Generative AI Is Reshaping Scientific Ideation, but Experience Determines Who Benefits

A 2026 study published in Research Policy provides compelling evidence that generative AI is fundamentally altering how scientific ideas are born—and that its impact is strikingly uneven across career stages. In a randomized experiment involving 310 researchers, Matthias Trobinger, Anil R. Doshi, and Sen Chai examined how scientists responded when AI systems proposed new research directions based on their prior work. The findings reveal a technology that may narrow gaps in idea production while simultaneously widening the importance of evaluative judgment.

AI Moves Feedback Earlier in the Scientific Process

The study found that AI influenced the content of ideas researchers produced, but experience strongly moderated that influence. Less-experienced researchers were considerably more likely to incorporate AI-generated suggestions, interpret them as novel, and perceive the resulting ideas as impactful. More-experienced researchers, by contrast, were more likely to regard the same suggestions as obvious, insufficiently advanced, or merely another tool.

"The most important implication is not that one group trusted AI and the other did not," the authors note. "It is that generative AI appears to be moving scientific feedback earlier in the process." Previously, researchers had to decide whether an idea was coherent enough to present to another person. AI substantially lowers that threshold, allowing a researcher to externalize an idea while it is still incomplete, contradictory, or poorly articulated.

The Emergence of a Pre-Supervisory Layer

For early-career researchers, generative AI partially compensates for structural disadvantages—smaller professional networks, less confidence approaching established experts, and less accumulated experience transforming vague questions into defensible projects. It offers on-demand preliminary discussion without scheduling constraints, social pressure, or reputational risk.

This creates what the study describes as a pre-supervisory layer. AI does not replace the supervisor; instead, it changes what reaches the supervisor. If researchers begin submitting more coherent and developed proposals, senior scientists may spend less time helping formulate basic questions and more time evaluating strategic importance, methodological rigor, and resource allocation.

However, the study also identifies a significant risk: a polished AI-assisted proposal may create the appearance of intellectual maturity without the underlying reasoning skills normally developed through constructing it. "Supervisors may receive better-presented ideas while gaining less visibility into how well the researcher understands them," the researchers caution. The result could be an emerging gap between proposal quality and researcher capability.

Expertise Shifts From Generation to Rejection

When possible research directions were expensive to generate, producing a strong idea was itself a defining capability. Generative AI dramatically lowers the cost of producing options, and that abundance changes the function of expertise. The more valuable capability becomes rejection.

Experienced researchers in the study frequently dismissed AI suggestions not because they were incoherent, but because they were not sufficiently important. The ideas were sometimes logical extensions of prior work, yet still failed to cross the threshold of being worth pursuing. This reflects a form of expertise that current AI systems struggle to reproduce—tacit knowledge about which questions have already been explored informally but never published, which theoretically attractive projects are operationally impossible, and which methods will fail under real-world constraints.

"Even an LLM with broad access to published research may generate a plausible direction while lacking the tacit and unpublished knowledge explaining why it was never pursued," the study notes. This suggests that the future role of senior researchers may become increasingly similar to that of portfolio managers, allocating attention and resources among an expanding supply of AI-generated possibilities.

The Scientific Monoculture Risk

A parallel concern emerges from a 2026 Nature study that used a pretrained language model to identify AI-augmented research across 41 million natural-science papers. That analysis 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 did not. Yet AI use was also linked to a 5% reduction in the range of topics studied and a 22% drop in collaboration.

If large numbers of researchers use similar models to generate research questions, those researchers may receive overlapping suggestions—creating a form of scientific convergence that expands the number of ideas while reducing intellectual diversity. Individually, each suggestion may appear novel. Collectively, the system may concentrate attention around the same methods, variables, and theoretical frameworks.

Designing AI as an Adversary, Not Just an Assistant

The study's authors argue that scientific ideation may require a different design philosophy than most current generative AI products, which are optimized to be helpful. The most valuable research assistant may not be the system that generates the most convincing proposal, but the system that most effectively tries to destroy it.

A research-focused AI, they propose, should operate in multiple modes: as a hypothesis generator proposing competing explanations, an adversarial reviewer identifying why a proposal may be trivial or infeasible, a novelty auditor checking whether an idea is genuinely underexplored, a constraint simulator testing against practical limitations, a provenance tracker documenting which elements originated from the researcher versus the model, and a divergence engine intentionally searching outside dominant literature for analogies from distant disciplines.

Rethinking How Researchers Are Trained

The emergence of AI-assisted ideation also creates a training problem. Scientific education has traditionally assumed that researchers develop judgment through repeated exposure to uncertainty—generating weak ideas, misunderstanding literature, receiving criticism, and discovering why apparently promising directions fail. AI can remove some of that friction. Removing unnecessary friction is beneficial; removing developmental friction may not be.

The study suggests universities should avoid framing AI policy as a binary choice between permission and prohibition. Possible requirements could include mandating that researchers critique AI-generated proposals before adopting them, compare AI-assisted ideas with independently generated alternatives, document why specific AI suggestions were rejected, and have supervisors evaluate the reasoning process rather than only the final proposal.

The organizations that benefit most, the researchers conclude, will not be those that deploy AI to generate the greatest number of ideas. They will be those that build the strongest systems for criticism, selection, provenance, and intellectual diversity.