Does Artificial Intelligence Advance Science?

Artificial intelligence is rapidly diffusing across scientific research. This expansion raises a central tension in the science of science. While AI may increase researcher productivity, there are concerns it might favor incremental work over truly original ideas. The fundamental question is whether AI adoption is associated with advances in scientific creativity.

Current debates often treat AI as a monolithic force. Some view it as a tool that accelerates workflows. Others worry it might reduce intellectual diversity. There has been little empirical clarity on how AI use relates to the nature of scientific breakthroughs. This paper addresses that gap. It investigates the relationship between AI adoption and multiple dimensions of scientific creativity. The study analyzes a massive corpus of over one million publications.

The difficulty of measuring scientific creativity

Measuring creativity in a laboratory setting is straightforward. However, measuring it in the global scientific record is notoriously difficult. Researchers often conflate "novelty"—producing something new—with "creativity." True creativity requires that the new idea is also meaningful and useful.

Existing bibliometric approaches (methods used to measure scientific literature) often fail to distinguish between these two. For instance, a paper might be highly novel because it combines two obscure fields (recombinant novelty). Yet, it might fail to have any lasting impact on the community. Conversely, a paper might introduce a brand-new concept or term (object novelty) that eventually reshapes a field.

Previous studies have struggled to scale these measurements to millions of papers. Many rely on simple keyword counting. This approach fails to capture the actual semantic depth (the underlying meaning) of a research contribution. Without a way to separate how an idea is built from how much it is valued, we cannot determine if AI use is linked to deeper thinking or simply more efficient mixing of existing ingredients.

A multi-modal framework for AI adoption

To move beyond the "AI vs. non-AI" binary, the authors implement a sophisticated classification system. This system treats AI as a layered infrastructure rather than a single tool. As illustrated in the dataset construction pipeline, the researchers use a two-stage GPT-SciBERT pipeline.

This process utilizes semantic embeddings (mathematical representations of text meaning) to filter the OpenAlex database. This ensures that papers are categorized by how they actually use AI.

The authors categorize AI-related research into three analytically distinct modes:

1. Tool-oriented mode: This involves applying existing, "off-the-shelf" AI models to specific domain tasks. An example is using a standard computer vision model to analyze biological cells.
2. Adaptation-oriented mode: This involves modifying or reconfiguring existing AI architectures to meet specific, domain-driven problems. One might tweak a transformer model to better handle genomic sequences.
3. Foundational mode: This focuses on the development of new AI methodologies themselves. These serve as the building blocks for other researchers.

By decomposing AI usage this way, the study tests whether different "modes" of engagement are associated with different types of creative outcomes.

Divergent pathways to novelty and impact

The empirical results reveal that AI adoption is broadly associated with higher levels of creativity. However, the mechanism of that association depends on the research mode. The authors report that AI-adoption publications are significantly more likely to rank in the top decile (the top 10%) of creativity compared to non-AI research.

Crucially, the paper finds a "crossover" pattern in how these gains manifest [, Figure 3]. Tool-oriented research is more strongly associated with recombinant novelty. Because these researchers bring computational models into traditional domains, they naturally bridge disparate literatures. This creates atypical combinations of knowledge. As shown in, Tool-mode research exhibits the largest gains in this area.

However, Tool-mode research shows the weakest association with object novelty. This is likely because these researchers often import AI models wholesale from other fields. When a model is imported without modification, the conceptual vocabulary of the receiving domain does not necessarily change. The novelty lies in the connection, not in the creation of new concepts.

In contrast, Adaptation-oriented research follows a different trajectory. By modifying AI to fit specific scientific inquiries, these researchers are more likely to generate new conceptual constructs and terminology. The authors find that Adaptation-mode research is associated with higher object-based novelty and higher long-term citation impact. Essentially, while tools are linked to mixing existing knowledge, adaptation is more closely linked to the introduction of new scientific concepts.

Limitations of the bibliometric lens

Several constraints must be acknowledged. First, the study focuses exclusively on AI as a methodological tool. It does not capture the "hidden" use of generative AI for administrative or communicative tasks. Examples include drafting manuscripts or polishing prose. These uses might influence perceived creativity. However, they leave no trace in the formal bibliographic record.

Second, the measures of creativity are indirect. Recombinant novelty (based on reference lists) and object novelty (based on noun phrases) are proxies. They are substitutes for the profound, often tacit, conceptual shifts that characterize true paradigm changes. A machine can detect a new word. It cannot yet fully grasp the weight of a revolutionary physical theory.

Finally, the results are tied to the peer-review ecosystem. The observed advantages may partially reflect how current evaluation systems reward the specific types of interdisciplinary "recombination" that AI facilitates.

The verdict: A conditional association

Does AI relate to scientific advancement? The evidence suggests a qualified yes. AI is not a uniform force that elevates all research in the same way. Instead, it acts as a powerful engine for expanding the boundaries of what is possible.

If the goal is to accelerate the synthesis of existing knowledge, Tool-oriented AI is more strongly associated with success. It excels at exploring new combinations of known variables. However, if the objective is to catalyze conceptual innovation, the data suggests that Adaptation-oriented AI is the mode most closely linked to the birth of new scientific frameworks. For policymakers and funding agencies, the takeaway is clear. Evaluating scientific merit in the age of AI requires a nuanced toolkit. Assessment frameworks must distinguish between the efficiency of recombination and the profundity of conceptual invention.

Figures from the paper