How to Pilot an AI Research Assistant Within Your University’s AI Policy (2026)
By Sam de Bruyn 10 min read
A university pilots an AI research assistant inside its own AI policy by doing five things in order: putting the tool through the security and privacy approval the institution already runs, writing down which features count as permitted use under the rules already in force, choosing a small cohort with a named academic owner, telling students what to declare and keep, and agreeing the measures that will decide whether the pilot continues. None of that needs a new AI policy. It needs the existing policy applied to one named tool, in writing, before any student uses it.
An AI tool pilot inside a university AI policy is a time limited, formally approved trial in which a named group of students and staff use one tool under the institution’s existing rules on academic integrity, disclosure and data protection, with the measures for judging it agreed before the trial starts.
What does a pilot inside your AI policy look like?
A pilot inside an AI policy looks like five written decisions with a named owner against each: who approved the tool, which features are permitted, who is in the cohort, what students must declare, and what will be measured. Recording those five decisions before the tool is switched on is what separates a pilot from students quietly using a tool anyway.
Jisc’s own pilot of an AI student support platform shows the shape. That project ran between August and December 2024 with twelve institutions taking part, eight further education colleges and four higher education institutions, among them Bath Spa University, Liverpool John Moores University and the University of Westminster. One semester, a defined set of participants, a stated set of questions, a written report at the end.
A pilot that starts with "let’s see how it goes" produces anecdotes instead, and anecdotes cannot be taken to a learning and teaching committee. What to ask a supplier beforehand is a separate question, covered in our guide to what universities should expect from an AI research assistant.
Source: Jisc National Centre for AI, LearnWise end of pilot report, 28 April 2025.
Who has to approve an AI tool before students can use it?
In most universities the academic owner cannot approve a tool alone, because the sign off sits with IT security and privacy too. The University of Auckland is explicit about the sequence for research tools: all digital applications, services and systems undergo a security assessment by Digital Services, a Privacy Impact Assessment comes first where personal information is involved ("Complete a PIA before collecting or using personal information with generative AI tools"), and researchers are told to plan for the University approval process to take three to five weeks. Auckland adds that ethics applications and IT procurement cannot complete until a tool is confirmed as University approved.
Data protection law sets the floor underneath that process. The University of the Witwatersrand’s AI policy document (G-C-2026-030) states in section 4.5, on data security, that "Sensitive and / or personal data should not be entered into public AI systems", and section 4.2 holds that the person using the tool "remains fully responsible for the originality, accuracy, and integrity of their work".
As of October 2026, three to five weeks is the planning number that matters most here. A pilot aimed at a cohort starting in February has to begin its paperwork in December.
Sources: University of Auckland, generative AI tools for research, University of the Witwatersrand AI Policy Document G-C-2026-030.
How do you map a tool’s features against your AI policy?
Map the tool feature by feature rather than as a single yes or no, because one product usually spans several categories your policy already treats differently. The method is a four column table, filled in by the academic owner and signed off by whoever owns academic integrity: each row names a feature, what the student is actually doing, the question the existing policy already answers about it, and the record kept if it is permitted.
| Feature | What the student is actually doing | Question your existing policy answers | Record kept if permitted |
|---|---|---|---|
| Finding papers | Searching for literature to read | Is search assistance permitted, and does it need declaring? | Search terms and dates run |
| Summarising uploaded papers | Reading support on sources the student chose | Is reading support permitted on the student's own sources? | Which papers, which tool |
| Answering questions with cited passages | Locating and checking what a source says | Does a citation to the student's own PDF satisfy your referencing rules? | The passage behind each claim |
| Drafting text | Producing prose that could be submitted | Is drafted text permitted, and at what level of declaration? | Prompt log and dated drafts |
| Checking citations | Verifying that references exist and are correct | Is verification permitted without declaration? | The checker output |
Stellenbosch University’s Faculty of Science guidelines for students, version January 2026, provide a sourced way to label the rows. The guidelines split use into weak and strong: tools in the weak category "do not necessarily need to be declared, although declaration is recommended", while strong use "should be fully disclosed, including the purpose for which it was used" and "counts as plagiarism/AI-giarism if not declared". The guidelines also give the tie breaker for rows you cannot classify: "When in doubt, treat the use as a strong use to be declared."
For a published scale rather than a local one, the revised AI Assessment Scale by Mike Perkins, Jasper Roe and Leon Furze, dated December 2024, sets out five levels: No AI, AI-Assisted Planning, AI-Assisted Task Completion, Full AI and AI Exploration. Its authors are clear that institutions "should adapt it to their specific contexts and needs".
Sources: Stellenbosch University Faculty of Science, Guidelines for using GenAI tools in Learning and Assessment, January 2026, Perkins, Roe and Furze, The AI Assessment Scale Revisited, December 2024.
Who should be in the pilot cohort, and how long should it run?
Pick a cohort with a shared deadline and a single supervisor or programme coordinator, and run the pilot across a full semester rather than a few weeks. A cohort with a shared deadline gives comparable work to look at, and a single academic owner means someone can answer questions in week three instead of referring them to a committee.
The argument for a full semester comes out of a pilot that was shorter. Jisc ran an AI student support platform across twelve institutions between August and December 2024, and reported that interaction levels stayed limited, which the end of pilot report attributes largely to the short duration of the project. Participants also asked for clearer communication about offboarding at the end. By the close, eight of the twelve institutions had committed to or were exploring paid subscriptions. The honest caveat is that the tool there was a student support platform, not a research assistant, so what carries over is the pilot design rather than the product.
Size matters less than comparability. One postgraduate cohort working to a single submission date will tell you more than several hundred volunteers scattered across faculties, because you can read a sample of the finished work and ask one supervisor what changed.
Source: Jisc National Centre for AI, LearnWise end of pilot report, 28 April 2025.
What should students be asked to keep as a record?
Ask for three things: a declaration in the submission itself, a log of prompts and outputs available on request, and dated drafts kept until after feedback. Those three cover almost every disclosure rule currently in force, and asking for them from day one costs students very little.
The University of Suffolk’s guidance shows what that looks like written down. From the 2026/27 academic year, students there must include an AI declaration as part of every assignment submission, and a declaration is required even when no AI was used. Suffolk’s templates are plain: "I used [insert tool name] to [insert description of AI function]." Transcripts and prompt logs go in an appendix or are made available on request, and students are advised to keep notes, supplementary materials and draft versions until after they have received feedback.
For postgraduate researchers, Stellenbosch’s Faculty of Science guidelines go further and treat the log as data: "detailed records of the prompt history should be downloaded and saved as research data." Wits asks for declarations to be made to the University through a comprehensive template, so a pilot there inherits a disclosure format rather than inventing one. Students who have never written one can start from our guide to writing an AI use disclosure statement, and the dated drafts Suffolk asks for are the evidence described in how to prove you wrote your own work.
Sources: University of Suffolk, Acknowledging AI use, updated 16 September 2026, Stellenbosch University Faculty of Science guidelines, January 2026, Wits AI Policy Document G-C-2026-030.
How do you judge whether the pilot helped?
Agree three measures before the pilot starts: citation accuracy in a sample of the work, supervisor judgement of whether the students read more carefully, and a short student survey. Jisc evaluated its own AI pilot with focus groups across all participants, institutional usage statistics and an end user survey distributed where possible, which is a reasonable template to copy. Detection software is not one of the three, for reasons set out in why process based evaluation beats AI detection.
Citation accuracy is the measure that can be checked rather than felt, and there is a published method for it. In a study in JMIR Mental Health on 12 November 2025, Jake Linardon and colleagues extracted 176 citations generated by GPT-4o across six literature reviews and verified each against Google Scholar, Scopus, PubMed, WorldCat and publisher databases. Thirty five of the 176, 19.9%, were fabricated outright, and among the 141 real citations, 64, or 45.4%, contained errors. A pilot can run a smaller version of that check: pull a sample of references from the work students produced and verify each against a database.
Where a tool cites only the PDFs a student uploaded, the check is different and faster: open each citation and read the passage behind it. The claim and the passage either match or they do not, and a supervisor can test that in minutes. Our comparison of the free tools that check whether references are real sets out what each database catches.
One caution about published pilot results, including the hopeful ones. The pilot implementation of the AI Assessment Scale at British University Vietnam, reported by Leon Furze, Mike Perkins, Jasper Roe and Jason MacVaugh in the Australasian Journal of Educational Technology on 16 October 2024, states that results "indicate a significant reduction in academic misconduct cases related to GenAI". That is a direction, not a figure you can transplant into your own business case, so treat it as a reason to count misconduct cases yourself.
Sources: Linardon et al., JMIR Mental Health, 12 November 2025, Furze, Perkins, Roe and MacVaugh, Australasian Journal of Educational Technology 40(4), 16 October 2024, Jisc LearnWise end of pilot report.
What should a pilot never claim about a tool?
A pilot should never describe a tool as compliant or approved in general terms, because institutional approval is specific, conditional and time limited. The University of Auckland states it bluntly: "Previous or current use of a tool does not mean it is currently University approved." Auckland also records that Digital Services’ approval runs for a period of time aligned with the University’s data classification, and may carry conditions such as being limited to one research project or to prescribed settings.
Borrow that phrasing for the pilot’s own paperwork. "Approved for this cohort, for this purpose, until this date" is a sentence a university can defend. "Approved" on its own outlives the conditions attached to it, and tends to reappear in a prospectus two years later.
The same restraint applies to suppliers, including us. No supplier can tell a university that a product meets that university’s AI policy, because the policy is the institution’s to interpret and the use is the student’s to declare.
Source: University of Auckland, generative AI tools for research.
Frequently asked questions
Do we need an AI policy before we can pilot an AI research tool?
No. A pilot needs the rules you already have, applied to one named tool and written down. Most universities already hold policies on academic integrity, plagiarism, data protection and IT security, and the pilot mapping document is where those rules meet a specific product and a specific cohort. A pilot is also one of the better ways to find out what a standalone AI policy should eventually say.
Who signs off an AI tool, the academic owner or IT?
Both, in most institutions, with privacy alongside them. At the University of Auckland, digital applications undergo a security assessment by Digital Services, a Privacy Impact Assessment is required before personal information is used with generative AI tools, and researchers are advised to plan for three to five weeks. Treat the academic case and the security and privacy review as parallel tracks starting the same day.
What should students declare during a pilot?
Which tool they used, what they used it for, and where the output appears in the work. The University of Suffolk's template wording is "I used [insert tool name] to [insert description of AI function]", with transcripts or prompt logs supplied in an appendix or on request. Stellenbosch's guidelines add the default to apply when a case is unclear: "When in doubt, treat the use as a strong use to be declared."
How do we check whether a tool's citations are real?
Take a sample and verify each reference against a database such as Crossref, Scopus, PubMed or Google Scholar, the method used in the published literature on citation fabrication. Where a tool cites only the papers the student uploaded, open each citation and compare the claim against the passage it points to. A supervisor can run both checks on a handful of submissions without special software.
Can a pilot tell us whether students learned more?
Not on its own, and claiming otherwise will not survive a committee. A single semester pilot can measure citation accuracy, usage, supervisor judgement and student-reported experience, which is enough to decide whether to continue, extend or stop. Learning gains need a comparison group, which is a separate piece of work from a procurement pilot.
Where Litrevu Fits in a Pilot Like This
Litrevu is an AI research assistant that reads the papers a researcher adds, answers questions about them, and cites the passage behind each answer. For a pilot the part that matters is checkability: a student adds their own papers, and every citation points back to a passage in one of them, so a supervisor can open a claim and read the sentence it came from.
Our position with universities, said plainly: we would rather shape Litrevu around your policy than ask you to fit around ours. Today that means agreeing together which features your students use during a pilot and which they do not, and keeping that mapping in your documentation rather than ours. Litrevu can also draft cited text, and many universities do not allow AI-written text in submitted work even after editing, so drafting is a row in your mapping table like any other.
Here is the shape of a cited answer, so you can see what a supervisor would be checking:
Question asked of the uploaded library: What do these papers say about [topic]?
Answer: [Statement of what the uploaded papers say] (Author, Year). [Second statement] (Author, Year). Where the uploaded papers do not cover a question, Litrevu says so.
Opening the citation (Author, Year) shows: [the passage from the uploaded PDF, word for word, with the paper it came from]
That block is a format illustration, not output from a real query. The bracketed text marks where content from a researcher’s own papers appears, and it deliberately contains no findings, authors or numbers.
As of October 2026, Litrevu is free to start with 25 messages and 2 drafted sections every 30 days, no card needed, which is enough for a librarian or supervisor to upload a few papers and judge the citation behaviour before committing a cohort to anything.
Try It on Your Own PapersRelated Guides
- The Ethics of AI in Research Writing: What Universities Should Expect From an AI Research Assistant
- How to Write an AI-Use Disclosure Statement for Your Assignment
- AI Detection Is Dead. Universities Are Switching to Process-Based Evaluation.
- Best Tools to Check If Your References Are Real
- Litrevu for Universities