As artificial intelligence continues to advance, policymakers face a critical regulatory dilemma surrounding AI model weights—the underlying intellectual components that enable systems to converse, generate images, and write code. While major frontier labs typically keep these weights private and accessible only through APIs, open-weight models with publicly available parameters are surging in popularity. This creates a regulatory challenge: open models reduce industry concentration, but they also make it significantly harder for the government to oversee technology as capabilities grow more dangerous.

To address potential risks, regulators may soon seek to license the publication and distribution of capable AI model weights. Because the federal government lacks a dedicated statutory framework for this task, the export control regime stands out as its most realistic and immediately available tool. Applying this mechanism to domestic technology raises serious legal and structural questions, highlighting the need to evaluate model weights through the lens of constitutional protections.

The Export Control Regime as a Regulatory Vehicle

If the government moves to license AI model weights, the process will likely rely on the Export Administration Regulations (EAR), administered by the Bureau of Industry and Security (BIS) within the Commerce Department. A precedent for hasty licensing already exists; in June 2026, the Commerce Department effectively banned Anthropic’s capable closed-weight models, Mythos 5 and Fable 5, using export control licenses.

With minor adjustments, the export control apparatus could be configured to require pre-publication licensing for model weights. Although the EAR currently exempts published information from its controls, BIS could carve model weights out of that exemption. Under current rules:

  • Releasing model weights to a foreign person—whether located abroad or inside the United States—constitutes an export.
  • Publishing weights on the open internet makes them instantly accessible to foreign persons globally, meaning labs would struggle to restrict access solely to Americans.
  • Under deemed-export rules, sharing controlled technical information with foreign-person employees at domestic companies can also trigger licensing requirements.
  • Broader mechanisms, such as the Foreign Direct Product Rule, could potentially be extended to restrict foreign model weights trained using American algorithms, chips, or data.

Risks of Unchecked Regulation and Market Concentration

Employing the export control regime for domestic AI oversight carries profound risks. Designed primarily as a foreign policy tool, the export control framework lacks the procedural safeguards traditionally demanded of speech-licensing regimes. If utilized without adaptation, it would allow the government to restrict model weights quickly, without published standards, stated reasons, or immediate judicial checks.

This lack of oversight creates several dangers for the broader technology ecosystem:

  • Arbitrary Delays: The absence of enforceable review deadlines means the government could stall the release of model weights indefinitely.
  • Entrenched Monopoly Power: Because closed-weight providers never distribute their raw weights, they would avoid many of these restrictions. Halting open-weight distribution for even a year or two could cripple domestic momentum and cement closed-model dominance.
  • Foreign Alternatives: Stifled domestic open-weight development could inadvertently drive users toward foreign alternatives, such as Chinese open-weight models operating outside U.S. jurisdiction.
  • Regulatory Capture: Closed-weight providers could become the chief economic beneficiaries of a licensing bottleneck, reducing industry pressure to build a balanced, modern regulatory framework.

A First Amendment Approach

Because model weights are trained on speech and are capable of generating speech, they may themselves be protected as speech. Whether or not courts ultimately codify that classification, treating model weights as though they are protected by the First Amendment offers a practical path forward.

Building First Amendment guardrails into any licensing framework would require the government to ensure transparency, issue determinations in a timely manner, and provide judicial recourse for developers facing denied license applications.