Research AI Engineering

LERU Digital Education policy group programme meeting

James Hetherington

2026-11-12

European AI-assisted discovery: the role of research engineers

Thank Kathryn Woods. She has talked about AI in education; I’ll follow up on research, and in particular on the people and institutions that make computational research trustworthy, and what has to change for AI.

Twenty minutes, then questions.

The craftsperson and the scholar

Image credit: Beinecke Library, used by permission, composed by the Software Sustainability Institute.

UCL ARC

• Research Software Engineers
• Data Scientists
• Data Stewards
• Research Infrastructure Engineers
• Research Managers

Over the last decade and a half, research software engineering went from a name we made up to a profession: thousands of RSEs worldwide, national societies across Europe, career paths in our universities.

At UCL, ARC and its predecessor started with three people. ARC is now about 120 people, about 60 of them research software engineers.

The point: universities have absorbed a digital transformation of research before, and digital research technology platforms like ARC are how we did it.

What digital research technology platforms do

Platforms, engineering, research into methods, and education, each feeding the others. Professional staff with an academic mission.

This is the machinery a university uses to put new digital methods into the hands of researchers. AI is the next, and biggest, such method.

Europe’s dilemma

Image credit: Generated with ChatGPT image generation, OpenAI (November 2026).

We can’t be left behind. AI capability is becoming a determinant of research excellence and competitiveness. And our researchers aren’t waiting for us: in Wiley’s survey of about 2,400 researchers last year, 84% already used AI, up from 57% the year before, but only 41% felt their institution supported them adequately. The train is leaving with or without the institution.

But many in our communities are sceptical, and not without reason: copyright and provenance, energy and environment, labour, and whose values are built into these systems.

In ARC we wrote a policy for AI-assisted software development that starts from both facts: AI may help, and it is here for years to come; used unwisely, it risks the epistemic grounding of research; and colleagues with reservations must be respected alongside enthusiastic adopters. We tried to disagree well.

That tension decides how fast we can move.

Values alignment is speed

So why do values matter for a leader who just wants their institution not to be left behind?

Because the weight of opposition in our own communities will slow adaptation. If the AI we offer is opaque and unaccountable, the sceptics are right to resist, and they will.

Digital research technology platforms already have the answer in our own values: FAIR, findable, accessible, interoperable, reusable. Applied to AI, FAIR points to open-source and open-weights models, run where we can see them, with AI use attributable and auditable. Our ARC policy prefers these where capability allows.

Values-aligned AI is not a brake. It’s what wins the consent that lets us move fast.

The openness spectrum

Closed

Weights, data and training private

Open weights

Run it yourself; training data private

Fully open

Apertus EuroLLM Marin

Weights, data, code and recipes all public

More capable today ⟵ ⟶ more inspectable, more sovereign

Closed frontier models: the most capable, but nothing to inspect.

Open weights: download and run on your own hardware, which matters for sovereignty and data protection, but you can’t see what went in. This is where Mistral, Qwen, DeepSeek and NVIDIA’s Nemotron sit.

Fully open: weights, training data, code and recipes all published. Apertus from the Swiss AI Initiative, ETH, EPFL and CSCS; EuroLLM; and from the US, Ai2’s OLMo and Marin from the Open Athena lab. This is FAIR for AI.

Fully open, but behind

Among the fully open models, Europe holds its own: Apertus has the best multilingual score of any fully open model, ahead of the US’s OLMo and Marin.

But the fully open models as a group trail the open-weights ones by a long way: Qwen 3.5 at 27B is far ahead of Apertus at 70B.

Part of that gap is a deliberate choice: Apertus trains only on data whose owners allow it, and limits memorisation. That’s the price of values alignment, and worth paying.

And the dashed line is the closed frontier: Gemini, Claude and GPT score 90 to 93 on MMLU-Pro. That’s not quite like-for-like, since they’re finished, post-trained products and the bars are base models, but post-training doesn’t close a gap that size. Epoch AI estimate the best open models have trailed the frontier by about four months on average this year.

So the honest position: closed and open-weights models for the hard work now, and fully open, especially European, models as the investment we’re making, for provenance, sovereignty and trust. And Europe needs to invest to close the gap: the EUROPA consortium’s 400-billion-parameter open model, OpenEuroLLM’s first models due next month, and the AI Gigafactories.

And running FAIR models where we can see them means our own compute: universities need to host them locally, not just rent access.

Fake barns

Image credit: Generated with ChatGPT image generation, OpenAI (November 2026).

That’s the institutional choice. Now to what AI does to the knowledge we produce.

Alvin Goldman’s thought experiment: a county full of fake barns, with one real one. A visitor points at the real barn and says “that’s a barn”. It’s true, and justified by what they see, but it isn’t knowledge: in that county they could just as easily have been fooled.

Computational research is very good at building fake barns: results that look and feel genuine, but rest on hidden bugs, wrong assumptions or irreproducible runs. Generative AI makes the facades more convincing: polished, plausible code, at scale.

Running code isn’t knowledge

Shingleton, Hetherington & Basiri, Assertion Errors: Why AI-Generated Research Code Still Needs Epistemic Infrastructure, preprint under review at CACM

With colleagues at Glasgow we argue that research software engineering is part of the epistemic infrastructure of computational science: the practices, norms and social arrangements through which software-mediated claims become warranted.

The usual defence of vibe-coding: we already rely on NumPy without reading it. True, but we trust NumPy because of the community practices behind it: its tests, review history and bug tracker. AI output earns trust the same way, or not at all.

So AI doesn’t make RSE less important. As the supply of plausible code goes up, the work of telling real barns from fake ones becomes more important.

Cognitive debt

The degree to which the community responsible for a technical artefact are unable to understand or explain it, due to accumulated uncomprehended machine construction

After Storey (2026), How Generative and Agentic AI Shift Concern from Technical Debt to Cognitive Debt; Kosmyna et al. (2025), Your Brain on ChatGPT

By analogy with technical debt: technical debt lives in the code; cognitive debt lives in our heads. Margaret-Anne Storey describes it as the team losing its shared theory of the system when AI lets it go fast. The MIT Media Lab’s EEG study found people accumulating cognitive debt when they wrote with an AI assistant. Our ARC policy calls the same thing comprehension debt.

For research, the cost isn’t only that no one can change the code safely: it’s that no one can stand behind the conclusions. So our engineers must be able to make unassisted changes to code they’re responsible for, and code quality, not just behaviour, stays the standard: we oppose “code is throwaway” for research.

And inherited debt: collaborators now arrive with projects designed with an AI. Our staff review them, and can decline work where the debt can’t be controlled.

Two futures

Image credit: Generated with ChatGPT image generation, OpenAI (November 2026).

Two futures for research software engineering.

In one, RSE becomes irrelevant: why hire an engineer when a postdoc can generate the code in an afternoon? And we get a generation of vibe-coded papers, standing on code nobody understands. Bad science, at scale.

In the other, RSE becomes the epistemic infrastructure for AI-assisted science: the people who make AI-written code trustworthy, and who teach researchers to do the same.

Which one we get depends on whether RSE communities adapt now. Our policy expects every engineer to get hands-on with these tools: to keep claiming to be experts, we have to be expert in this too.

Research AI Engineers

So here’s my proposal: a new profession, the Research AI Engineer, alongside research software engineers, research infrastructure engineers, data stewards and data scientists.

It’s the profession that answers both halves of this talk: the people who run open, FAIR models where we can see them, and who tell real barns from fake ones.

Not prompt engineers. People who build the epistemic infrastructure for AI-enabled research: who choose and run the models, frontier and open; who make AI-assisted research reproducible, attributable and auditable; and who teach researchers to tell real barns from fake ones.

RSE became a profession because a community named it, built careers for it and found the money for it. We need to do the same, faster.

Faculty AI Changemaker Teams

What we’re doing at UCL. Credit Tim Scanlon, who leads it.

Research AI Engineers embedded in faculties, close to the research, but coordinated through ARC for platforms, a community of practice and professional development: the model that grew ARC.

They give hands-on help, training, compute and seed projects, and co-supervise students. They’re open-ended ARC posts, not short-term project roles: a career.

Mathematical and Physical Sciences is the pathfinder, recruiting now; Life Sciences next; then further faculties as the evidence comes in.

In UCL’s AI strategy consultation this was the most strongly supported initiative. One colleague called it “a no brainer”.

A call for action

• Re-tool your digital research technology platforms for AI, now
• Use frontier models and invest in open, European ones
• Invest in local compute to host FAIR models
• Recognise and hire Research AI Engineers

Four asks of leaders in this room.

Re-tool your digital research technology platforms now; they’re the machinery for this transformation, and they need to change.

Values-aligned AI is how you keep your community with you: use frontier models, and invest, together, in open European ones.

Invest in local compute to host FAIR models, so that open models run where we can see them, under our own governance.

And recognise the profession: Research AI Engineers, with real career paths. Back to where Kathryn started: they are educators too, teaching our researchers to use AI in ways that make knowledge, not fake barns.

Thank you.

References

• UCL ARC (2026). AI-Assisted Software Development Policy for UCL Advanced Research Computing. doi:10.5522/04/34070631
• Storey (2026). How Generative and Agentic AI Shift Concern from Technical Debt to Cognitive Debt. margaretstorey.com, 9 February 2026
• Kosmyna et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, arXiv:2506.08872
• Shingleton, Hetherington & Basiri (2026). Assertion Errors: Why AI-Generated Research Code Still Needs Epistemic Infrastructure. Preprint, under review at CACM. SSRN 7185258
• Goldman (1976). Discrimination and perceptual knowledge. J. Philosophy 73, 771. doi:10.2307/2025679
• Apertus v1.5 technical report (2026). arXiv:2610.15542
• Wiley (2025). AI Adoption Jumps to 84% Among Researchers as Expectations Undergo Significant ‘Reality Check’ (ExplanAItions study, 2,430 researchers, August 2025). newsroom.wiley.com, 7 October 2025
• Kaggle (2026). MMLU-Pro leaderboard, captured 23 October 2026
• Epoch AI (2026). Open and closed model gap