unclear what "data security" refers to: use of data willingly input into an LLM system is beyond the scope of the Data Protection Act (2018), as it is *de facto* consensually given, hence the previous point of guidance to not input personally identifiable information into them. Without a list of safe tools, or at least criteria to assess their data safety, this seems to expect students to adjudicate the terms of service of global technology companies themselves. Particularly frustrating is the line _"Use AI tools that are approved by your university"_, without further detail as to what these tools are. A student may rightfully expect *this very document* to list those tools, or point to an official resource that does. These unclear clauses hints at boundaries in "the right way" to use AI, yet do not draw them, an uncertainty likely to cause anxiety. Most objectionable to me, however, is the inclusion, in a paragraph reminding students to "Consider the Environmental Impact" (*ibid.*, p. 3), of a mitigating sentence: > As a student, you can make a difference by using AI thoughtfully and only when it adds real value to your work. *At the same time, AI can also support environmental solutions—helping researchers tackle climate challenges and improve energy efficiency.* Being mindful of your digital choices is part of being a responsible and sustainable learner. (*ibid.*, p. 3, emphasis added) AI's potential to help in climate research is point often made immediately after mentioning the considerable environmental harms of generative AI, in a prototypical example of *[False Balance](https://en.wikipedia.org/wiki/False_balance)* ("bothsidesism"). Climate and energy models are a completely different application of deep learning ("AI") than LLMs, using much smaller models, with footprints in training and inference (actual use) that are orders of magnitude lower. More importantly, this point has no relevance to *student use of AI* (for assessment). The triteness of this section and the irrelevance of this sentence made me suspect it could be un-edited synthetic text; [quillbot.com](https://quillbot.com/ai-content-detector) deemed this paragraph, and the whole end of the document, as AI-generated with moderate confidence. ![[Quillbot screenshot.png]] Evidence shows a AI users face a 'social evaluation penalty' in professional settings. No one likes a slop-monger, because it sends the signal the reader was not worth a writing effort. Maybe AI "should not do the work" for the University when it comes to policy drafting, for performative as well as substantive reasons. The farce continues. ### Academic guidelines The problem of UR offloading its policy writing work to LLM systems is less ostensible, but manifests with deeper implications in the guidelines for Academic staff published at the same time [@universityofroehamptonAcademicGuidelinesUse2025]. In particular, this document equates type 1 assessment design with 'assessment *for* learning', type 2 with 'assessment *of* learning', and type 3 with 'assessment *as* learning'. This is for the avoidance of doubt, complete nonsense. The types of assessment design being considered are all assessment *of* learning: summative assessment necessary for passing the module. Mapping N things in one domain onto N things in another, prioritising neatness over meaning, is typical of synthetic text - if the author of the document knew the actual meaning of these terms, they would not have drawn this triple equivalence. This should have been caught in review: the authors may not have the domain knowledge to check LLM output, but, at the very least, faculty within the School of Education does, and hopefully any knowledgeable enough educator. So it is extremely concerning that this document was approved by the University's Senate, in which sit members who do know the meaning of these terms, and who should not have let this text to be published as such. How was this allowed to happen? I understand that, particularly in large institutions, it is often easier to go with the flow than make waves or rock the boat. Still, this seem to speak to a complete lack of countervailing voices - something I hope to remedy by the present. It is concerning that people who are meant to have a say felt, in this instance, unable or unwilling to speak up, and let such slop end up in approved university policy. This reflects very poorly on *all* university governance, because if this was the case for the approval of *this* text, this could have been the case for *any* of them. This is not an AI problem, but like in a lot of aspects of HE, AI is making a pre-existing but unacknowledged problem manifest. ### Research Office guidelines In contrast, the RO's guidelines on use of AI for research and knowledge exchange [@vallanceUseArtificialIntelligence] have been, if not completely human-written, at least meaningfully human-reviewed. They are sensible, and fulfil their role well, providing clarity to researchers whilst keeping an appropriate level of criticality. This is a positive item in the overall response of the University I would be remiss not to acknowledge, even though its quality merely stands out in contrast to the rest of the policy documents. This quality should be the baseline, not the exception. ## Assessment support bots (ASB) ### Autumn 25 In the first term of 2025/26, the University decided that, after a successful pilot in the School of Psychology, a pilot whose success has not been detailed, as far as I know, every module was to have their own assessment support chatbot. These would work from a simple custom prompt, shared by all instances, and an assessment brief Word document, specific to each module, and thus to each bot. The School of Education chose to commission PhD students to help with this work, which is how I ended up heavily involved in the project for the SoE, not so much a front-row seat as a gig as stage-manager of the farce, actually maybe just your regular roadie. The same questions as WriteNow arise: why choose OpenAI for this? Why *create* a system and have students be their passive user rather than *teach* them AI-literate use? In addition, issues specific to the use-case of the project also arise: 1) Whilst feedback on writing is a valid use case for LLM technology, using ChatGPT as an unreliable intermediary layer to help students *find information in a structured document* of, at most a dozen pages is, at least, questionable, at most outright insulting their intelligence. 2) That being said, mandating a pro-forma assessment brief for each module is, *itself, a welcome initiative.* The ASB project will have had this side effect, and here again, I want to acknowledge a positive. 3) The University's direction as regard AI and AI literacy also includes staff training, so the *fish-giving vs. fishing-education tension* applies here too. If time and energy had been instead invested in *educating staff* in creating bots of any kind, this would have solved this problem (generously assuming it needed solving, beyond a pro-forma brief), but also future ones. These points fit within the perception of the university's response as performative, not substantive. This should concern AI-critics and AI-supporters alike, because it satisfies neither of these groups. In my work setting up the ASB, I went beyond the core instructions, and *tested* the chatbots. I also made sure to lock the underlying model to GPT-4o-mini, on account of its smaller parameter count (what makes the largeness in *large* language models). A smaller model means a smaller compute footprint, and correspondingly lower energy and water usage. This model was since deprecated by OpenAI, and bots will now use whatever is the latest model. Some technical issues came to light in the autumn of 2025: 1) The university's insistence, in its paperwork, on Word documents using a one-column table for their formatting, is, here again, harmful. I say here again, because in spite of a working group to render doctoral milestones accessible, the RDCOM forms, which were last revised 4 years ago, still use this paradigm, which renders them inaccessible to screen readers. As regards the bot, the problem manifests in that any document needs first to be flattened into a linear text file (before being converted into a stream of tokens). This, in turns, affects the LLM's capacity to retrieve data, and, given the structural inability of these systems to acknowledge ignorance or uncertainty, they will return misinformation to students. 2) This was particularly problematic for assessment deadlines, whose accurate retrieval was unduly sensitive to time and date formats. For instance, a deadline could not be found, because it was written as 12pm, when the rest were *e.g.* 14:00. Yes, 'artificial intelligence' is that stupid. 3) The table format means that when module leaders wanted to include a table (*e.g.* rubric) in the content, many included it as an image, further decreasing the accuracy of information retrieval. 4) Assessment briefs containing links to Moodle would results in bots that could not be made public because of OpenAI's privacy policy. ### Spring 26 In January, it emerged the University had pivoted its implementation strategy with the assessment bots: Schools and Faculties were only asked to provide assessment brief, with Paul and Yannis setting up the bots from these files. This bot was to be a Microsoft Copilot one, a welcome change, as it would be accessed via students university accounts. This would also simplify set-up, using a Sharepoint folder, rather than have to upload university documents to OpenAI. A more concerning change is that the number of bots was to be reduced, with the aim of having one bot per School/Faculty, in charge of assessment support for all active modules for it. The moment I learnt of this proposal, from Miles, I responded with a link to research on "context rot": the degradation in semantic search performance as LLMs "context window" (the total text input) grows in size [@liuLostMiddleHow2024;@hongContextRotHow2025]. There was no way this idea would work, and anyone keeping abreast of LLM research should know this. My job, that term, was merely to collate and check assessment briefs, liaising with module conveners, which I dutifully did, chasing my hierarchy for news of these new bots. The Copilot test links Paul sent to Miles and I would not work, because they required a more expensive Microsoft licence the University did not have. In April, emails were sent to schools, acknowledging this, and falling back onto plan B, setting up, on OpenAI, custom GPTs still under the one-bot-per-School model. To this effect, Yannis asked for help from the schools, as, ChatGPT having a maximum number of document for 'knowledge bases' of their custom GPTs, the assessement briefs needed concatenated into fewer word files, whilst keeping the size of individual files below fifty pages. He also offered Schools to take back full ownership of the bot setup, on which the school of education took him up. I first followed the prescription, and concatenated Word files into larger document, and using all of them as the 'knowledge base' of a single bot bot. As expected, this did not pass testing, with the system returning outright disinformation about every module tested, or claiming not to know about a module if its brief information was too far down in one of the 50 pages word files. Farcical, again. Yannis' Plan C was to do what we did the previous term: one custom GPT per module. Seeing as the school of Education recovered some latitude in design, I proposed, instead, to have one bot per *programme*. This turned out to be a level of aggregation that works well in terms of LLM performance, with bots backed by a dozen individual briefs; it makes for a simpler set-up, requiring neither Word concatenation, or repeated effort for each module. More importantly, it is directly relevant to students mode of use, with them being able to get assessment support for all modules of their programme from a single bot. This kind of design choice does not require any domain knowledge of LLM operations, it is a self-evident level of aggregation, rather than the School/Faculty level, particularly if this initiative is notionally for the sake of the student. Unlike performance issues in large-context LLM use-cases, this design decision was well within the expertise and capacity of the decision makers, and it was not taken. In testing the bots, we noticed that, some time between November 2025 and April 2026, OpenAI changed their access policy so that anonymous ChatGPT sessions could not upload documents. This also affects documents attached as knowledge base in custom GPTs, meaning, as mentioned above, that both WriteNow and ASBs would not work from an anonymous session, forcing students to have an OpenAI account to use them. This is all quite concerning: 1) We are told of "successful pilots" in the School of Psychology, but evidence for this success has not been provided. What metric for success was used? Was student voice sought? 2) From my perspective, the decisions made about ASBs this term seem to betray a lack of understanding of the technology and its limitations. Is there enough substantive expertise at the decision making level? 3) The change to accessibility of files-backed bots to anonymous sessions of ChatGPT is the same issue of trying to build an information system on top of a service susceptible to significant change in functionality without warning. AI, as many seemingly never tire to say, is here to stay. Is there a long term strategy for AI tooling at the University? ## Moving forward I think the custom GPT were particularly offensive to me as *a software engineer*. Just because one can design and deploy an information system by writing one page of A4 doesn't mean they should, because *there is more to making safe and effective systems than design and deployment.* Maybe it is because I am a software engineer that I even think of them as software. The ease with which they can be built is comparable to any digital artefacts like many of the ones we make daily: a document, a slide deck, an infographic, etc. so it is understandable that many will think of creating an AI system as a similar endeavour, particularly these with no experience of building software. But an AI chatbot is an information system. We are not just "making the Word doc interactive", we are creating an software stack - even if the layer we have written is in English - with which our students are going to interact. As an institution, we have a duty to ensure said information system is *safe* first, *effective* second, and ideally *efficient* third. Implicitly, to make ensuring these easier, it helps if the system is reliable and behaves consistently. As it happens, LLM-based systems have yet to prove themselves on of these criteria, but it behoves us not to compound the problem in our choices as an institution. It is not all bad, though, I appreciate I am pouring scorn by the bucket on these endeavours, but they have had the merit to have taken place with minimal financial outlay and future commitment to technology vendors by the university: as far as I know, a couple of ChatGPT Pro licences. This is, actually, *very good news*. Compare this with Oxford and the LSE having struck deals with OpenAI and Anthropic, Edinburgh's complicated relationship with the former [@roswellEdinburghStaffUrge2026], or the **CHECK** Universities in the Learnwise Pilot **LINK**: Roehampton's slow response has turned an advantage, in that it has not committed itself to anything yet - as far as I know. Though this lack of speed was not out of caution, as once the response arrived it did not shine by the sensibility of its approach, so my fear is this could continue with a contract or partnership announced next year. Although, again for the lack of awareness of available options displayed by leadership thus far, this may very well be upgrading the MS 365 licence to roll out Copilot, which would make for infinitely safer and more reliable student-facing AI, but also for giving staff a safe alternative to uploading, potentially indiscriminately, work data to arbitrary third parties. Still, whether these systems will be efficient for pedagogy, or effective for administrative and teaching work has yet to be evidenced: in fact, recent research found the contrary. As regards pedagogy, because AI systems never refuse to help a learner, aside for safety reasons, they prioritise in their tutoring short-term collaboration, when a human may encourage the learner to persist rather than provide help. The causal impact on learner persistence, and the decrease in unassisted performance because of this, has been evidenced on more than a thousand participants across a variety of tasks [@liuAIAssistanceReduces2026]. A recently developed benchmark for pedagogical safety (not giving the answer away, but guiding the learner to it, correcting misconceptions rather than sycophantically reinforce them) across Maths, Physics and Chemistry AI tutor scenarios found all major foundation models to undermine learning through 2over-disclosure, reinforced misconceptions, and collapsed scaffolding." [@hazraSafeTutorsBenchmarkingPedagogical2026 p. 9]. As to efficiency, a report on a pilot of the very same Microsoft 365 Copilot in the Department of Business and Trade found small time savings, but no productivity gains [@departmentforbusinessandtradeEvaluationM365Copilot2025]. **TODO WORKSLOP** The discourse of AI's inevitability has inflated as the same time as the investment bubble, all the while as the many harms of the technology, to environment, societies, and individuals, came to light. the above are but a small selections of papers offering a much needed factual counter-narrative to the hype. ### Recommendations: A University needs more than AI policies, it needs AI *governance*, that is transparent, democratic, and grounded in evidence and expertise. 1) Decisions should be made after meaningful consideration of the options available. 2) This should be done with domain-expertise on board. 3) This should be done considering the potential for future vendor lock-in, and subsequent rent extraction. 4) This should be done considering alternative options with models running on UR-owned (or at the very least controlled) infrastructure, which beyond vendor dependency is also a matter of data protection, and of visibility/control over energy expenditure. 5) This should be done with student, faculty, and staff representation, in such a way they feel free to raise concern that will be heard and actioned. 6) Initiatives should be piloted transparently, with meaningful evidence and feedback gathered, and findings reported publicly to student, faculty, and staff. 7) Decisions should be informed by available evidence, which is more than sending an email claiming one has read it, but properly citing any literature in which the rationale for the initiative is grounded. 8) The general principle of *not giving the fish*, but teaching angling should inform the philosophy. 9) Pedagogical or administrative benefits of AI systems should either be evidenced before they are adopted ('research' from AI *vendors* is not evidence), or piloted so evidence can be garnered. None of this should be controversial, yet seemingly, it needs outlined.