Artificial intelligence / UK careers / STEM
How AI Is Changing STEM Jobs in the UK
UK evidence on AI and STEM careers: which technical tasks change, what it means for graduates, and how employers should test a workflow.
AI is changing STEM work in the UK task by task. It is not yet replacing STEM jobs wholesale.
UK research published in January 2026 describes improving AI capabilities in coding, cybersecurity and research, alongside some weaker hiring in occupations with high AI exposure. It does not establish that AI caused those hiring patterns. Employers still report difficulty filling AI roles.
At a glance:
- Tasks change first. Code drafting, log summaries and literature searches are already assisted work.
- Hiring signals are mixed. High exposure is not proof of adoption or redundancy.
- Verification is the skill. The professional decides whether an output is fit to use.
This article separates what the tools can do, what employers use them for, and what the hiring evidence shows. The workflow examples and preparation suggestions are editorial analysis, not measured employment outcomes or forecasts.
UK evidence shows task change, not proven job losses
Improving capability differs from everyday work
The Department for Science, Innovation and Technology's January 2026 assessment of AI capabilities and the UK labour market describes improving capabilities in coding, cybersecurity and research.
It also stresses a gap between controlled evaluations and everyday work. In real projects, requirements are ambiguous and tasks need iteration.
Hiring is weaker in exposed roles, but the cause is unproven
The assessment reports signs of weaker hiring in some occupations with greater AI exposure. It says these patterns do not establish that AI caused the decline. Economic conditions and differences between sectors also matter.
Exposure measures potential overlap between AI capabilities and work tasks. It does not prove adoption or redundancy.
Skills shortages sit alongside hiring uncertainty
The AI Labour Market Survey 2025 executive summary, published on 28 January 2026, reports that 97% of respondents identified at least one AI labour-market skills gap. It also reports that 35% of organisations struggled to fill AI roles.
These are findings from the organisations surveyed, not percentages of every UK employer or every STEM occupation. Recruitment can become harder for some applicants while employers still struggle to find particular kinds of expertise. A skills shortage is not a promise that every graduate will find a job easily.
AI assists five STEM task types; verification stays human
STEM is a broad category. A laboratory technician, a statistician and a structural engineer face different constraints. The table shows a possible use in each area and the check that remains with the professional. It does not claim that every UK workplace has adopted these tools.
| Area | Example of AI-assisted work | What a professional still needs to establish |
|---|---|---|
| Software development | Draft a function, explain unfamiliar code, suggest tests | Does it meet the requirement, handle failures and fit the system? |
| Data science and mathematics | Suggest analysis code or explore modelling approaches | Are the assumptions valid, and does the evaluation avoid leakage? |
| Engineering | Explore design alternatives or draft technical documentation | Are the constraints, units and physical behaviour correct? |
| Scientific research | Help organise literature or draft an analysis workflow | Do the cited papers exist, and can the result be reproduced? |
| Cybersecurity | Summarise logs or suggest investigation steps | Does the evidence support the conclusion, and is the action authorised? |
The common issue is verification. Generating an answer cheaply moves effort towards deciding whether that answer deserves to be used.
Software: producing code and owning a change differ
Imagine a developer using an assistant to draft an image-upload handler. The first version may look convincing. The professional task also includes checking file limits, malformed input, permissions, error handling and how the change affects existing users.
If drafting becomes faster, a team might deliver more work with the same staff, change the mix of roles, or reduce some hiring. Which outcome occurs depends on demand, budgets and management decisions. A coding demonstration alone cannot tell us the employment result.
For a candidate, a useful project therefore includes a short explanation of the decisions, tests and failure cases. Explaining why a generated implementation was rejected can reveal more understanding than a large quantity of unexplained code.
Science and engineering: keep the link to physical evidence
A generated summary can help organise a problem, but it is not an experiment. A suggested design is not a verified physical system. Measurements, instruments, operating conditions and domain assumptions remain essential to assessing the output.
An illustrative materials project might use AI to suggest candidate formulations. The team would still need a credible selection method and experimental evidence. Likewise, a design assistant's suggestion needs checking against the actual loads and constraints before anyone relies on it.
Subject expertise plus the ability to evaluate computational assistance is the useful pairing. Learning the tool without learning the underlying discipline leaves the user with fewer ways to recognise a plausible mistake.
Entry-level roles need a new learning ladder
One concern is how juniors build judgement. Early tasks often mean writing simple functions, cleaning data, preparing reports and checking calculations. If teams automate some of that work, they need another way for new colleagues to practise and receive feedback.
This is a risk to manage, not evidence that all entry-level STEM roles will disappear. Make learning visible: ask a junior colleague to explain a model's assumptions, reproduce a result independently, or compare an assisted solution with a simple baseline.
The AI Skills for Life and Work summary report, published in January 2026 from research conducted between November 2023 and March 2025, distinguishes AI experts, specialists, implementers and general users. It treats judging output accuracy and reliability as a foundation across all four groups. That is a more useful starting point than assuming every STEM graduate must become a machine-learning researcher.
Prepare with five checks
These suggestions follow from the task-level analysis above. They are not a ranking of qualifications or a guarantee of employment.
- Keep the fundamentals strong. Reason about the code, mathematics, measurements or engineering principles behind an answer.
- Pick one relevant workflow. Try an assistant on a bounded task in your discipline, using public or synthetic material suitable for the tool.
- Build an independent check. Use tests, a baseline calculation, held-out data or a reproducible experiment. Do not let the same system be the only judge of its own answer.
- Record failures as well as successes. Explain what went wrong, how you noticed, and whether correction removed any apparent time saving.
- Show ownership. In a portfolio or interview, separate what the tool produced from the decisions you made and the evidence you checked.
Employers: pilot the workflow, not the job
A small pilot answers a better question than "Can AI do this job?" Compare a defined workflow with and without assistance. Count review and correction time, inspect quality, and decide who owns the final result. If junior work changes, build supervised learning into the redesign.
AI is not yet replacing UK STEM jobs wholesale
There is no single defensible answer for every STEM occupation. Some tasks are candidates for automation. Others become easier to perform with assistance. Neither observation specifies how many people an employer will hire.
The January 2026 UK evidence supports taking the transition seriously while staying careful about causation claims and predictions. For an individual, the practical aim is to get better at doing and checking meaningful work in their field. For an organisation, it is to measure the full workflow and keep the expertise needed to take responsibility for its outputs.
Sources and scope
- DSIT and the AI Security Institute: Assessment of AI capabilities and the impact on the UK labour market, published 28 January 2026.
- DSIT: AI Labour Market Survey 2025 report, executive summary, published 28 January 2026. The report title refers to the survey year.
- AI Skills for Life and Work: summary report, published 28 January 2026. This commissioned research reflects its authors' methods and findings, rather than government policy.
Sources checked on 26 October 2026. This is an explanatory article, not a live vacancy tracker. The cited studies describe their own research periods and populations.