What a UK AI or data science proposal must show
UK doctoral programmes admit and progress candidates on the strength of a research proposal. Whatever the local template, the proposal has to persuade readers that the project can lead to an original contribution to knowledge, the defining feature of a UK doctorate (QAA, 2020), and that it can be done with the data, skills and time available. In artificial intelligence (AI) and data science, many applicants have strong technical skills but write proposals that describe a system to be built rather than a question to be answered.
A strong proposal connects five things: a specific problem, questions that can be tested, data you can lawfully obtain, an evaluation plan that would convince a sceptic, and an ethical case. This article takes each in turn, using lessons from UK-based research, and ends with a worked example and a final check. Read your programme's own guidance as well; proposal length and headings vary between universities and doctoral training centres.

What UK-based research teaches proposal writers
Several influential critiques of machine learning research come from UK institutions, and reviewers of proposals know them. Each points to a design decision you can make at proposal stage.
| Study and UK setting | What it found | What your proposal should do |
|---|---|---|
| Roberts et al. (2021), University of Cambridge | None of 62 studies using ML on chest images for COVID-19 was judged of potential clinical use; problems included poor documentation, small or biased data and no external validation | Plan external validation and describe data provenance |
| Nagendran et al. (2020), Imperial College London | Most non-randomised studies comparing deep learning with clinicians were at high risk of bias, yet many abstracts claimed comparable performance | Match claims to the strength of the design |
| Pendlebury et al. (2019), King's College London and Royal Holloway | Unrealistic class balance and ignoring time made malware classifiers look far better than they would be in use | Split data by time and use realistic base rates |
| Thomas et al. (2017), Cambridge Cybercrime Centre | Research using leaked or scraped data often lacked a documented ethical justification | Justify the source and use of every dataset |
| Tomašev et al. (2019), DeepMind, London | Continuous prediction of acute kidney injury from electronic health records of US veterans | State the population your data represent and where the model may not transfer |
Taken together, these studies show that the weaknesses examiners find in AI theses are usually decided at the design stage, long before any model is trained (Roberts et al., 2021; Nagendran et al., 2020; Pendlebury et al., 2019; Thomas et al., 2017; Tomašev et al., 2019). Security studies with UK co-authors have generalised these lessons into a list of pitfalls that recur across machine learning research, from sampling bias to evaluation only in the laboratory (Arp et al., 2022).
Frame a problem, not a system
A problem statement that is too broad ("AI for healthcare"), out of date or unimportant will not survive review. Define the problem in the terms of the people who face it, show where current methods fall short and explain why learning from data is needed rather than a simpler rule or model. UK priorities such as health, cybersecurity, sustainability and public services can show relevance, but relevance is not a gap: the gap must come from the literature.
Then turn the problem into objectives and questions that can be tested. "To develop an intelligent system" is an activity; "Does a model trained on hospital A's records keep its discrimination and calibration at hospital B?" is a question. Check that each objective has a method and an evaluation attached to it, and that the literature review argues towards those questions by comparing methods and identifying unresolved disagreements rather than summarising papers one by one.
Data, computing and feasibility
Feasibility is where many AI proposals fail. State which data you will use, who controls them, under what licence or agreement, and how long access will take. Patient data in the NHS normally require Health Research Authority approval (Health Research Authority, 2026); commercial data need a written agreement; scraped data need an ethical and legal justification (Thomas et al., 2017). UK research councils expect data to be managed and, where possible, shared, so include a short data management plan (EPSRC, 2022) and plan to document the dataset with a datasheet (Gebru et al., 2021).
Estimate computing needs (GPU hours, storage) and check what your university provides. Name the main modelling families you expect to compare, the preprocessing and feature engineering steps, and the tools (for example, Python or R libraries), but keep the choice open where the literature does not yet favour one approach. List the main risks, such as delayed data access or too few events, with a fallback for each.
An evaluation plan reviewers will trust
- Primary metric chosen from the decision the model supports, with secondary metrics. Accuracy, F1 and AUC suit classification; RMSE and MAE suit regression. See choosing and reporting evaluation metrics.
- Splits fixed before any preprocessing, by patient, site or time where relevant, to avoid leakage (Kapoor & Narayanan, 2023; Pendlebury et al., 2019).
- Baselines: a simple model and the best published method on the same data.
- External validation on data from another site or period where possible (Roberts et al., 2021).
- Fairness: the subgroups you will report and why.
- Reproducibility: code, versions and seeds to be shared where permitted.
Viewpoints on clinical AI make the same point from the user's side: a model that performs well in development is only the start of the evidence needed before it can be used in practice (van Hartskamp et al., 2019; Nagendran et al., 2020).

Worked example (illustrative)
| Element | Weak version | Stronger version |
|---|---|---|
| Problem | Use deep learning to improve NHS efficiency | Emergency departments lack reliable early warning of patients who will need intensive care within 24 hours |
| Question | Can AI predict deterioration? | Does a model trained at two hospitals keep its discrimination and calibration at a third, and for which patient groups does it fail? |
| Data | Hospital data (to be obtained) | Routinely collected records from three named trusts, access route and approvals identified |
| Evaluation | Accuracy, precision, recall, F1, RMSE and AUC | Primary: discrimination and calibration at the external site; baselines: an existing early warning score and logistic regression |
| Ethics | Will follow ethical guidelines | Approvals named, data minimisation, subgroup analysis by age, sex and ethnicity |
Significance, timeline and final checks
Explain the contribution twice: to knowledge (what will be known that is not known now) and to practice (who could use it, and what further evidence they would need). Avoid promising deployment you cannot deliver. A three-year timeline might allocate the first months to the literature review, data access and approvals; the middle period to data preparation, experiments and validation; and the final year to external validation, papers and the thesis, with milestones for each.
Before submission, check grammar and clarity, technical consistency between sections, the referencing style, your programme's formatting rules, ethical approval needs and a clear statement of originality.
Common mistakes
- An overly broad or generic AI topic.
- Unrealistic system-building goals for the time available.
- Ignoring ethics, privacy or algorithmic bias.
- Weak alignment between objectives, methods and evaluation.
- A descriptive literature review without critique.
- No clear evaluation plan, baseline or validation strategy.
- No statement of originality or contribution.
- Data access assumed rather than planned.
- Poor academic writing and formatting.
For the general structure of a proposal, see how to write a PhD research proposal; for finding the problem, how to identify a PhD research problem in computer science; for a neighbouring field, developing a PhD proposal in UK engineering; and for design pitfalls once the project starts, methodology mistakes in machine learning PhDs.
Key takeaways
- A UK doctorate requires an original contribution, so frame a question, not a system (QAA, 2020).
- UK-led reviews found missing external validation and overstated claims in many ML studies (Roberts et al., 2021; Nagendran et al., 2020).
- Split data by time or site and use realistic base rates to avoid optimistic results (Pendlebury et al., 2019; Kapoor & Narayanan, 2023).
- Plan approvals, data management and documentation before you depend on the data (Health Research Authority, 2026; EPSRC, 2022; Gebru et al., 2021).
- Justify the ethical basis of every dataset (Thomas et al., 2017).
