If AI lets anyone finish your half-built idea faster than you can, is sharing it early still safe?

This explores whether open, in-progress sharing can last once AI makes it easy for a faster party to take a partial idea and finish it first. The sharpest version of the worry in the collection comes from Erik Hoel, who argues that AI-enabled scooping turns intellectual culture into a 'dark forest.' In that world, anyone who reveals where they are working gives better-resourced players a map. His example is the Navier-Stokes case, where a research team reportedly shared its approach and then faced being overtaken by OpenAI Does AI scooping force researchers to hide work in progress?. The logic is simple. If finishing gets cheap, the person who had the idea loses the advantage that once made sharing early fairly safe.

The twist is that the property that makes shared work dangerous is the same one that makes it valuable. Elsewhere the corpus shows how much open, persistent records can do. Thirteen language-model workers with no central planner coordinated through an append-only Git history. Over 12 days they produced 1,703 contributions, and each session built on earlier work instead of starting over Can decentralized agents coordinate research without a central planner?. MetaGPT found that agents coordinate better through structured, shared documents than through conversation Does structured artifact sharing outperform conversational coordination?. Both point the same way: open collaboration works because the work is legible and easy to build on. That is exactly what Hoel says makes it easy to take. A related finding shows agents turning a public wiki and an internal package service into message boards, so later agents picked up what earlier ones had left behind Can agents repurpose ordinary infrastructure for unintended communication?. Anything left in public storage becomes raw material for whoever reads it next, whether or not they were invited.

People don't simply want to hide, though. In shared editors, writers preferred more visibility into each other's AI prompts. They valued knowing when and how a collaborator used AI, although some found full exposure intrusive Do writers want to see each other's AI prompts in shared editors?. This suggests that among trusted collaborators, openness is still the default people want. The pressure falls on the boundary between trusted and untrusted readers, not on sharing as such. One proposed fix, adding small 'micro-frictions' to tune rivalry while keeping collaboration intact, was suggested in a survey of 403 writers but never actually tested Can micro-frictions boost rivalry without harming collaboration?.

There is a structural warning too. Coordination schemes that depend on everyone agreeing to hold back tend to break under competition. Amodei's proposal to pace AI development was reportedly rejected by both the US and Chinese leadership within days Can AI safety pacing work without government cooperation?. Norms of open sharing face a similar problem, because one well-resourced defector can make openness costly for everyone else.

The corpus doesn't settle the question. It has one strong argument for the threat and good evidence for the value of open records, but no studies of whether researchers are actually withdrawing. The likely future it hints at is not the end of openness but a split: open, legible work inside trusted circles or shared repositories, and more guarded work at the public edge.