Feb 2, 2026
Modelling the impact of verification in cross-border AI training projects
Fabio Marinello
🏆 1st Place Winner
This paper develops a stylized game-theoretic model of cross-border AI training projects in which multiple states jointly train frontier models while retaining national control over compute resources. We focus on decentralized coordination regimes, where actors publicly pledge compute contributions but privately choose actual delivery, creating incentives to free-ride on a shared public good. To address this, the model introduces explicit verification mechanisms, represented as a continuous monitoring intensity that improves the precision of noisy signals about each actor's true compute contribution. Our findings suggest that policymakers designing international AI governance institutions face a commitment problem: half-measures in verification are counterproductive, and effective regimes require either accepting some free-riding or investing substantially in monitoring infrastructure.
When your theorems predict experimental results within 1-6% error, that's not a coincidence; that's a framework that actually works. The math seems to checks out and the experiments back it up across three different models.
But the practical concerns are real. The detector needs to know which tokenizer generated the text. In a world where every model uses a different tokenizer, that's a deployment headache that the paper waves away. You'd need some kind of registry or metadata standard, and that's a whole separate problem.
The overlap story is also worrying. At 0% overlap you get perfect detection. At 10% overlap, detection craters to 35%. There's no graceful middle ground. You're either getting full watermark strength with constrained vocabulary, or you're relaxing vocabulary and losing the watermark almost entirely. For production use, that cliff is a problem.
And the robustness testing needs to be tougher. Replacing random words with "masked" isn't what a real adversary does. A real adversary paraphrases, back-translates, or runs the text through a second LLM. The 30% replacement survival rate is encouraging, but it's answering an easier question than the one that matters.
The sparse watermarking idea they mention at the end (only watermark every N-th token) is where this probably needs to go for real-world use. That's worth a whole follow-up paper.
The introduction frames cooperation as potentially safety-enhancing, but I would like to see more discussion on, or explicit assumptions about, how this project contributes to AI safety (if that is the objective).
Currently, the project findings support international AI development efforts. This could diffuse race dynamics , and democratise frontier AI by involving middle powers in development. It would be good to be more explicit about whether, and under what conditions, these findings could also accelerate AI development or increase race dynamics.
Cite this work
@misc {
title={
(HckPrj) Modelling the impact of verification in cross-border AI training projects
},
author={
Fabio Marinello
},
date={
2/2/26
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
