If sharing a half-finished research idea means AI might finish it first, do people just stop sharing?

This explores whether AI-enabled 'scooping', where an AI system finishes someone's partly shared idea before they can, narrows which research questions get worked on in public. The short version is that the collection has a strong argument that scooping pushes work out of public view, and separate evidence that AI is already narrowing research topics. No study here directly connects the two. That missing link is worth knowing about before you read further.

The scooping argument comes from Erik Hoel, who describes culture turning into a 'dark forest' Does AI scooping force researchers to hide work in progress?. Once AI can take a half-public idea and complete it faster than the person who had it, sharing early becomes a risk. Hoel points to a team working on the Navier-Stokes equations who allegedly got scooped by OpenAI after describing their approach. His prediction is about visibility rather than topic choice: researchers keep working on what they care about but stop showing work in progress. Even so, that changes what the public record contains. The open record ends up showing finished results, while the early, speculative ideas that others could build on stay private.

The topic-narrowing evidence comes from a different direction. A large study found that scientists who use AI publish about three times more papers and get nearly five times more citations, yet science as a whole covers about 4.6% fewer topics and has 22% less collaboration Does AI help individual scientists while narrowing scientific focus?. The mechanism isn't scooping. AI pulls work toward problems that already have plenty of data. Other notes point to the same pull toward familiar ground. When frontier AI agents are given long research tasks, they mostly recombine known techniques and rarely invent new ones Do frontier AI agents actually conduct novel research or just optimize?. And a fine-tuned model can now predict which of two research ideas will work better than expert researchers can, by learning from past results Can machines learn to predict which research ideas will work?. If people use tools like that to choose projects, everyone gets nudged toward ideas that resemble past successes.

Putting these together gives a sharper answer than either note gives alone. Scooping risk is probably highest for the kind of work AI does well: incremental, recombination-style ideas in areas with lots of data, where an AI can finish a partial idea quickly. That suggests the pressure to hide work falls hardest on ideas that are already crowded. Truly unusual ideas may be safer to share because AI is bad at completing them. DeepMind researchers add a twist. They argue that the bottleneck in AI-driven science is now physical lab capacity rather than ideas, and propose a market that would license ideas and pay for validating them Could markets allocate scarce lab resources to AI-generated research ideas?. If ideas really are cheap and execution is scarce, scooping matters mainly where execution is also cheap, such as computational work, and much less in wet-lab science.

There is also a feedback loop. A survey of 230 publications describes AI's effects on research production and peer review as a linked arms race: more papers, automated review, manipulation, defenses, and evasion Does AI create a coupled arms race in research production and review?. Hidden prompts that tell AI reviewers to praise a paper are one small, concrete example Are hidden AI prompts in preprints a deceptive research practice?. Secrecy driven by scooping fear would be one more move in that loop. Overall, the corpus supports 'AI narrows topics' and 'scooping pushes work out of public view' as separate claims. Whether scooping itself narrows topics remains an open question that nobody here has measured.