Machines Reading the Team: AI Reveals How Scientists Slowly Learn to Speak the Same Language

When dozens of researchers from wildly different fields are thrown together to tackle a problem as sprawling as the roots of unhealthy urban development, one of the hardest questions is deceptively simple: are they actually understanding each other? A team at the University of Bristol has now tested whether artificial intelligence can answer that question, by pointing natural language processing (NLP) tools at six years of internal interviews from a large UK research consortium. Their findings, published in PLOS Sustainability and Transformation, offer both an encouraging proof of concept and a sobering lesson about how long it really takes for a multidisciplinary team to build a shared language.

The consortium in question was assembled to investigate the upstream determinants of non-communicable disease in urban settings, drawing on 57 academic staff across six universities and spanning thirteen disciplinary areas grouped under the UK’s three main research councils: economic and social sciences, engineering and physical sciences, and medical research, plus a professional services team. From 2019 to 2025, a dedicated research-on-research sub-team conducted 63 semi-structured interviews with 39 consortium members in three rounds, timed in Years 2, 3 and 5 of the programme. Those transcripts, normally the raw material for slow, expert-led qualitative analysis, became the test bed for a very different kind of scrutiny.

The researchers adopted what they call a dual approach. The first strand used co-word clustering, an NLP technique that groups words appearing together across texts, to identify emergent themes in the full interview transcripts. The second strand was more targeted: the team built a dictionary of 245 tokens deemed to represent transdisciplinary language, extracted from foundational project documents including the peer-reviewed programme protocol, and then tracked how those words were used over time and across disciplinary boundaries. The dictionary was split into tokens describing the upstream problem under investigation, the mid-downstream built environment and health outcomes, and the research problem itself, that is, the challenge of actually running such a complex collaboration.

The co-word clustering produced eleven initial clusters, refined down to nine coherent themes, ranging from new large-scale research approaches for real-world problems to local government case studies, health and decision-making, mission and integration, shared understandings, project structure, and urban development systems. Yet despite the apparent richness of these themes, the technique yielded no clear findings about how transdisciplinary understanding developed. The authors are candid about this: the themes could be described and tentatively interpreted, but whether those interpretations were accurate or merely what they call wishful thinking remained uncertain, particularly given the low volume of data and the wide variation in the number of tokens within each cluster.

The vocabulary analysis told a far more compelling story. In the first interview round, almost half of the 245 transdisciplinary tokens were not used by any of the disciplinary groups, and most of the rest were used by only one, two or three of the four groups rather than all of them. Over the six years, however, a clear convergence emerged. Overall use of the transdisciplinary vocabulary declined, but the words that remained in circulation became more widely shared. By the end, 135 words were used by at least one discipline, 40 by all four, and the trajectory pointed towards a shared lexicon of roughly 15 to 25 percent of the original dictionary. In other words, the team did not adopt more jargon as it matured; it shed most of it, converging on a small set of useful, jargon-free terms.

The disciplinary breakdown added intriguing nuance. Economic and social scientists used the greatest diversity of transdisciplinary vocabulary in the first round, across by far the widest range of interviews, at 45 percent across 15 interviews, falling to 30 percent and then 23 percent in later rounds. Public health researchers used the second most, followed by engineers, with professional services staff using the least, and none at all in the first round. That ordering is striking because the engineers were arguably the group most familiar with the vocabulary from the outset, having led the systems science and research-on-research work. One explanation offered is that the engineers focused on process-oriented systems analysis rather than the wider research themes; another is that, being most fluent, they simply expressed the same ideas in different words.

Perhaps the most consequential finding is about time. It took the consortium essentially the full six years to develop its shared vocabulary, and the convergence accelerated only towards the end of that timeline. The authors note that the team once spent six months of considerable discussion simply agreeing what the word health meant. For funders and research leaders, the implication is uncomfortable but valuable: the development of shared understanding in large, newly formed transdisciplinary teams may not be easily short-cut. It is a matter of exposure and accumulated interaction, and expectations of rapid conceptual alignment should be managed accordingly. The extensive early management activity in this consortium, including co-developing a glossary, adopting mission-orientation, and building shared conceptual mental models, may have helped, but it may also have attempted to fast-track a process that could not be forced.

The study is equally instructive about the limits of NLP itself. The approach was labour intensive, combining automated analysis with manual, expert-led contextual interpretation, and the authors judge that it could not have provided rapid course corrections during the live programme. Interviewer words were included in the data, interview questions varied between rounds, the consortium restructured between phases, and normalisation across unevenly sized disciplinary groups posed non-trivial challenges. Any confident use of such methods, the authors argue, would require a dedicated validation exercise, ideally involving the interview participants themselves, along with minimum thresholds for token numbers and theme coherence.

Still, the authors see genuine promise. If datasets are larger, better formatted, and analysed with more efficient scaling, NLP could surface patterns that traditional qualitative analysis misses, and its semi-automated nature may reduce researcher bias. The clearest lesson, that shared language in transdisciplinary science takes years rather than months to emerge, and that less jargon ultimately proves more useful, is one that any team hoping to solve complex global challenges would do well to hear before the funding clock starts ticking.

Subject of Research: Using natural language processing to assess the development of shared transdisciplinary understanding in a large research consortium

Article Title: Using Natural Language Processing (NLP) to assess changes in transdisciplinary understandings across a large research consortium

Article References: Black, D., Gopsill, J., Rosenberg, G., Silvonen, T., Kukreja, A., & Hicks, B. (2026). Using Natural Language Processing (NLP) to assess changes in transdisciplinary understandings across a large research consortium. PLOS Sustainability and Transformation, 5(10), e0000222. https://doi.org/10.1371/journal.pstr.0000222

Image Credits: AI Generated

DOI: 10.1371/journal.pstr.0000222

Keywords: natural language processing, transdisciplinary research, interdisciplinary collaboration, research-on-research, co-word clustering, shared vocabulary, research consortium, team science, urban health, research evaluation, qualitative interviews, PLOS Sustainability and Transformation