Generative AI Is Changing How Scientific Ideas Begin

For most of modern science, the first stages of a research idea were largely invisible.

A researcher noticed an anomaly, connected two concepts, questioned an established assumption, or identified a gap in the literature. That idea was privately refined before being tested through conversations with supervisors, colleagues, reviewers, and funding bodies.

Generative AI is inserting itself into that sequence.

Researchers can now present an incomplete thought to a large language model before discussing it with another person. They can ask the system to generate hypotheses, challenge assumptions, identify adjacent fields, or convert an intuition into a more coherent proposal.

This may appear to be another productivity improvement. It is more significant than that.

Generative AI is creating a new intermediary layer between private thought and human scientific judgment. That layer could change how ideas are formed, who receives early intellectual support, how researchers are trained, and where expertise creates value.

A 2026 study¹ published in Research Policy provides evidence that this transformation is already unevenly distributed. Less-experienced researchers were considerably more likely to incorporate AI-generated suggestions and view resulting ideas as novel or impactful. More-experienced researchers were more likely to regard the same suggestions as obvious, insufficiently advanced, or merely another tool.

The findings point to a broader shift in scientific work. As AI makes idea generation abundant, the scarce capability may no longer be producing possibilities. It may be knowing which possibilities deserve to survive.

AI Is Changing the Sequence of Scientific Thinking

The study by Matthias Trobinger, Anil R. Doshi, and Sen Chai examined how researchers responded when generative AI proposed new research directions based on their prior work.

In a randomized experiment involving 310 researchers, participants either received no visible AI suggestions, one AI-generated suggestion, or three AI-generated suggestions. A follow-up study with 28 researchers explored how they used AI interactively during early-stage ideation.

The results showed that AI influenced the content of the ideas researchers produced. However, experience strongly moderated that influence.

Less-experienced researchers were more likely to incorporate AI suggestions, interpret them as novel, and perceive the resulting ideas as impactful. More-experienced researchers were less influenced and often believed that the suggested directions had already been considered, were insufficiently ambitious, or lacked the context required to identify a genuinely valuable research problem.

The most important implication is not that one group trusted AI and the other did not.

It is that generative AI appears to be moving scientific feedback earlier in the process.

Previously, researchers often had to decide whether an idea was coherent enough to present to another person. AI substantially lowers that threshold. It allows a researcher to externalize an idea while it is still incomplete, contradictory, or poorly articulated.

That creates a fundamentally different workflow.

The Real Disruption Is Pre-Supervisory AI

Early-career researchers often face a structural disadvantage that has little to do with intelligence or effort. They have smaller professional networks, less confidence approaching established experts, and less accumulated experience transforming vague questions into defensible projects.

Generative AI partially compensates for that disadvantage.

It offers on-demand preliminary discussion without scheduling constraints, social pressure, or reputational risk.

This creates what could be described as a pre-supervisory layer.

AI does not replace the supervisor. Instead, it changes what reaches the supervisor.

That distinction has major implications for research organizations. 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.

In theory, this could improve the efficiency of mentorship. Experienced researchers could focus their limited attention on higher-value judgment rather than preliminary structuring.

However, it could also conceal weaknesses.

A polished AI-assisted proposal may create the appearance of intellectual maturity without the underlying reasoning skills normally developed through the process of constructing it. Supervisors may receive better-presented ideas while gaining less visibility into how well the researcher understands them.

The result could be an emerging gap between proposal quality and researcher capability.

Expertise May Shift From Generation to Rejection

Generative AI dramatically lowers the cost of producing options.

That abundance changes the function of expertise.

When possible directions were expensive to generate, producing a strong idea was itself a defining capability. When possibilities can be generated almost without limit, 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.

Senior researchers often know:

• Which questions have already been explored informally but never published.
• Which theoretically attractive projects are operationally impossible.
• Which methods will fail under real-world constraints.
• Which findings would be technically valid but scientifically unimportant.
• Which research areas appear open only because the relevant work is underway but not yet public.
• Which questions are unlikely to survive peer review or attract funding.

Much of this knowledge is tacit. It exists in professional networks, failed experiments, conference discussions, abandoned grant proposals, reviewer comments, and years of exposure to the field.

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.

This suggests that the future role of senior researchers may become increasingly similar to that of portfolio managers. Their value will lie not only in producing ideas, but in allocating attention and resources among an expanding supply of AI-generated possibilities.

AI Could Flatten Access While Widening Judgment Gaps

Generative AI is often described as a democratizing technology because it gives more people access to capabilities previously associated with specialists.

That is partly true in science.

Researchers without elite networks or highly available supervisors can now access immediate brainstorming, explanation, synthesis, and criticism. This could help researchers at smaller institutions, in underfunded regions, or outside dominant academic networks formulate stronger projects.

Yet equal access to AI does not produce equal outcomes, creating a potential paradox: AI may narrow the gap in idea production while widening the importance of evaluation skill.

The researchers most influenced by AI-generated suggestions may also be the least equipped to identify their limitations. Meanwhile, those best positioned to evaluate the suggestions may perceive less immediate value in using them.

The result could be a new form of inequality in scientific work: not unequal access to ideas, but unequal ability to filter them.

The Scientific Monoculture Problem

There is another risk that extends beyond individual researchers.

Large language models generate outputs by learning patterns from existing literature. 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.

The study does not directly demonstrate this system-level outcome. However, its findings make the possibility more credible. Less-experienced researchers were more likely to incorporate AI suggestions, and early-career researchers represent a substantial share of future scientific production.

If AI becomes an intermediary during the formative stages of their careers, the models may influence not only individual projects but also the boundaries of what emerging researchers consider worth studying.

Research AI Should Be Designed as an Adversary

Most generative AI products are optimized to be helpful. They answer questions, expand ideas, and produce polished output.

Scientific ideation may require a different design philosophy.

The most valuable research assistant may not be the system that generates the most convincing proposal. It may be the system that most effectively tries to destroy it.

A research-focused AI should operate in several modes:

Hypothesis Generator

The system should propose competing explanations rather than a single preferred direction. It should vary assumptions, theoretical frameworks, and methodological approaches.

Adversarial Reviewer

It should identify why a proposal may be trivial, infeasible, already explored, statistically underpowered, ethically problematic, or unlikely to produce meaningful evidence.

Novelty Auditor

It should retrieve current papers, preprints, patents, clinical trials, grant databases, and conference proceedings to determine whether the idea is genuinely underexplored.

Constraint Simulator

It should test the proposal against practical limitations such as sample access, equipment requirements, cost, timelines, regulatory approvals, and data availability.

Provenance Tracker

It should document which elements originated from the researcher, retrieved literature, external collaborators, or model-generated suggestions.

Divergence Engine

It should intentionally search outside the dominant literature and present analogies from distant disciplines rather than continually reinforcing the most statistically likely path.

This would reposition AI from an automated idea factory to an intellectual stress-testing system.

Universities May Need to Rethink How Researchers Are Trained

The emergence of AI-assisted ideation creates a training problem.

Scientific education has traditionally assumed that researchers develop judgment through repeated exposure to uncertainty. They learn by generating weak ideas, misunderstanding literature, receiving criticism, revising assumptions, 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.

If junior researchers rely on AI to structure every early idea, they may produce stronger short-term outputs while accumulating less independent problem-formulation experience.

Universities should therefore avoid framing AI policy as a binary choice between permission and prohibition. The more important question is how AI use should be integrated into the development of expertise.

Possible requirements could include:

• Researchers must critique an AI-generated proposal before adopting it.
• AI-assisted ideas must be compared with independently generated alternatives.
• Students must document why specific AI suggestions were rejected.
• Supervisors should evaluate the reasoning process, not only the final proposal.
• Research methods courses should include calibration exercises using plausible but flawed AI output.

The objective would not be to preserve an artificial AI-free model of research. It would be to ensure that AI accelerates the development of researchers rather than substituting for it.

Conclusion

Generative AI is not simply helping scientists generate more ideas. It is changing when ideas become social, who receives early intellectual support, and where expertise creates value.

For researchers earlier in their careers, AI can function as an always-available preliminary collaborator. It can reduce the social cost of presenting incomplete thoughts and help transform intuitions into proposals that are ready for human discussion.

For experienced researchers, the value may lie elsewhere. Their advantage increasingly rests in recognizing which plausible ideas are unimportant, which elegant projects are infeasible, and which apparent gaps are not gaps at all.

This points toward a new division of labor in science.

AI will make research possibilities abundant. Human expertise will determine which possibilities deserve time, funding, and attention.

The organizations that benefit most 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.

The future of AI-assisted science will depend less on whether machines can suggest research questions and more on whether humans preserve the ability to reject the wrong ones.

References:

^{1. M. Trobinger, A. R. Doshi, and S. Chai, How experience moderates the impact of AI suggestions on researchers’ perceptions of their ideas, Research Policy 55 (2026) 105575, https://doi.org/10.1016/j.respol.2026.105575}