Rethinking Research: Why Humans Still Matter in the Age of LLMs

Large language models (LLMs) should not replace human researchers; they should amplify what we do best: critical thinking, domain expertise, and imagination. Whether that happens depends on whether we choose to treat LLMs as partners rather than shortcuts or competitors.

From supervisor to co‑creator

Over more than three decades as an academic, I have supervised and mentored nearly 100 PhD students and postdoctoral researchers. Those relationships – thinking together, debating, arguing over drafts, planning experiments and learning – have been one of the great joys of my career and the engine of my own research.

That dynamic is changing. LLMs have entered the supervision process almost silently: students use them to find research problems, summarise papers, write code, and draft sections of their theses. When I first experimented with these tools in my own research, I watched them lay out plausible ‘open problems’, suggest research questions, and list highly cited papers in seconds.

My first reaction was a big surprise. If a model could do in minutes what I had spent years doing with my students, what exactly was my role as a supervisor?

Why humans still matter

Looking more closely at those model outputs reminded me why human expertise remains central. The ‘most influential’ papers suggested by the model were real, but other experts would reasonably choose differently. The research questions sounded original, but many were repackaged versions of existing work rather than original contributions. Some of the proposed problems sounded plausible but were wrong, and some were technically correct but irrelevant.

Without domain expertise, it would be very hard to see that. With it, I could ask basic but crucial questions: Is this really an open problem? Has this already been done in another form? Is this direction likely to lead to new knowledge or just another incremental paper? The model’s fluency did not remove the need for judgement, it simply changed where that judgement had to be applied.

This is where critical thinking comes in. As engineers and scientists, we are trained to solve problems, but not always to deeply interrogate the reasoning behind a solution. Working with LLMs forces us to question why a model proposes one hypothesis rather than another, which assumptions are built into its suggestions, and where its ‘reasoning’ stops being evidence and becomes speculation.

A three‑way partnership

I now think of PhD supervision as a three‑way partnership between student, supervisor, and model. In this partnership, the roles need to be clear.

• LLMs act as co‑creators, not co‑supervisors or co‑authors (at least not just yet!). They can summarise papers, propose alternatives, draft text, and challenge our arguments, but they are not responsible for the integrity or originality of the work.
• Supervisors focus on what cannot be automated. We set standards for rigour, help students formulate problems, and decide when a research direction is worth pursuing, grounded in good human values such as collaboration, empathy and integrity. Supervisors don’t only provide discipline expertise, but also provide very important non-academic support for students through the often challenging PhD journey for successful completion.
• Students learn to work with LLMs critically. When I supervise a new PhD student, I now suggest a simple task: read three key papers and write a summary without an LLM, then bring both your own summary and your prompts to our meeting. Only afterwards do we compare your summary to what the model produces and discuss where it adds value and where it misleads. To embed this three-way partnership as practice, I do the same task asked of the student and we compare the findings, including the prompts used to engage with the LLMs. The difference we see in prompts used may reflect our lived experiences.

This practice builds trust on ground we can verify. Over time, as students and researchers use models for problems where no one knows the answer in advance, that early training in ‘peer‑reviewing’ the LLM becomes essential.

Hybrid intelligence, not full automation

There is growing excitement about automating scientific discovery.

In my view, full automation may be technically possible, but it is neither realistic nor desirable as the dominant model of science. Training and running huge models carries significant environmental costs, undermining the very societal benefits we hope research will deliver.

The alternative is what I call hybrid intelligence. Human ‘biological’ intelligence – our intuition, values, and ability to reframe problems – narrows the search space and decides which questions matter. Artificial intelligence then explores that space at a speed and scale we cannot match. Used in this way, LLMs become force multipliers rather than shortcuts: they help us tackle problems that would otherwise be out of reach, but they do not choose the problems for us.

What we do with the time we save

LLMs have shortened parts of the research journey. Tasks such as literature reviews, research ideas, first drafts and coding can be accelerated or partially automated. The important question is not just ‘what will be automated?’ but ‘what do we do with the time that is freed?’.

If our answer is simply ‘produce more papers’, we will have missed the point. I believe we should use that time discussing and debating ideas. We should think more seriously about impact – social, cultural, environmental – rather than treating publication counts as the main measure of success. We should build research teams that use LLMs to surface diverse prompts and perspectives instead of converging too quickly on the most plausible‑sounding answer.

LLMs now influence how research problems are framed, how quickly we move from idea to draft, and how students experience supervision. I do not want AI to become a shortcut for research; I want it to become a partner. I do not want it to limit the intellectual development of my students or my own; I want it to accelerate what we can do with an intelligent machine.

The progress of LLMs has been extraordinary and will continue. One outcome I hope we can avoid is that researchers fail to adapt to this changing journey of discovery, mentoring and supervision, and humans drift from being the centre of advances in knowledge creation to mere peripheral actors.

This reflection only sketches a few of the questions I raised in my recent talks at King’s College London (Tuesday 9 December 2025) and Newcastle University (Thursday 11 December 2025) on how researchers should work with intelligent machines. For those who want to go deeper, the full talk at King's Rethinking Research: The Role of Humans in Scientific Discovery in the Age of LLMs is available on YouTube and the PowerPoint presentation is available here in PDF format.

I have been continuing this conversation at universities across the UK, including:

• Imperial College London, Wednesday 4 February 2026
• University of Bristol, Monday 9 February 2026
• Cranfield University, Tuesday 10 March 2026
• University of Southampton, Wednesday 22 April 2026
• University of Exeter, Wednesday 13 May 2026
• Queen Mary University of London, Tuesday 26 May 2026
• Loughborough University, Thursday 28 May 2026
• The London School of Economics and Political Science, Monday 1 June 2026
• Northeastern University London, Friday 12 June
• University of Liverpool, Tuesday 7 July 2026

I will also be delivering talks on the topic at Hong Kong Polytechnic University on Tuesday 5 May, Southern University of Science and Technology (SUSTECH) on Thursday 7 May, and online for The World Academy of Sciences' Distinguished Speaker series on Monday 22 June.