How to Choose a Computer Science Dissertation Topic in the UK (2026)

Short answer: A computer science dissertation is marked on the report, not the code. The artefact is evidence; the marks are for problem definition, design justification, evaluation method and critical reflection. The two things that sink projects are the compute ceiling, which you can plan around, and dataset licensing, which most students do not know is a problem until it is one. “Publicly accessible” is not “public domain”, and it is not a lawful basis under UK GDPR.

A working system is not a dissertation

The most common failure in an artefact-based project is a system that runs, demonstrated with a screenshot, and no evaluation. No baseline, no metrics, no comparison, no user study. It reads as a coursework submission, and it is marked accordingly.

Decide your evaluation design before you write any code, and write the evaluation section of the report first if you can. Three questions have to have answers: what am I comparing against, what am I measuring, and how do I know the difference is real. If the honest answer to the first is “nothing”, the project needs rescoping, not more features.

Mark schemes increasingly require a reproducibility package alongside the artefact: a versioned repository, pinned dependencies, seed-controlled experiments, documented hardware, and a licence statement covering any dataset used. Assemble it as you go. Reconstructing it in the final week is miserable and usually incomplete.

The licensing problem, which is the serious one

Kaggle and Hugging Face are indispensable and also full of traps. Many hosted datasets were scraped without a lawful basis, are licensed non-commercially, or contain personal data. If you train on scraped social media content, you are the data controller for a UK GDPR processing operation, and the ICO’s position on web scraping for generative AI training is that legitimate interests can in principle be a lawful basis but the developer must demonstrate necessity and pass the balancing test. A student project cannot meaningfully evidence that.

The practical rule: use an already-published, properly licensed research dataset, or scrape only non-personal data having checked the site’s terms and robots.txt, and document both in your ethics form. Check the licence on every dataset you use and state it in the report. A dataset with no licence at all is not permissive by default; it is unlicensed, which is worse.

Note also that a purely artefact-based project on public non-personal data usually gets a light-touch or exempt ethics review, but the moment human participants evaluate your artefact in a usability study or survey, full committee review applies. Build that lead time into the plan.

The compute ceiling, and designing under it

The familiar arc: propose to fine-tune a large model, discover the department has no GPU allocation, fall back to a free Colab tier that disconnects, then run a smaller model against an evaluation designed for the larger one. The fallback is fine. The mismatched evaluation is not.

Find out what compute you actually have in week one, then design for it. Several genuinely good dissertation shapes need no cluster at all: evaluation and benchmarking work, efficiency and energy measurement, accessibility auditing, reproducibility studies, and anything using an existing pretrained model rather than training one.

Data and tooling worth knowing about

Live territory for 2026-27

• AI evaluation. The AI Security Institute, renamed from the AI Safety Institute in February 2025, publishes research and released its inaugural Frontier AI Trends Report in 2026. Its open-source Inspect framework makes real evaluation work feasible on a student budget, which is rare in this area.
• UK AI regulation. The UK remains on a sectoral, regulator-led, principles-based approach rather than a single AI Act. Compliance tooling, auditability and model documentation are live and under-researched. The Law Commission’s new project on public sector automated decision-making, added to its 14th Programme in September 2025, is a strong law-and-computing crossover.
• Accessibility. Public sector website accessibility regulations plus WCAG conformance auditing. An automated auditing artefact has real UK public sector sites to test against and a clear evaluation design.
• Model energy use. Measurable on a laptop with instrumentation such as CodeCarbon. A defensible small-scale empirical dissertation that needs no GPU at all.

Sources

• ICO, lawful basis for web scraping to train generative AI models: https://ico.org.uk/about-the-ico/what-we-do/our-work-on-artificial-intelligence/response-to-the-consultation-series-on-generative-ai/the-lawful-basis-for-web-scraping-to-train-generative-ai-models/
• ICO, Data (Use and Access) Act 2025 guidance: https://ico.org.uk/about-the-ico/what-we-do/legislation-we-cover/data-use-and-access-act-2025/
• AI Security Institute research: https://www.aisi.gov.uk/research
• Inspect Evals: https://ukgovernmentbeis.github.io/inspect_evals/
• BCS Code of Conduct: https://www.bcs.org/membership-and-registrations/become-a-member/bcs-code-of-conduct/
• UCI Machine Learning Repository: https://archive.ics.uci.edu
• Hugging Face datasets: https://huggingface.co/datasets
• Find Case Law API: https://nationalarchives.github.io/ds-find-caselaw-docs/public
• data.gov.uk: https://www.data.gov.uk