Generative AI and Large Language Models in research

Purpose and scope

The guidance on this page is for the responsible use of Generative AI (GenAI) and Large Language Models (LLMs) during the research process. GenAI and LLMs can assist at different stages of the research process, including ideation, literature review, grant writing, data analysis, coding, and administrative tasks. However, the challenges and risks vary significantly depending on the specific application. This guidance sets out key principles and practices to help everyone who is engaging with research activities to use GenAI and LLMs ethically, responsibly, and proportionate to the risks associated with each task.

After reading this guidance, you should be able to:

  • identify and mitigate risks to accuracy, reliability, data protection, confidentiality, and intellectual property.
  • critically assess when and how it is appropriate to use GenAI and LLMs.
  • document and disclose GenAI and LLM use transparently in grant applications, publications and other written outputs, and project records.
  • make informed decisions about whether to use these tools or avoid them entirely in particular contexts.

Important: While this guidance helps identify potential challenges and considerations when using GenAI and LLM tools and how to mitigate them, it is individuals who are ultimately responsible for the appropriate use of GenAI and LLM tools and their outputs. Both internal UWE Bristol policy and national laws governing AI use (copyright, GDPR, AI Act) are rapidly evolving. Anyone engaging with research activities should regularly check for updates and ensure compliance with current regulations.

Context

The recent years have seen a quick expansion of Generative AI (GenAI), producing creative content, and Large Language Models (LLM), which use natural language processing (NLP) to understand and generate humanlike text-based content. These tools are used to create new text, images, code, video, or audio (e.g. ChatGPT, Copilot).

GenAI and LLM can present opportunities for research, they can accelerate certain processes, save time, and advance discoveries, however, they can also present substantial risks that cannot yet be fully anticipated or comprehended today. In line with The Concordat to Support Research Integrity (PDF), any opportunities presented by GenAI and LLM must be balanced against the need to conduct research activities with high integrity in order to uphold honesty, accountability, rigour, transparency, and respect. Improper, unreflective use of GenAI and LLMs can lead not only to unethical, poor-quality research but ultimately to loss of trust in research, and it can have personal, societal and environmental implications.

Tools

UWE Bristol has approved Microsoft Copilot as the designated LLM tool. Please make sure you are using the Enterprise version of Copilot. You can confirm this by checking that you are on the Work tab and that the green shield with a tick is visible.

Important: While Copilot provides a more secure environment than public LLMs, challenges such as hallucinations, bias, incomplete data coverage, and explainability limitations remain (see below). All outputs must still be critically evaluated.

Key challenges and risks

Quality and accuracy

Hallucinations and reasoning

  • GenAI and LLMs write convincingly but may present incorrect information as fact. Because they are ‘statistically trained on language’ rather than trained on any ‘ground truth’, they cannot differentiate between correct and incorrect information and present all outputs as equally valid.
  • Accordingly, when there is insufficient training data to inform an answer, GenAI and LLMs make things up (this is known as ‘hallucinations’). For example, they create false references, which may look authentic, but where the underlying text does not exist. 
  • Similarly, while ‘reasoning’ LLMs can break down queries into smaller components to make logical leaps when asked to solve a given problem, this is not the same as the deductive reasoning or intuition used by humans.

Limited and biased data resources

  • Most LLMs are primarily trained on openly available information with limited access to peer-reviewed journals, meaning that their understanding of more complex academic literature and concepts is limited. 
  • GenAI and LLMs that are not web-connected may have a cut-off date on the data with which they have been trained. They may therefore be using incomplete data set or information which is not up-to-date.
  • Training GenAI models on poorly curated datasets creates instances of ‘data poisoning’ , or ‘false data injection’, as do attacks on GenAI systems, putting reliability of GenAI systems at risk. 

Therefore, researchers should:

  • always cross-check outputs against authoritative sources.
  • never rely on Copilot or any LLM to generate references or evidence without independent verification.
  • use LLM outputs as a starting point, not as definitive research material.

Reliability

Reproducibility and explainability

  • LLMs are ‘black-box’ mathematical models of language, which means that their outputs are not always fully explainable. The rules by which a given input results in a given output are not transparent, which can therefore limit the degree to which systematic errors, biases, or flawed reasoning can be identified and corrected in the research process.
  • In addition, due to variabilities in input conditions (e.g. user profile, conversation history, model settings, system updates), LLMs can produce different responses to similar or identical inputs. This non-deterministic behaviour makes replication of experiments or LLM use across different users difficult, undermining fundamental principles of peer review and academic reproducibility.

Plagiarism

Copyright and attribution

  • AI companies have been scraping the internet to find data to train their models, using copyrighted material, personal data and trademarks. GenAI and LLM re-present information written by others. Without the means of assigning credit to other people’s work, researchers are at risk of plagiarism (text, image, code etc.) and copyright infringement. The legal responsibility for infringement is likely to lie with the user and not the GenAI/LLM tool.
  • GenAI and LLM providers may claim ownership over your inputs and outputs and can use the data, code, videos, images etc. for model training purposes and even sell them to third parties for data mining purposes.

Therefore, researchers should:

  • appropriately acknowledge all use of Copilot or any other LLM (see referencing format in Referencing AI use section below).
  • never present AI-generated material as wholly original research.
  • check outputs and sources carefully to avoid unintentional plagiarism.

Ethics and bias

Systemic bias

  • Training data for GenAI and LLMs reflects historical inequalities, cultural prejudices, and systemic discrimination present in source materials. Since models are trained on vast internet content, they inevitably incorporate and may amplify societal biases related to gender, race, ethnicity, religion, and other protected characteristics. The opacity of training datasets makes it difficult to identify or mitigate specific biases. 
  • Bias can have serious negative implications, including further marginalisation of underrepresented groups, perpetuation of harmful stereotypes, and reinforcement of existing inequalities in research findings. 
  • Prompt sensitivity: Specific terminology and framing significantly influence outputs. Seemingly neutral language can trigger biased responses, while those engaging in research may unknowingly introduce bias through prompt design.

Environmental and social impact

  • The environmental impact of GenAI and LLM data centres is significant, particularly at a local level, as they require large amounts of energy and water. The rapid turnover of hardware also generates electronic waste and raises concerns about environmental contamination.
  • Many GenAI systems depend on large-scale data labelling, which is often carried out by low-paid workers under precarious conditions, raising ethical concerns around fairness and working standards.
  • Publicly funded or community-created datasets are frequently repurposed by commercial providers without adequate recognition, compensation, or reinvestment into the original communities.
  • Both intentional and unintentional misuse of GenAI systems can amplify misinformation, enable unethical applications, or lead to societal and environmental harm through misinterpretation or poor oversight.

Therefore, researchers should:

  • critically evaluate outputs for bias and unfair assumptions.
  • refrain from using GenAI and LLMs for sensitive topics without additional safeguards.
  • consider environmental and social impacts of AI use when choosing methods.

Privacy, intellectual property and data protection

GDPR compliance

  • Under the UK Data Protection Act (2018), derived from GDPR 2016, individuals have rights to rectification, erasure ("right to be forgotten"), and data portability that LLMs cannot technically fulfil once personal data is embedded in training models. 
  • LLM providers often cannot identify what personal data was used in training or ensure complete deletion from trained models. Cross-border data processing may trigger additional transfer restrictions under GDPR Chapter V.

Confidentiality and IP risks

  • Personal, sensitive, confidential, or proprietary information should never be input into public GenAI and LLM tools. This includes:
    • Research participant data
    • Unpublished research findings or methodologies
    • Grant application details or funding strategies
    • Commercially sensitive information
    • Personal data of colleagues, students, or collaborators
  • Cross-system data integration may enable re-identification of supposedly anonymous individuals when LLM outputs are combined with other datasets. 
  • LLM-generated outputs may inadvertently contain personal data - researchers remain responsible for identifying and protecting such information under data protection principles. 
  • Research ideas, methodologies, or preliminary findings input into public LLMs may be deemed as ‘public disclosure’ and therefore compromise patent applications.

Therefore, researchers should:

  • reduce/never input identifiable participant data, unpublished findings, or proprietary information into any LLM.
  • review data ownership, storage, and retention policies before inputting content.

Decision-making framework

To help you consider whether to use GenAI or LLMs, it is worth considering these critical questions:

Purpose and necessity

  • Why am I using AI for this task? Is it essential, or could I accomplish it another way?
  • Will this use enhance the rigour and usefulness of my research, or does it compromise it? 
  • Am I actually saving time, considering the need to verify outputs?

Transparency and accountability

  • How will I document and disclose my AI use in publications, grant applications, or reports?
  • Would I be comfortable explaining my AI use to reviewers, funders, or ethics committees?
  • Am I contributing original ideas and critical thinking, or am I depending on the AI to generate ideas in my place?

Verification and reliability

  • How will I check the accuracy of AI outputs?
  • Can I reproduce consistent results?
  • Do I have the expertise to validate the AI-generated content?

Ethics and legal compliance

  • Am I respecting institutional, funder, and legal requirements for copyright, privacy, and data protection?
  • Could my AI use introduce hidden bias or reinforce inequities?

Confidentiality and security

  • Am I exposing unpublished research, participant data, or proprietary information to risk?
  • Have I reviewed the tool's terms and conditions regarding data ownership and storage?

If in doubt, GenAI or LLMs should not be used for:

  • processing personal, sensitive, or confidential data.
  • tasks requiring original critical analysis or novel theoretical insights.
  • when you cannot adequately verify the accuracy of outputs.
  • situations where bias could significantly impact research validity or ethics.
  • contexts where institutional policies explicitly prohibit use.

Referencing AI use

Increasing disclosure requirements

Academic institutions, research funders, and publishers are rapidly implementing mandatory disclosure requirements for LLM use. What was once optional is becoming required across the research ecosystem:

  • Research funders: UKRI, Wellcome Trust, ERC, and other major funders now require explicit declaration of AI assistance in grant applications and reports.
  • Academic journals: Major publishers (Elsevier, Springer Nature, Wiley) have implemented AI disclosure requirements as part of submission processes, with potential rejection for non-disclosure.
  • Institutional policies: Universities increasingly mandate AI use disclosure in thesis submissions, ethics applications, and research progress reports.
  • Professional standards: Academic societies are establishing AI disclosure guidelines with potential disciplinary consequences for non-compliance.

How to reference

Any material wholly or partially produced by LLMs must be appropriately referenced. Presenting LLM-generated responses as your own work constitutes academic misconduct and could result in disciplinary action or funding withdrawal. It is therefore key to reference outputs appropriately.

Current best practice treats LLMs as private correspondence. While referencing styles are still evolving, all citations must include:

  • the specific platform and model used.
  • date of interaction.
  • your name as the recipient.
  • a brief description of the task where relevant.

Citation examples

Harvard style:

  • In-text: (Microsoft Copilot, 2025)
  • Reference list: Microsoft Copilot. 2025. Microsoft Copilot Response to [Your Name], 26 September 2025. 

How to record your AI prompt use

  • Date/Time: [DD/MM/YYYY, HH:MM]
  • Platform/Model: [e.g., Microsoft Copilot]
  • User: [Your Name]
  • Project: [Research title/reference]
  • Task Purpose: [e.g., Literature synthesis, brainstorming, coding assistance]
  • Initial Prompt: [Full text of your query]
  • Key Follow-up Prompts: [Additional questions asked]
  • Output Summary: [What the AI provided]
  • How Used: [Direct use, modified, inspiration only]
  • Verification Method: [How you checked accuracy/relevance]
  • Human Contribution: [Your modifications/additions]

Sources used and further reading

For more information and views on AI and research integrity, you can also refer to AI working group bibliography (PDF) prepared by the UK Committee on Research Integrity (it includes academic and opinion pieces written up to October 2024). 

You may also be interested in

Researchers look at a screen

UWE Bristol Code of Good Research Conduct

UWE Bristol has set in place a Code of Good Research Conduct and a Policy on Good Research Conduct to promote good research practice, and prevent research misconduct.

Researcher looking at tissue under microscope

Human tissue research and teaching at UWE Bristol

Information on human tissue in research and teaching at UWE Bristol.

Bee sitting on a flower

Policy statement on the use of animals in research

UWE Bristol strongly supports research that meets the highest ethical standards, including that which involves animals.

Researcher looks at data on a computer

Research data management

Important information for researchers at UWE Bristol about research data management, including our official policy.