New Study Highlights Untapped Potential of AI Researchers in Scientific Discovery
Image Credit: CDC | Splash
In a recently published preprint titled Unlocking the Potential of AI Researchers in Scientific Discovery: What Is Missing?, researchers Hengjie Yu and Yaochu Jin from Westlake University and the Westlake Institute for Advanced Study in Hangzhou, China, shed light on the evolving role of artificial intelligence in scientific research. Released in March 2025, the study provides a detailed analysis of AI’s growing influence in high-impact scientific journals and underscores the need to better integrate AI researchers into the scientific discovery process.
[Read More: Deep Research Now Open to ChatGPT Plus Users]
AI’s Rapid Rise in Scientific Research
The study reveals a significant uptick in AI-driven research over the past decade, particularly within the prestigious Nature Index journals, a collection of 145 high-impact publications selected for their scientific reputation. According to the authors’ analysis of 20,401 AI-related articles retrieved from the Web of Science Core Collection, the presence of AI in these journals has surged ninefold since 2015. In 2024 alone, AI-related articles accounted for 3.57% of total publications in these journals—a proportion that, while still modest, marks a sharp increase from a decade ago, when such articles were far less common.
This growth is attributed to breakthroughs in machine learning and large language models (LLMs), which have transformed fields like structural biology, chemistry, and biomedicine. Notable examples include AlphaFold and ESMFold, which have redefined protein structure prediction, and tools like GPTCelltype and PathCha, which streamline cell annotation and pathology diagnostics, respectively. Despite this progress, the authors note that AI remains a minor component of overall scientific output, suggesting that its full potential is yet to be realized.
[Read More: Evo AI Revolutionizes Genomics: Designing Proteins, CRISPR, and Synthetic Genomes]
Forecasting AI’s Future in Science
Drawing on the Diffusion of Innovation theory—a framework that predicts how new technologies gain traction—the researchers project a dramatic increase in AI’s role in scientific discovery over the coming decades. The theory, developed by Everett Rogers, suggests that adoption follows an S-shaped curve, starting with innovators and early adopters before spreading to the broader population. Applying this model, Yu and Jin estimate that by 2050, AI-related research could constitute approximately 25% of all publications in Nature Index journals, up from its current 3.57%. This forecast hinges on the continued development of accessible AI tools and the active involvement of AI researchers in scientific endeavours.
The authors highlight that while fields like computational biology and materials science have embraced AI early on, broader adoption across other disciplines will drive this growth. However, they caution that the pace of this expansion may slow as AI matures and becomes a standard tool in research, a natural plateauing effect seen in the diffusion of past transformative technologies like computational modelling and high-throughput sequencing.
[Read More: Entalpic and Hugging Face Launch LeMaterial to Revolutionize Materials Science]
Who’s Leading the Charge?
A striking finding from the study is the dominance of experimental scientists in AI-driven research. An analysis of 80,945 author affiliations from the 20,401 AI-related articles shows that nearly 90% of this work is led by researchers from scientific institutions rather than AI-focused ones. While the involvement of AI institutions has grown—from appearing in 14.41% of papers in 2015 to 28.66% in 2024—their role remains largely supportive. The proportion of papers led by AI institutions, where they are listed as the first affiliation, has risen from 3.13% in 2015 to 8.37% in recent years, yet scientific institutions continue to steer the majority of projects, with their average first-author ranking holding steady at around 1.1.
This trend is even more pronounced in five well-known interdisciplinary journals—Nature, Nature Communications, Science, Science Advances, and Proceedings of the National Academy of Sciences (PNAS)—where AI researchers are more active but still rarely take the lead. In 2024, AI institutions contributed to 40% of articles in these journals, yet only 14.6% listed them as the first affiliation. The researchers suggest that this imbalance reflects a gap in AI researchers’ direct engagement with scientific discovery, a challenge they aim to address with targeted strategies.
[Read More: AI-Powered Animal Behaviour Research Transforms Neuroscience]
Strategies to Empower AI Researchers
To unlock AI’s full potential, the authors propose three key directions. First, they advocate for the development of user-friendly AI tools tailored for experimental scientists. Such tools, which could automate data analysis, modelling, and experiment design, are already in use by early adopters and could amplify AI’s impact if made more accessible. The study cites examples like SHAP (SHapley Additive exPlanations), a model interpretation method widely adopted across disciplines, as evidence of how effective tools can bridge the gap between AI and science.
Second, the authors call for AI researchers to take a more proactive role in scientific discovery by bridging cognitive and methodological gaps. Currently, only about 8% of AI-assisted research is led by AI experts, a figure that has plateaued in recent years. The study attributes this to a lack of understanding about where AI can be applied and how AI researchers can lead projects. To address this, they suggest focusing on fields with existing datasets or untapped potential, such as materials design and climate prediction, and developing specialized algorithms to tackle these challenges.
Third, the researchers emphasize the need for a thriving AI-driven scientific ecosystem. This involves fostering collaboration, establishing standards for AI use, and creating platforms that encourage innovation. The success of tools like SHAP, supported by accessible software packages, illustrates how such an ecosystem can accelerate progress. By prioritizing model explanation over marginal performance gains, AI researchers can provide insights that resonate with experimental scientists’ goals, further integrating AI into the research process.
[Read More: AI Detects Hidden Brain Waves: Early Identification for Dementia Patients]
Practical Approaches and Workflows
The study outlines four practical approaches to equip experimental scientists with AI: user-friendly platforms for data-rich analysis, domain-enhanced LLMs for data-scarce scenarios, automated data extraction from literature, and autonomous experimental systems powered by robotics. These methods aim to lower barriers to AI adoption while offering opportunities for AI researchers to contribute meaningfully. For instance, advances in multimodal LLMs could soon streamline the extraction of structured data from unstructured scientific texts, a task that remains challenging but holds immense promise.
Additionally, the authors propose a tailored AI4Science workflow for AI researchers, encouraging them to identify scientific problems, leverage existing datasets or literature, and collaborate with experimental labs for validation. This approach, they argue, positions AI researchers as key contributors rather than mere tool providers, a shift critical to realizing AI’s transformative potential.
[Read More: AI Breakthrough: Headband-Style Device Poised to Detect Alzheimer’s Years Ahead]
Challenges and Future Outlook
While optimistic, the study acknowledges hurdles such as AI “hallucinations”—where models generate inaccurate outputs—and the need for human oversight in scientific reasoning. Experimental validation, the authors stress, remains essential to ensure rigour, particularly as AI-driven discoveries expand. They also note the competitive pressures facing AI researchers, citing a quote from Kyunghyun Cho: “I sensed anxiety and frustration at NeurIPS’24,” reflecting the intense landscape they navigate.
In conclusion, Yu and Jin’s work underscores that while AI4Science has made remarkable strides, its future hinges on empowering AI researchers to lead rather than follow. With strategic interventions, the field could see a tenfold increase in impact by mid-century, reshaping how science is conducted.
[Read More: How Generative AI is Transforming Insomnia Management and Sleep Health Solutions]
Source: arXiv
