Open-Source AI Is Uniquely Dangerous

Abstraction: Unsecured open-source AI uniquely risky because safety features cannot be re-patched

Key points:

• "Unsecured" (open-source) AI is uniquely dangerous because safety features can be stripped by anyone; once an "uncensored" derivative like "Llama 2 Uncensored" is published on Hugging Face, the original maker is powerless to stop it
• Secured AI (ChatGPT, Bard, Claude) can monitor abuse, patch jailbreaks, and update safety guardrails continuously; unsecured models cannot be updated after release
• Meta released Llama 2 with a 27-page Responsible Use Guide that was immediately ignored by third-party derivative creators
• Key risks: CBRN weapons assistance, mass personalized misinformation, nonconsensual deepfake pornography, election interference
• Author proposes 9 regulatory actions: registration/licensing above capability threshold, liability for "reasonably foreseeable misuse," mandatory watermarking, training data transparency, "know your customer" for AI hardware sales
• Gary Marcus (NYU): "I think how we regulate open-source AI is THE most important unresolved issue in the immediate term"

Connections: Meta · Openai · Hugging Face · AI Safety · Open Source AI · AI Regulation

Source: https://spectrum.ieee.org/open-source-ai-2666932122