Lyndon Drake, Research Fellow in AI, Faculty of Theology and Religion
I have an academic and professional background in Computer Science and software development, so I am used to writing code and enjoy it. There's no doubt, though, that as it is no longer my focus, writing software has become more time-consuming. Alongside this, there are increasing challenges in reviewing the vast quantities of new publications in many fields, especially as my work involves cross-disciplinary scholarship.
A key example where Gemini helped was in creating a prototype of an LLM-based virtual viva voce examiner, an idea that I worked on with two colleagues and have documented in an arXiv paper. From a research point of view, the mundane work of coding the proof-of-concept web app was not itself significant, as it was obvious that it was possible to write such an app once we had conceived of the idea. Being able to use Gemini to quickly produce it meant I was able to focus on the research task, and Gemini also provided the backend LLM, which was ideal for a proof-of-concept.
A second example has been using Gemini’s Deep Research function for literature surveys, where I have been able to identify key pieces of literature in relation to research questions as they arise. The feature of Deep Research that is especially useful is that it identifies higher-quality results (which still need checking, as responses still include non-existent but plausible works, mischaracterised works i.e. real literature which does not match the LLM summary, or real works but where the link provided is not to the work in question). Despite these flaws, there is no question that Deep Research has typically provided a high-quality survey of a new area for me in a matter of hours and has surfaced important literature that I am reasonably sure I wouldn't have encountered if working in a more traditional approach.
For my specific use cases, I estimate several hours saved in each example. In the case of the prototype LLM viva tool, while not a complex piece of software development, I have not developed a web app for some years and it would certainly have taken me at least a day to complete by myself, instead of the minutes it took using Gemini. In the case of literature surveys, taking into account the time needed to check results (often far longer than the time Gemini took to provide a response), I would say that it turned a week's work into about a day's work, but it's important to note that as well as saving time it produced better results than I think I could have achieved by myself.
In my view, the key to using GenAI tools has been to identify those tasks which (a) take time as a human and (b) are not what I consider to be central to the research work I am undertaking.