Translating AI into scientific impact: Field context, career position, and institutional capability in AI-enabled research

Unequal Returns to AI: How Field, Career, and Institution Shape Scientific Impact

Using massive amounts of research data, scientists found that using AI in research does not benefit everyone equally. Senior researchers see more "prestige" from just mentioning AI. Younger researchers must use it deeply to succeed. Furthermore, universities with mid-level AI skills actually see the largest proportional gains in citations by translating AI for other fields.

Artificial intelligence is rapidly becoming a foundational tool in scientific research. It acts as a general-purpose input (a tool applicable to many different tasks). This can be applied to everything from predicting protein structures to analyzing text. Historically, the scientific community has assumed that increased access to powerful new technologies leads to a uniform rise in impact. However, the mere availability of a tool does not guarantee that every researcher or institution can extract the same value from it.

A new study from Peking University addresses a critical gap in our understanding of this technological transition. While previous research has focused on how much AI is being adopted, this paper asks who actually benefits from that adoption. The researchers investigate how the translation of AI knowledge into measurable scientific impact is associated with three distinct layers: the scientific field, the researcher's career stage, and the institution's existing technical capability.

The limits of simple adoption

Current discourse around AI in science often treats the technology as a monolith. This assumption overlooks the complex social and evaluative structures that govern how scientific work is recognized. Scientific impact is not just a function of technical prowess. It is a product of legitimacy and audience reception. A technique that is considered cutting-edge in computer science might be viewed as peripheral in a traditional field like medicine.

The authors argue that the "return on investment" for AI knowledge is not a constant. As shown in, while AI integration is visible across many disciplines, its prevalence and intensity vary wildly.

Mathematics shows high levels of AI integration. Conversely, fields like Medicine and Chemistry show much lower levels of AI-referencing papers. Simply adopting AI does not solve the problem of how that knowledge is integrated into existing disciplinary frameworks. This creates a risk where the benefits of AI are concentrated in specific pockets of the scientific landscape.

A multilevel framework for translation

To move beyond simple adoption metrics, the authors develop a multilevel framework to track how AI knowledge is linked to citation impact. They categorize AI integration along two margins: the "extensive margin," which tracks whether a paper cites any AI-related work, and the "intensive margin," which measures the actual share of AI-related references in a bibliography.

The mechanism of impact operates through three primary channels:

1. Field Legitimacy: Disciplinary filters determine if AI-based methods are seen as credible. The authors find that the returns to AI vary by field. For example, while AI is positively associated with impact in Sociology, it can be negatively associated with impact in Engineering and Economics .

2. Career Signaling: For senior scholars, the mere presence of AI references acts as a signal of intellectual renewal. For junior scholars, however, the "signal" of just mentioning AI is insufficient. They require "intensive" engagement—using AI deeply and citing high-quality, recent AI papers—to build technical credibility .

3. Institutional Translation: This is a nuanced layer. The authors suggest that impact is associated with "translational capacity" (the ability to make AI knowledge meaningful to non-AI audiences).

Instead of a simple linear progression where more AI skill equals more impact, the authors propose a non-monotonic model. They hypothesize that institutions with intermediate AI capability hold a unique advantage. These institutions may bridge the gap between specialized AI technicality and broader scientific utility.

Evidence of the translation advantage

The study utilizes a massive dataset from SciSciNet V2. It encompasses approximately 12.8 million focal papers from 360 institutions. Using Poisson pseudo-maximum likelihood (PPML) models (a statistical method for handling skewed data), the authors report that, on average, integrating AI knowledge is associated with higher five-year citation impact. Specifically, the paper finds that a higher share of AI-related references (the AI score) correlates with more citations. This relationship faces diminishing marginal returns [Table 1].

The institutional analysis reveals a non-linear pattern regarding citation returns. The authors report that the largest proportional citation returns are observed among institutions with intermediate AI capability (the Q2 and Q3 groups). This differs from the absolute leaders in AI research (the Q4 group) [Table 3].

The researchers suggest this pattern may be driven by two distinct pathways: * Epistemic Positioning: Top-tier AI institutions (Q4) tend to produce work that is deeply embedded in AI-centered "neighborhoods." Their papers are highly technical and frequently cited by other AI researchers .

• Audience Diversity: Intermediate institutions (Q2/Q3) exhibit higher "audience entropy" (a measure of how diverse a group of followers is). Their AI-enabled research attracts citations from a much wider variety of scientific subfields. They may act as brokers, translating complex computational methods into language that the broader scientific community can use.

Identifying the blind spots

While the study provides a robust macro-level view of AI diffusion, it has notable limitations. First, the authors rely on "observable traces"—specifically, citations in a reference list—to measure AI integration. This is a proxy. It cannot capture the actual use of AI tools, proprietary code, or the tacit expertise a researcher might employ. A researcher might use a sophisticated transformer model without ever citing the seminal paper that describes it.

Second, the study is observational. While the authors use various controls and fixed effects, they cannot definitively claim that increasing AI capability will cause a specific increase in citation impact. There may be unobserved factors, such as funding levels, that drive both AI capability and citation success.

Finally, the institutional metric is based on CSRankings. This primarily measures academic strength in computer science. This may not fully capture "applied" AI capability. This includes expertise found in industry-linked research labs or institutions focusing on applying AI to biology.

The role of translational capacity

The findings suggest that the value of AI in science is not determined solely by technical depth. Instead, the impact depends on how well that knowledge is translated across different communities.

If the objective is to push the mathematical and algorithmic frontier, top-tier AI institutions remain the primary engines of progress. However, for maximizing broad scientific influence, intermediate institutions appear to play a vital role. These institutions seem to excel at bridging the gap between specialized AI research and diverse scientific applications.

For policymakers and university administrators, the takeaway involves more than just providing technical resources. To truly harness the AI revolution, institutions may need to support the "translational capacity" required to make machine learning meaningful to chemists, geologists, and sociologists alike.

Figures from the paper