Law Reform Institute

My new piece in @lawfare explores how export controls can and cannot reduce the risks of open-weight models. More thought needs to go into how existing authorities apply to open weights. As the risks become more acute, there may be a temptation to improvise, reaching for export controls or related authorities to address the immediate security concern without having thought through a durable approach in advance. The June directive to Anthropic relied on the government’s leverage over a developer that still controlled access to its models. Open-weight AI presents a more complex problem. The article argues that export controls can help reduce the risks of open weights, but they are not a general regime for governing them. Managing those risks will necessarily require other domestic tools and increased international cooperation, particularly with China. Many thanks to Katherine Pompilio and all the folks at Lawfare for their thoughtful edits.

LRI posted a draft of my new white paper on its website this morning. The paper addresses the question of whether and how export-control and related authorities can reduce the risks posed by open-weight models. Export controls have taken on an unusual role in the Trump administration’s pre-release review of closed-weight frontier models. The administration has developed a voluntary framework for such models under EO 14409, and the June 2026 Anthropic export-control directive suggests that this framework is backed by coercive export-control authorities. Recent reports indicate the voluntary framework may be expanded to include open models as their capabilities narrow the gap with the closed frontier. However, a framework of pre-release review backed by the possibility of an export-control directive has less practical force with open-weight models. Any restraint directed at publication would raise statutory and First Amendment concerns, and foreign developers generally remain beyond the reach of such authorities. In addition, this approach loses most of its practical leverage at publication. After the weights are released, an export-control directive cannot recall downloaded copies or restore centralized control. This is not to say that the administration shouldn’t expand voluntary engagement to include advanced open-weight models; it clearly should. It should also carefully consider how it can effectively use export-control authorities to reduce the risk of open models. Under current conditions, the EAR’s “published” exclusion should remain available for general-purpose open-weight releases, and existing export-control authorities should not be stretched to restrain publication. If U.S. open models begin to provide threat-relevant capabilities beyond foreign alternatives, Congress should create purpose-built pre-publication authority, with clear standards, appropriate limits, and prompt judicial review. Existing authorities are most effective when focused on high-consequence output disclosures and on downstream services that materially assist WMD or military-intelligence activities or create demonstrated risks to U.S. systems. Such conduct-based controls can reduce risks relating to both U.S. and foreign open-weight models. The paper explains how the administration should—and should not—deploy these authorities. I encourage everyone with an interest in the topic to read the paper, not just this X post. We’re releasing it in draft form because we’re soliciting feedback and plan to post an updated version in about two weeks. I’m certain I haven’t gotten everything right, even after spending a lot of time thinking about these issues. I know others have been thinking hard about them too, so please send any comments you have by email to [email protected]. A link to the paper is in the comments.