Guidance on AI use in research and innovation

This guidance is designed to support researchers to use artificial intelligence (AI) responsibly and effectively in their research.  Additional guidance is available for PhD students on Responsible Use of Generative AI in Doctoral Research.

This guidance will be updated as required by the University working group on AI in Research and Innovation to reflect changes in the development of AI tools and sector norms.  

What is AI?

AI is the capability of computational systems to perform cognitive functions that we normally associate with human minds. AI models underpin generative AI (GenAI) tools, which is an umbrella term for an AI tool that creates new content. Such tools include large language models (LLMs) such as CoPilot or ChatGPT that can generate text, sound, images and video.

Related terms are machine learning, a subfield of AI that enables systems to learn patterns from data without explicit programming, and deep learning, a subset of machine learning that uses multi-layer “neural networks” to perform data analysis in an attempt to mimic human brain function. A broad term for machine learning and deep learning models used when the model is developed solely from data (without the use of physical principles) is data-driven models. These models are typically employed for approaches such as regression, classification, or predictive processes to assist with decision making.     

Types of AI usage in research and innovation

There are several ways that AI can be used in research and innovation. Two broad classes are as follows:

1. AI use to support research: examples include (i) discovery/understanding of the research landscape through distillation and summarisation journal articles and other material; (ii) refining inputs (text, images, sound), including adapting text to different audiences such as for teaching or the general public, and improving clarity and accessibility; (iii) data cleansing and other data processing activities; (iv) supporting code writing and debugging; and (v) brainstorming and planning tasks. These tasks typically use GenAI tools with the aim of improving efficiency. 

2. AI use as a fundamental component of the research or as a research tool: examples include the use of data-driven tools such as neural networks to analyse research data, verification of the outputs of machine learning tools, monitoring and execution of research and experimental plans, sociological research that explores how people use AI tools, explainable AI (XAI) research that aims to improve understanding and trust in the outcomes of AI models, computer vision algorithms that analyse images and videos for object detection and face recognition, and development of new AI tools or models including ones that can be used to narrow down options before starting in-depth research or sift data to find patterns that highlight promising research paths.   

Sustainability and environmental considerations

The training and operation of AI models, including the LLMs used in GenAI tools, is typically carried out in vast data centres that require significant energy and use water for cooling. Rare earth minerals are also required to produce the computer chips and other hardware with their own environmental costs. While it is difficult to collate accurate data on environmental impacts, one estimate is that the most common type of ChatGPT query may use 0.0029 kWh of energy, 30 ml of water and lead to the emission of 0.69 g of CO₂; as a comparison, the same article states that Streaming Netflix in HD for one hour uses more than 20 times as much energy (around 0.077 kWh) [https://nationalcentreforai.jiscinvolve.org/wp/2025/05/02/artificial-intelligence-and-the-environment-putting-the-numbers-into-perspective/]. Another estimate describes a single ChatGPT query as requiring the same amount of energy as running an oven for a little over 1s, and about one-fifteenth of a teaspoon of water [https://www.businessinsider.com/how-much-energy-does-chatgpt-use-average-query-watts-altman-2025-6].  Online AI carbon footprint calculators also exist.

Researchers should consider the University of Reading’s strategic focus on sustainability when using AI tools.

Ways to limit the cost of GenAI tool use includes careful writing of prompts, requesting text rather than image outputs, avoiding repeated similar requests for small gains (e.g., stylistic improvements), limiting the length of conversations, and limiting usage of large GenAI models e.g., considering whether a GenAI tool is needed and not using a GenAI tool when a non-GenAI based search engine will provide the same information. Related guidance applies to the use of other (non-GenAI) AI tools: consider whether AI use is necessary, use the smallest possible AI model, and use transfer learning to fine-tune existing models rather than training new models from scratch. Open-source sharing of models and code can also reduce environmental costs. Some vendors have also made sustainability commitments e.g., Microsoft. 

The environmental costs of the use of AI tools should be balanced against their potential benefits including more efficient ways of working, the rapid analysis of large datasets, and enhanced pattern recognition and modelling of complex systems. The environmental cost of using pre-trained data-driven models for inference can be substantially less than that of alternatives e.g., physics-based models. AI tools can also contribute to research that has socio-economic benefits.  

Ethics and integrity considerations

The use of AI tools raises new ethical considerations. These tools should support the judgement of researchers, rather than replace it, with mandatory human oversight. As with all research, research using AI tools must be traceable and reproducible and, where possible, consistent with open research principles. Consider using explainable AI approaches to interpret AI-derived outcomes and quantify uncertainty. Researchers should consider possible bias in AI training sets and AI use should be disclosed and acknowledged in research outputs (following guidance from the relevant publisher where available). Projects developing AI models and applications for innovation and commercialisation should also support open research principles with traceability and reproducibility where appropriate. At the same time, researchers should consider the protection of intellectual property of AI innovation components, including datasets, trained models, and software development, to enable effective translation, innovation commercialisation, and impact. Some research contracts may include constraints on where AI tools can be used or developed for the associated research projects.

Researchers must follow policies on intellectual property rights  and data protection. Intellectual property rights considerations include sources of training data: was it obtained under license, from open source, or instead they scraped in breach of copyright? Data protection considerations include ensuring that data is only shared with commercial GenAI tools where appropriate and that those tools are setup to prevent them from using the data for training if required.     

Researchers should seek ethical approval if personal data of any kind is to be used or there are other ethical concerns.

Ways to use GenAI tools to support your research

Used appropriately GenAI tools can improve research efficiency and aid creativity. GenAI tools can be used to:

1. Plan: help generate ideas and frameworks, and summarize relevant information e.g., from a set of published articles.

2. Improve: refine and polish initial drafts including tailoring for specific audiences, restructuring and improving conciseness, support translation between languages, comment or optimize model code, and simplify and summarise technical material.

3. Co-develop: generate first draft text or model code, help debug model code, and contrast viewpoints or data. 

The Research Funders Policy Group statement on the use of Generative AI tools in funding applications and assessment states that AI tools must be used responsibly when developing funding proposals and should be acknowledged in any outputs. This policy group includes many of our major funders e,g., UKRI, Wellcome and the Royal Society; see also the UKRI specific guidance. These principles should also be adopted in the development of applications for internal university funding calls.

Many journals and publishers also have their own policies on the use of AI e.g., Springer and Wiley. While the specific policy should be checked for the intended publisher of an output, a common theme is that GenAI tools do not currently satisfy the authorship criteria and human accountability is essential for the final version of the text. The use of GenAI images is typically prohibited with a few use exceptions.    

While it is typically not a requirement to report use of AI tools to assist copy editing, there may be a requirement by funders and publishers (and their journals), to report some types of usage of AI tools in research outputs and proposals. 

Doctoral research students should also refer to the Guidance on the Responsible Use of Generative AI in Doctoral Research.

Use of GenAI tools in research assessment

The policies of funders and publishers (and their associated journals) also address the use of AI by peer reviewers. Peer reviewers are accountable for the accuracy and views expressed in their reports, there are known limitations of AI tools, and there may be sensitive and proprietary information in proposals or outputs/papers that should not be shared with remote cloud-based AI tools. Consequently, reviewers are typically instructed not to use AI tools for research assessment. Peer reviewers should also consider whether the material that they review contains errors or inconsistencies that suggest that AI tools may have been inappropriately used in content generation.  

The use of AI tools in the preparation of the university’s submission to the Research Excellence Framework exercise is addressed in the University of Reading REF2029 code of practice (further information will be added when available).

Risks and limitations of GenAI tools

While there are potential benefits from the use of GenAI tools, users must also consider their limitations and the associated risks. Well-known risks related to their use in research include the hallucination or fabrication of material (plausible but incorrect information may be generated such as fake citations), biases created by the choice of data on which the model is trained, sensitivity of the outputs to the details of the prompt entered, and the risk of violating copyright or other policies by entering material into remote cloud-based AI tools. As stated above, these tools should support the judgement of researchers, rather than replace it, with mandatory human oversight. See also the section on Ethics and integrity considerations.

University approved GenAI tools

The University recommends the use of Microsoft CoPilot Chat (as part of our M365 licence) for day-to-day use queries, productivity enhancement and creativity because the Terms and Conditions of use do meet the necessary requirements of data protection laws and have, therefore, been approved by the University’s Legal, IMPS and Digital teams. However, you should not use it for any data set containing personal or personal sensitive data.