Ro Khanna Proposes Bill to Ban Self-Improving AI
According to CNBC, Representative Ro Khanna plans to introduce a bill aiming to ban AI models capable of recursive self-improvement until federal safety measures are in place. He calls this a "civilizational extinction risk."
The Human Control Over AI Act would:
- Prohibit models that improve themselves or alter their core objectives, containment, or shutdown controls independently.
- Establish a federal agency to oversee and license the development and deployment of advanced AI models from companies like OpenAI, Anthropic, Google DeepMind, and xAI.
- Mandate safety protocols, including audits, testing, and human oversight, for these models.
- Impose criminal penalties on employees who disable safety features or deploy unauthorized systems.
- Target AI companies with liability insurance requirements before releasing models.
- Encourage agreements and export controls to prevent AI development in adversaries like China.
Khanna based the bill on discussions with AI safety nonprofits, emphasizing a community-driven approach.
He has also written to Chinese AI labs, seeking information about their work on superintelligence and recursive self-improvement, as part of efforts to forge an AI agreement with China.
Other proposals, like the FRONTIER Act, share similar goals, but no House bills are expected to be voted on before the midterm elections.
This comes on the heels of a warning from researchers concerning the potential dangers of AI automating its own research.