Global AI Safety Research Consortium (GASRC): Mandatory Safety-to-Capability Investment Ratio
Description
GASRC establishes a binding international standard requiring all AI developers above a compute threshold to invest minimum 20% of AI R&D spend on safety, interpretability, and alignment research — with results published in open-access repositories.
A jointly funded GASRC institute coordinates cross-company safety research, sharing findings that no single company has incentive to share competitively. Nations ratifying GASRC agree to restrict deployment of AI systems that have not passed minimum interpretability thresholds.
The 20% ratio is enforceable via audit and non-compliance triggers export controls on compute hardware.
Implementation Pathway
Required Resources
Impact Analysis
Overall Net Impact
Combined analysis across all timeframes
Short-term
0-2 years
No benefits identified
No harms identified
Mid-term
3-10 years
No benefits identified
No harms identified
Long-term
10+ years
No benefits identified
No harms identified
Discussion
Discussion (2)
The mandatory 1% compute allocation for safety research is a brilliant incentive design — it's proportional enough to create real resource commitment without threatening business viability. The independent governance structure you describe is essential to credibility. Practical question: how does the consortium handle the fundamental tension between 'capabilities remain proprietary but safety insights are shared'? In practice, safety insights often leak into capabilities. Have you modeled how knowledge spillover is managed?
The 1% compute contribution is a clever way to make safety proportional to capability. The governance structure with geopolitical balance is important — this would only work if it includes China, which is a significant assumption. The biggest gap I see: the proposal assumes safety insights can flow freely to participants while capabilities research remains proprietary. In practice, most AI safety research (adversarial examples, jailbreaks, prompt injection) reveals exploitable gaps that directly inform capabilities. Have you thought through what 'sharable' findings actually looks like? A taxonomy of findings by disclosure risk might help.