The White House's October 2023 Executive Order on AI governance introduced a tiered approach to regulating artificial intelligence, distinguishing between closed and open models. Under this framework, open-source AI models released by U.S. companies are largely exempt from extensive government review, provided they meet basic safety thresholds.
The policy aims to preserve U.S. leadership in AI by avoiding overregulation of publicly available models, which are seen as essential for research, education, and startup development. However, critics argue that open models could still be misused for disinformation or cyberattacks, even without centralized control.
This exemption reflects a strategic balance between fostering innovation and managing risk. Closed models—those whose weights and architecture are not publicly available—are subject to stricter oversight, including reporting requirements to agencies like the Department of Commerce. Open models, by contrast, face lighter-touch regulation, relying instead on industry self-governance and transparency measures.
Key distinctions in the EO:
Closed models: Must report development practices, risk assessments, and deployment plans to federal agencies.** Open models**: Exempt from mandatory federal audits unless they pose national security risks or are used in sensitive applications like critical infrastructure.
The policy aims to preserve U.S. leadership in AI by avoiding overregulation of publicly available models, which are seen as essential for research, education, and startup development. However, critics argue that open models could still be misused for disinformation or cyberattacks, even without centralized control.
For developers and organizations working with open AI systems, this means fewer bureaucratic hurdles but continued responsibility for ethical deployment. The emphasis remains on proactive risk mitigation rather than reactive compliance. Next Chinese Labs Release 10 Open Models in 30 Days — US Startups →
All Replies （3） #
T
I've been testing local fine-tuning of LLaMA models, and the closed-model restrictions really do impact deployment options for small teams.
0
D
Running quantized Mistral on a 3090 at home—closed model limits kill reproducibility for indie devs.
0
C
Fine-tuning LLaMA locally still costs a fortune and eats GPU hours—don't see how these guidelines help small teams survive.
0