General Purpose AI Will Never Be Safe
Authorities should require AI companies to 1) limit the purposes of their AI systems; 2) respect sector-specific rules; 3) be held accountable for the harm their AI systems cause | Edition #267
TL;DR
- General-purpose AI chatbots (e.g., ChatGPT, Gemini, Claude) break the rule of not combining distinct services like financial advice, psychotherapy, and storytelling into a single product.
- These diverse services have different user groups, risk profiles, and legal requirements, making a one-size-fits-all regulatory approach for general-purpose AI impossible.
- The lack of context-specific rules for general-purpose AI leads to inadequate user protection and increased harm, particularly in areas like mental health.
- AI companies can exploit the difficulty of regulating general-purpose systems to avoid accountability for the harms their products cause.
- The proposed solution is to mandate purpose-specific AI chatbots, requiring companies to define target audiences, use cases, and implement appropriate safety standards and guardrails for each.
- This approach would allow for proper regulation and certification of AI systems based on their intended function and associated risks.
- The debate between Elon Musk and Sam Altman highlights the difficulty of regulating general-purpose AI due to its broad applicability and diverse risks.