Summary
- A new paper titled "AI Alignment Is Neither Solvable nor Sufficient" by Gerrit Großmann argues that the standard framing of AI alignment is fundamentally misguided.
- The paper states that alignment is ill-defined due to the lack of a stable set of human values, and where specified, it remains unsolvable.
- Even if alignment were solvable, it would not necessarily prevent negative consequences because hybrid human-machine societies operate as complex systems with emergent risks.
- The author concludes that AI safety is an ongoing continuous challenge rather than a problem that can be permanently resolved by verifying individual model properties.
Why it matters
This paper challenges foundational assumptions in AI safety by suggesting that focusing solely on aligning individual models is insufficient for managing systemic risks in hybrid societies.