The U.S. government took action on June 12 through an export control order directed at Anthropic, compelling the company to retract its powerful AI models, Mythos 5 and Fable 5. Officials believed these models could facilitate attacks on U.S. infrastructure by adversaries. Thus, the order restricted Anthropic from sharing these models with foreign entities or nationals, including those within the U.S.
Anthropic faced challenges in discriminating access based on nationality or citizenship status, leading to the withdrawal of the models from public access. Shortly after, at the behest of the administration, OpenAI restrained the launch of its latest model, GPT-5.6, fearing an export ban.
Within weeks, the administration partially lifted the restrictions on Mythos 5, allowing a limited group of U.S. users to access it. By June 30, Anthropic received communication that licences for Mythos or Fable were no longer mandatory. Anthropic released Fable publicly but continued to limit Mythos to select U.S. organizations.
OpenAI’s GPT-5.6 received approval for wide release on July 8. The Congressional intent behind export control orders aligns with national security and foreign policy goals, but in this instance, the aim was evidently to enforce a consumer product recall. Initially, Anthropic was asked to voluntarily retract the model and work with the government on security issues. An official noted export controls as a “last resort.” Anthropic claimed they had “90 minutes to take the model down.” This led to a “de-facto licensing regime.”
Public approval remains uncertain. Martin Chorzempa from the Peterson Institute remarked on the unpredictability of the U.S. government potentially halting future model access. Despite reestablishing access to Fable and Mythos, the administration maintained its right to reimpose licensing requirements.
Jessica Tillipman at George Washington University emphasized that whether Anthropic’s models posed a cyber threat should be resolved through fair processes, not stringent ultimatums. Law professor Alan Z. Rozenshtein argued that the Export Control Reform Act of 2018 allows government intervention without due process. While defining AI model access as legal “export” remains debatable, existing laws authorize secretive “is informed” letters requiring export licenses.
The export control system largely bypasses the Administrative Procedure Act’s public notice requirements. Consequently, affected parties cannot legally contest any arbitrary actions by the government. This unchecked authority undermines U.S. tech firms abroad, who struggle to assure customers of sustained AI model access.
A French presidential candidate described this secretive power over AI models as America’s Strait of Hormuz. Transparent and accountable procedures for export control decisions are crucial for U.S. AI companies’ global reliability. Furthermore, Congress should consider refining the export control statute to align with AI regulation, ensuring public notice and comment.
Implementing Administrative Procedure Act protections could mitigate arbitrary regulatory actions. Guidelines similar to those of safety administrations might inform policymakers in imposing controls on AI models. Establishing public accountability and review processes is essential to prevent misuse.
The administration highlighted the need for reform, revealing the extensive authority granted by export control statutes. Prompt Congressional action is necessary to address these issues. Mark MacCarthy, the author of “Regulating Digital Industries,” serves as adjunct professor at Georgetown University’s Communication, Culture and Technology Program and as a senior fellow at relevant institutions.
Leave a Reply