Scientific exploration, collaboration and labor division in the large language model era

Since the widespread release of tools like ChatGPT, scientists are changing how they work. They are exploring more diverse research fields, collaborating with a wider variety of experts, and dividing tasks within teams in more specialized, modular ways.

While many discussions around Large Language Models (LLMs) focus on their ability to write better prose, the actual impact on the structural organization of science remains poorly understood. Traditionally, scientific progress relied on deep specialization—the "exploitation" of a narrow niche—and highly integrated research teams where members shared many overlapping responsibilities. Before the LLM era, moving into a new field was a high-cost endeavor. It required a massive investment in learning new vocabularies and navigating unfamiliar disciplinary norms.

A new study from the University of Wisconsin–Madison suggests that LLMs may be lowering these cognitive and social barriers. The researchers report that the period following the mass diffusion of LLMs coincides with a fundamental reorganization of how scientists choose their topics, pick their partners, and distribute their labor.

Beyond the niche: expanding research portfolios

Historically, scientists faced a sharp tradeoff between exploitation (refining known lines of work) and exploration (opening new problem spaces). Moving into an unfamiliar area carries a "pivot penalty." This is the cost of acquiring new knowledge, which can stall productivity. The authors of this study argue that LLMs may mitigate this by helping researchers navigate unfamiliar literature, terminology, and code.

The study finds that this transition is already happening. The authors report that after 2022, scientists increasingly published across a broader range of fields. Specifically, the average number of distinct primary fields represented in scientists' papers rose from 2.3 in 2022 to 2.8 in 2025 [Figure 1a]. This means researchers are spreading their work across more specialized domains. The researchers also measured the Rao-Stirling index—a metric that combines field shares with the "intellectual distance" (the cognitive gap) between them. This index rose from 0.176 to 0.199 in the same period [Figure 1c]. This indicates that research is becoming more interdisciplinary.

This expansion isn't uniform. The authors report that the most pronounced increases in "pivot size"—the degree to which a researcher moves away from their established expertise—occurred among established, advanced scientists [Figure 1g]. Interestingly, the study also notes that non-English-speaking authors from low- and middle-income countries showed the clearest increases in exploration [Figure 1i]. This suggests that AI tools may be acting as a linguistic and technical bridge to global scientific discourse.

Linking AI signals to interdisciplinary movement

To determine if these shifts are actually tied to AI usage, the authors developed a way to estimate an "AI-writing fraction" for millions of papers. Rather than using unreliable third-party detectors, they used a population-level framework based on word frequency. They modeled each paper as a mixture of human-written and AI-generated sentence distributions. They used maximum likelihood estimation (MLE)—a statistical method to find the best parameters for a model—to calculate the proportion of text resembling LLM-generated patterns.

The researchers then compared "high-AI-writing" authors (those with an average fraction above 0.15) against "low-AI-writing" authors (those below 0.05). They used Coarsened Exact Matching (CEM) to ensure the groups were comparable in terms of career stage, field, and productivity. This helped isolate the association between AI usage and research strategy.

The results reveal a strong correlation between AI-assisted writing and research breadth. The authors report that high-AI-writing authors are significantly more interdisciplinary and exploratory .

However, they offer a crucial nuance. High-AI-writing authors were already more exploratory before the widespread adoption of LLMs [Figure 2g]. The paper suggests a "selection-plus-reinforcement" pattern. People who are naturally inclined to take risks and explore new fields are also the ones most likely to adopt LLMs. The tools then reinforce those existing behaviors.

Changing the social fabric of collaboration

If scientists are studying more diverse topics, they must also change whom they work with. The study finds that collaboration networks became more interdisciplinary after 2022. The average number of distinct fields represented among an author's collaborators rose to 5.2 in 2025 [Figure 3a]. This shows that research teams are pulling in more varied expertise.

However, a surprising mismatch emerges regarding how AI-using scientists interact with their peers. While the overall trend shows more diverse collaborations, the authors find a decoupling in the high-AI group. Among high-AI-writing scientists, research interdisciplinarity is less closely tied to the diversity of their collaborators [Figure 3h].

This suggests a shift in how knowledge is integrated. In the traditional model, interdisciplinary research required collaborating with an expert from another field to bridge the gap. The authors propose that high-AI-writing scientists may increasingly integrate knowledge across fields using AI tools as a mediator. This could reduce their reliance on human cross-field collaborators to perform that translation.

Modularizing the research team

The final dimension of this reorganization is the division of labor within teams. Using the CRediT (Contributor Roles Taxonomy) system—a standardized way of reporting who did what in a paper—the authors analyzed how roles are distributed among coauthors.

They report that after 2022, the division of labor became more "modular" and differentiated. Individual contributors reported narrower sets of roles on average [Figure 4a]. Coauthors also shared fewer roles in common [Figure 4f]. Instead of everyone on a team doing a bit of everything, members are taking on more distinct, specialized responsibilities.

The study highlights a specific shift in the types of tasks being reported. The share of papers reporting "software" and "validation" roles increased. Meanwhile, "conceptualization" and "management" roles saw a decline [Figure 4c]. The authors suggest this may reflect a reality where LLMs automate or accelerate certain technical and drafting tasks. This allows human researchers to focus more on high-level supervision and verifying outputs.

Assessing the reorganization

The authors are careful to note that this study is descriptive. They do not attempt to prove that LLMs cause these changes. They also do not track private tool usage. Furthermore, the data is heavily weighted toward biomedical research due to the use of PubMed Central. The "AI-writing fraction" is also a textual proxy. It does not specify exactly which tasks the AI performed.

Is this reorganization a net positive? The evidence suggests a more flexible and expansive scientific landscape. However, the authors warn of a dual-edged sword. While AI can expand a scientist's reach, it could also inadvertently narrow the collective focus of science. This might happen if researchers gravitate toward "data-rich" problems that are easier for models to process.

For the technical community, the verdict is clear. The integration of LLMs into science is not just about better writing. It is actively reshaping the social and structural architecture of how human knowledge is produced.

Figures from the paper