If sharing your work might leak to an AI company, what do you do? A simulation, run two ways.
Report 009 · Dr Dan Epstein · version 0.3 · signed 19 October 2026 A Strategy Soup Scenario
TLDR
In September 2026, a mathematician said OpenAI’s new maths result could have drawn on his drafts in OpenAI’s own AI tool. OpenAI said his recent prompts in its tool could not have influenced its system.
Researchers now write their work inside AI tools made by AI companies that are racing them. Work in those tools might reach the company, through its systems or its staff. Work you share with a colleague sits in their tool too.
We built a computer simulation of a three-year race between university groups and AI companies, where that can happen and nobody sees it.
In the live run, each university group chose its own moves and gave its reasons.
With no fixes on offer, the groups kept sharing and published their work. The AI companies, with more computers, read it and got to the answers first: 7.9 of the 8.5 steps on the problem reached in an average run.
With five real fixes on offer, the groups moved to safer AI tools, recorded their work privately and kept sharing with each other. Together they reached 6.0 of the 8.9 steps in an average run, against 0.6 without the fixes.
The cost: they published much less until the very end.
How the simulation works
We are The Long Game Project, a small independent studio. The simulation is a model, not a forecast.
Three made-up university groups race two made-up AI companies for three years on one hard problem. The problem is solved in steps, and whoever reaches a step first gets the credit for it. Each quarter, each group chooses what to share and which AI tool to use. Every run starts just after news like the real dispute.
Live run, no fixes8
Live run, with the five fixes8, with the same chances
Rule-based run, groups following rules we wrote200 per condition
The five fixes
University accountThe same AI tool, on terms that say your work is not used to train the company’s AI
Local modelAn AI model on your own computers, slower but private
SwitchingMoving to the other company’s tool, if you trust it more
Open projectEveryone posts everything publicly into one shared project and splits the credit
TimestampA digital record that proves when your work existed, without publishing it, so you can show you were first
What happened
Old movesThe groups keep sharing and publish often. The AI companies reach almost every result.
With the fixesThe groups move to university accounts and local models, timestamp their work and share it only with each other. They rarely switch or use the open project.
End of the simulationWith the fixes, the groups reach most results and publish everything in the last quarter.
The numbers
Who reached the resultsFig. 1
With the fixes, the university groups reached most of the results.
Results reached in three years, average per run, by who reached them.
What it costFig. 2
With the fixes, the groups published far less until the very end.
Work the university groups had published, in units of work, average per run.
Both figures come from the same 16 runs. Every rule in the simulation is a choice we made.
What it means
The fixes protected the groups without breaking up their teamwork. But the timestamp only works because journals in the simulation accept it as proof of who was first, and real journals do not yet. And the groups published far less until the end, which would hurt any researcher outside the three groups. The rule-based groups, which stop sharing when worried, lost about two steps a run to the fear, and no fix brought their sharing back. This is a model of choices, not a prediction about people.
The line
Guard the tool before you stop trusting your colleagues.
A Long Game Project lab