Loading issue research
A dangerous-capability threshold should trigger safety evaluation before an irrevocable open-weight release, since a closed model can still be patched afterward and an open one cannot.
Verification Status
AI-researched, unverifiedLast Reviewed
Jul 4, 2026
Cited Sources
8
What is failing, what we would change, and the conclusion we are willing to defend.
Current federal policy treats open-weight release as geostrategically valuable. The 2025 "AI Action Plan" explicitly frames open models as tools for extending American standards globally, and export-control rules have historically exempted openly-released model weights from the controls applied to closed ones. The party's position doesn't fight that framing. It adds one precise safeguard current policy is missing: release timing. A closed model that turns out to have a dangerous capability can be patched, restricted, or pulled after the fact, as the 2026 Fable 5/Mythos 5 case showed, however messy that process was. An open-weight model cannot. Once weights are public, there is no recall. Researchers demonstrated exactly what that means in practice: stripping Llama 3's safety training took minutes to half an hour on a single consumer GPU, and WormGPT, a tool built specifically for phishing and malware, started in 2023 as someone's fine-tune of an open-weight model. A closed model can be patched after a demonstration like that. An open one can only be watched.
Use the same capability-based evaluation threshold from AI-02 to trigger pre-release review: not a separate, redundant "open-weight tax" applied regardless of what a model can do.
For a model that crosses that threshold, require the evaluation to complete before an irrevocable open-weight release specifically: a narrow, capability-triggered release-timing rule with no blanket restriction on open release.
Continue supporting open-weight competitiveness as policy, consistent with the platform's core value of open-source development and collaboration. A specific release narrowing a geopolitical capability gap is a national-security monitoring question for AI-07. Domestic open-weight releases remain generally permitted.
Support narrow, procurement-specific restrictions on federal use of adversary-origin open-weight models, distinct from restricting Americans' ability to release or use open-weight models generally.
Fund independent, ongoing monitoring of the open-weight risk/benefit balance, so "benefits currently outweigh risks" receives continuous review on a defined cadence.
A dangerous-capability threshold should trigger safety evaluation before an irrevocable open-weight release, since a closed model can still be patched afterward and an open one cannot.
Turn frustration into useful pressure.
If this position misses evidence or a lived consequence, challenge it. If it holds up, help test it locally and connect it to the issues around it.