Artificial intelligence is changing from an auxiliary tool into an integral part of the infrastructure for science. Tasks that once demanded large, interdisciplinary teams, such as literature review, experimental design and model building, can increasingly be handled by smaller groups equipped with strong judgement and effective AI systems.

The question is no longer whether AI will increase science production, but how it will reshape the questions that scientists choose to ask. For instance, a 2026 Nature study used a pretrained language model to identify AI-augmented research in 41 million natural-science papers (Q. Hao et al. Nature 649, 1237–1243; 2026). It found that scientists who engaged in AI-augmented research published three times as many papers and received nearly five times as many citations as those who didn’t. Yet AI use was also linked to a 5% reduction in the range of topics studied and a 22% drop in collaboration.

The uncritical adoption of AI in science is alarming — we urgently need guard rails

Thus, AI might make it easier to do science while, at the same time, narrowing the questions and styles of reasoning that are collectively pursued.

I have seen this tension in my own cross-disciplinary work, which spans algorithm design, biological-data analysis and clinical studies. For example, decades of careful research have not resolved the question of whether depression is one disorder or a collection of conditions with similar symptoms. This is, in part, because the problem requires translation across disparate domains, including clinical symptom assessment, brain-imaging preprocessing, algorithm design and clinical validation. Conventionally, each step depends on different expertise, and progress often slows at the handovers between disciplines. AI can make this chain less fragmented by helping researchers to read papers outside their field, compare methodological choices and translate statistical patterns back into clinical terms.

That gain, however, is also where the risk lies. Once the chain becomes easy to automate, it can become a template for feeding ‘paper mills’ — an AI tool can pilot a literature search, identify significant associations and eventually write a polished manuscript. Instead of deeply investigating a question, researchers, or automated systems, can run the same pipeline across various data sets or topics and reliably produce publishable results. In effect, this makes it possible to industrialize research: producing many studies with similar methods and similar-looking conclusions. What is lost is the slower, more-critical work of questioning assumptions, exploring alternative explanations and asking whether the original question itself is well framed.

Why evaluating the impact of AI needs to start now

Competing Interests

The author declares no competing interests.