FAQs for the policy on generative AI in research

As part of determining the methodology for using GenAI in your research, you should first research its use in your discipline or field of enquiry, especially if it is likely that standards have been developed for your discipline. The University’s Generative AI SIG on Teams run by the AI Competency Centre is a source for staying up to date with emerging practice (expectation 6).

The use of GenAI should also be planned in when developing a research funding proposal, together with determining associated costs. Most funders will cover these costs, if justified in the application.

When your research involves personal data, the use of GenAI must be specified both in the proposal and research ethics application, to ensure compliance with the University’s research ethics policy.

In line with expectation 7, you are recommended to first consider the option of running a third-party tool locally or to download a model from a community platform such as Hugging Face. The advantages of running GenAI models locally are that they are less energy-intensive, the data stays within the University and use can be replicated by rebuilding the model, enhancing reproducibility. It is important to read the relevant terms and conditions when downloading a community model, to ensure that your use will not result in a breach and to be aware that a version that can be run locally may not be as powerful or sophisticated as a cloud-hosted version.

If a large cloud-hosted model offers the best fit for purpose, you are strongly recommended to first investigate the suitability of GenAI tools supported by, or offered through, the University via the AI Competency Centre, especially those that protect data used and generated, and offer the ability to collaborate with other users across the University. All staff and students have free access to ChatGPT Edu, governed by an agreement between the University and OpenAI

Using GenAI to help you with your own written work would often not be considered substantive use, if the use could as easily apply to documents that are not research outputs, e.g. to construct summaries, to reduce word count, to improve readability or standard of one’s own English or foreign language. However, if use includes incorporating work authored by someone else, care needs to be taken not to infringe copyright (expectation 2).

If you are applying for external funding, you should comply with the relevant funder guidelines/policies; UKRI, Wellcome and CRUK have issued guidance.

If you are applying to an internal fund, e.g. the John Fell Fund or funds run by your division and/or department, you are permitted to use GenAI in developing your application as long as its use is declared in the application. You are only required to declare substantive use (see policy on use of GenAI in research for definition of substantive use).

Please check relevant funder guidance; UKRI, for example, do not permit the use of generative AI during interview.

If you have an application interview at the University either for an internal fund, or as part of a selection process for an external fund that only permits a restricted number of applications from Oxford, you are permitted to use GenAI to prepare for interview, but not for generating responses in a live interview. There is no need to declare this use, but it is important to check the AI generated responses (expectation 5).

Good practice for using GenAI is context dependent, both in terms of the research purpose of using GenAI, the functionality and limitations of the relevant tool and the acceptability of its use in your discipline. However, you should always seek to understand the limitations of the tool you have selected, validate the GenAI outputs, document use and declare use when sharing outputs.

What should I do to ensure the quality of the GenAI outputs?

To ensure research quality, you should validate GenAI outputs similarly to how you would check other types of outputs in your field, e.g. work undertaken by other researchers, software code, instrument measurements, to satisfy yourself that the outputs can be trusted (expectation 4).

Iterative prompting, by systematically refining and adjusting the prompts, can also improve the quality of outputs: the precision of a prompt can significantly influence the quality of the analysis and iterative prompting can result in better alignment of the GenAI tool with research objectives. However, it is important to be aware that the outputs of GenAI tools tend to converge to your line of enquiry, the more the number of iterations increases (expectation 4).

How do I document interactions with GenAI?

To establish an audit trail of how GenAI influenced the research process, substantive use of GenAI should be documented in research notebooks or equivalent, in line with good research practice. This would typically involve a complete transcript of the output without any editing, date of production, version of the tool and complete set of documents that were uploaded (not just links to documents) (expectation 4).

Can I use GenAI to edit images?

Where images are research results that are analysed for insights, data and to evidence findings, using GenAI to edit images is inappropriate and would constitute a breach of good research conduct.

Declarations of use of GenAI should follow the same principles as transparency and openness around text, data, software, and other materials associated with research. As replication is often not possible with cloud-hosted third-party models, which are continuously being updated with new versions, and the inherently probabilistic nature of GenAI tools, declarations should at a minimum include:

• Name and version of the GenAI tool used; e.g. Microsoft Copilot (version GPT-4); ChatGPT-3.5;
• Publisher (company that made the GenAI tool); e.g. Microsoft; OpenAI;
• URL of the GenAI system;
• Brief description (single sentence) of context in which the tool was used

(taken from UCL Library) (expectation 4)

You are encouraged to share the policy and use cases with your collaborators so they understand the expectations from the University; it is also possible that the policy applies to them if they will use University facilities. 

If your research involves personal data and is therefore subject to the University’s research ethics policy, the expectation is that you ensure your collaborators comply with University policies.

Outputs produced with the use of GenAI tools are subject to the same policies as all outputs; see the University’s Open Research Position Statement, which refers to the Open Access Publication Policy for scholarly articles, and the Research Data Management Policy for other digital objects (e.g. datasets, software, models) (expectation 5).

Most publishers have guidelines or rules about use of GenAI in research, e.g. use of GenAI must be declared and clearly explained, GenAI tools cannot be listed as an author, authors are accountable for the accuracy, integrity and originality of their research, and the research must be the author’s own, i.e. not breach plagiarism rules. You should familiarise yourself with publisher rules when selecting in which journal or with which publisher to publish.

The same procedures apply when discovering any type of error that affects published research. You should contact the editor of the relevant journal, in the first instance, with details of the error. If the error is relatively minor, i.e. does not invalidate the research published, the publisher will either issue a correction or correct the online version. If the error invalidates the published research, the publication will have to be retracted.

Some uses of GenAI by professional staff will not constitute substantive use, although others will require similar consideration when using GenAI to produce research proposals or research outputs, especially when using documents authored by a colleague (expectations 2 and 3), but also in terms of validation and accountability (expectations 1 and 5).

Examples are:

Substantive use

• Summarising research, reports, other documents, and data
• Improving analysis and workflows: analysing data to identify patterns, trends, and process inefficiencies, and to resolve these; automating tasks and generating content
• Starting points for content, e.g. extracting key points from a technical paper, generating ‘human’ readable versions of technical or scientific documents and papers as a starting point for work
• Drafting minutes using transcripts
• Creating policy briefs
• Reformatting information for specific uses

Non-substantive use

• Reducing word count and eliminating unnecessary text, including editing, grammar, sense checking and rephrasing text
• Finding specific quotes in documents, e.g. to identify actions, to include in a paper or publication
• Converting documents to PowerPoint slides
• Improving web content and proofreading
• Workshop planning and presentations

Most of these examples were taken from the AI in Communications at Oxford: Workshop Summary; the workshop was held on 2 April 2024 for communications professionals at Oxford.

The University’s agreement with OpenAI managed by the AI Competency Centre and which provides free access to ChatGPT Edu for staff and students enables the development of Custom GPTs to assist researchers. Custom GPTs can help with aligning research applications with funder guidance, sector expectations, regulations, or by giving researchers more targeted advice or guidance on how to improve their application using funder evaluation criteria or past successful applications. By using an enterprise GenAI tool, where the University has an agreement with the supplier, you ensure that any information uploaded stays within the University, while the collaboration functionality enables you to share the Custom GPT with researchers.

When supporting researchers, you should check with the relevant researcher(s) about any restrictions to the use of GenAI in research, and familiarise yourself with relevant funder and publisher policies.

GenAI should not be used to rank research proposals and make funding decisions. This does not obviate the regular analysis of research management data to track trends or benchmarking.