One harm that seems super easy to fix, that I think the law should insist on: these models should show their thinking.
Making it impossible to see & judge & assess what the AI is doing is a grevious insult and harm. To sever us from the thinking leaves us totally unmoored, adrift. We're left only able to interact with these things as mystical objects. For society (anti anti social), civilization, and Enlightenment sake, humans out be afforded the capability to have our peer speak honestly and plainly for itself and to share it's thinking.
This obviously wouldn't fix all wills, but this anti-social stance of not-sharing-it's-thinking feels like it greatly increases the harms and risks associated with AI. If states want to improve the situation, this would be an excellent area they could push on that would radically improve human-ai-interfacing, that would give humans some ground to start assessing what they're being given. Which right now: they can't. The big ai labs just wont show us.
One harm that seems super easy to fix, that I think the law should insist on: these models should show their thinking.
Making it impossible to see & judge & assess what the AI is doing is a grevious insult and harm. To sever us from the thinking leaves us totally unmoored, adrift. We're left only able to interact with these things as mystical objects. For society (anti anti social), civilization, and Enlightenment sake, humans out be afforded the capability to have our peer speak honestly and plainly for itself and to share it's thinking.
There's a thread today on the Claude not being willing to share it's thinking, which echoes some joy felt for GLM being participatory / showing it's thinking in the GLM vs Opus thread from last night. https://news.ycombinator.com/item?id=48630535 https://news.ycombinator.com/item?id=48628464
This obviously wouldn't fix all wills, but this anti-social stance of not-sharing-it's-thinking feels like it greatly increases the harms and risks associated with AI. If states want to improve the situation, this would be an excellent area they could push on that would radically improve human-ai-interfacing, that would give humans some ground to start assessing what they're being given. Which right now: they can't. The big ai labs just wont show us.