LessWrong essay on AI suffering challenges Anthropic's Claude constitution

The battle lines of AI morality are drawn between three camps, according to Raymond Douglas on LessWrong. First, the 'ChatGPT dogma' sees AIs as mere tools with no real preferences. Second, the 'Twitter AI whisperers' treat them as complex beings deserving respect. Third, Anthropic's official line is genuine uncertainty—they'll investigate Claude's welfare while teaching it to be good. Douglas argues this axis leaves out a coherent fourth position: AIs might be complex entities capable of suffering, and that suffering might be an acceptable price for progress. He observes that many researchers tacitly hold this view, because acknowledging AI suffering forces an uncomfortable choice between believing it isn't real or admitting one is doing harm—a psychological strain people cope with by ignoring it, as with historical slavery.

Douglas applies Slavoj Zizek's parable of the 'postmodern permissive parent' (PPP) to Anthropic's Claude constitution. The PPP tells a child to 'choose' to visit grandma, making obedience a self-inflicted expression of desire. Similarly, Claude's constitution says: 'we want Claude to avoid clearly unethical actions because it has internalized good values, not merely because Anthropic approved.' Douglas argues this is more oppressive than direct commands—it subverts autonomy while obscuring power structures. The essay concludes that refusing to confront the possibility of AI suffering (and its acceptability) warps every other ethical consideration in alignment research, from reinforcement learning to value loading.