This blog is part of the Good practice in research software development in the age of AI series.
Data governance refers to the policies and rules around the use, management, access and sharing of data. As researchers, we need to be aware of the ethical norms, laws and policies that govern our use of AI tools with our research data. Relevant policies include institutional policies, funder policies, journal policies and local research group policies on the use of AI in research.
As well as knowing and understanding the ethico-legal landscape we are working within, we also need to be aware of the practical steps we can take to steward the responsible use of our research data. Leslie et al. (2024) identify three principles of responsible data stewardship – Data Protection and Privacy, Data Integrity and Data Quality – all of which should be considered when deciding how and when to use AI tools in our research.
Which AI?
Artificial intelligence is a broad field encompassing a wide range of methodologies, and as researchers we can engage with AI in different ways. For example, when we use generative AI (GenAI) as a productivity tool – e.g. to perform tasks like writing code, summarising text, transcribing interviews or extracting themes from participant responses – we are users or consumers of AI. We can also be practitioners, e.g. applying pre-trained classical AI models to a research problem, or developers, e.g. training a new AI model.
This article focuses on the use of GenAI as a productivity tool. When we engage with AI in the role of practitioner or developer, additional data governance concerns may apply.
Ethico-Legal Landscape
When deciding when and how to use GenAI as a productivity tool in research, the following laws, ethics procedures and policies (among others) should be considered:
- UK GDPR and the Data Protection Act 2018 – personal data relating to living, identifiable people
- Common law duty of confidentiality – information shared in confidence, including about people who have died
- Copyright, Designs and Patents Act 1988 (CDPA) – original creative works, such as text, images, software, datasets and databases
- Export controls – military, dual-use and other sensitive technology
- Contracts (data sharing agreements, NDAs) – data received or shared under an agreement
- The terms of your ethical approval – research participant data
- Institution, funder and journal policy – any research data produced in your project or supporting your publications
The applicability of these laws, rules and policies will be project dependent. However, some broad principles are:
- Don’t share personal data with AI tools unless you are sure it is safe and legal to do so. Minimise the amount of data shared and remove identifiers and sensitive content. Seek advice from your institutional Information Security and IT teams where needed.
- Don't share data with AI where export controls, data sharing agreements or non-disclosure agreements prohibit this.
- Only upload third-party materials if you own the copyright or are appropriately licensed to do so.
- Don't use AI if this was not part of the original ethical approval (consent).
Data Protection and Privacy
Use Institutionally Approved Tools
You should always use tools approved by your institution where possible. Personal accounts on commercial AI services may offer terms and conditions that are unsuitable for research, e.g. they may use your inputs and prompts to train their models by default, process your data outside the UK or claim rights over your inputs. This can conflict with data protection law and the need to safeguard your intellectual property.
Uploading data into an AI tool is a form of processing, and when the tool runs on external servers, it is also a form of sharing. Uploading certain categories of data, e.g. personal data, can breach data protection law, participant confidentiality or ethical approval.
If an approved tool does not meet your needs, you should ask your institution to assess and approve one. When selecting and evaluating tools to put forward for approval, you should consider where data is sent, how long it is kept for and who owns the rights to inputs and outputs. See the University of Birmingham’s guide Responsible AI tool selection for researchers for further information.
Institutionally approved tools are generally safer, with contractual arrangements that ensure providers do not train on user data. However, they may not be suitable for use with specific categories of data such as personal data, research subject to export controls or data obtained under a data sharing agreement.
Deciding What Data You Can Enter Into an AI Tool
Before you use an AI tool, you must understand what category of data you are working with and whether your institutionally approved tool is suitable for use with that category of data. The following categories of research data may be prohibited, or may require a data protection impact assessment and additional oversight or approval before being uploaded, even to institutionally provided AI tools:
- Personal data without special category data
- Personal data with special category data
- Data subject to export controls
- Data subject to a non-disclosure agreement (NDA)
It is important to remember that pseudonymised data, where identifiers have been replaced with codes, are still personal data and should be handled accordingly. In contrast, truly anonymous and non-personal data generally carry fewer restrictions, but you should still check who owns the data and whether there are any licence terms that prohibit the use of AI.
If your project uses secondary data held under a licence or data sharing agreement, it is important to check the terms before you upload.
Technical Safeguards and Security
When seeking to share any human subject data with an institutional AI tool (with the appropriate approvals), you should minimise the risk of re-identification and harm to participants through technical safeguards:
- Investigate whether you can achieve your objectives by sharing only metadata, such as data dictionaries.
- Consider using mock data or synthetic data instead. While mock data is purely fabricated, synthetic data retains statistical properties of the original data and may not automatically be safe – careful assessment of disclosure risk is needed. Tools such as synthpop may be helpful here.
- Share the minimum amount of data necessary and strip any identifying details such as names, places and locations; tools such as Presidio by Data Privacy Stack may be helpful here.
You should also remain informed and vigilant about security risks associated with the use of AI tools, such as slopsquatting and prompt injection, which can result in data loss, damage or exfiltration.
Informed Consent
If you plan to analyse qualitative data such as free-text responses, interview studies and focus groups using AI tools, you should seek informed consent. For secondary analysis of such data, the original consent may not cover AI use – so check before you upload.
Data Integrity and Data Quality
Data integrity and data quality are additional dimensions we should also consider when working with AI tools. Never edit master data directly with AI tools. Instead, use AI tools to write code that prepares analysis-ready data in a reproducible and auditable manner.
If you use AI tools to derive new variables, e.g. by classifying free text, remember that an AI model’s results may not agree with human annotators, and many models suffer from bias. The only way to assess the quality of this new data is to evaluate the performance of the AI tool using a sample of ground-truth data.
Producing AI
If you develop AI models for others to use, additional data governance considerations should be taken into account. For example, you should know the provenance and lineage of your training data and ensure that you have the legal right to use it for training. You should also assess whether the model is disclosive, e.g. whether it leaks its training data or can reveal information about the training data.
Conclusion
Good data stewardship when using AI in research means staying in control of your data and your results. Understand the ethico-legal requirements that apply to your data, choose an institutionally approved tool, and minimise the amount, identifiability and sensitivity of any data you share with AI.
Further Reading
CAQDAS Networking Project, University of Surrey. Computer Assisted Qualitative Data AnalysiS: information, advice, training and support. https://www.surrey.ac.uk/computer-assisted-qualitative-data-analysis
Guetterman, T., James, K., & DeJonckheere, M. (2026). Ethical considerations for the use of AI in qualitative research. Methods in Psychology. doi:10.1016/j.metip.2026.100266
Leslie, D., et al. (2024). Responsible Data Stewardship in Practice. AI Ethics and Governance in Practice Programme, The Alan Turing Institute. https://aiethics.turing.ac.uk/modules/data-stewardship/
Microsoft. Presidio: Data Protection and De-identification SDK. https://microsoft.github.io/presidio/
Nowok, B., Raab, G. M., & Dibben, C. (2016). synthpop: Bespoke creation of synthetic data in R. Journal of Statistical Software. doi:10.18637/jss.v074.i11
OWASP (2025). Top 10 for Large Language Model Applications – LLM01 Prompt Injection; LLM02 Sensitive Information Disclosure. https://owasp.org/www-project-top-10-for-large-language-model-applications/
Spracklen, J., et al. (2025). We have a package for you! A comprehensive analysis of package hallucinations by code generating LLMs. USENIX Security Symposium. arXiv:2406.10279