What if AI safety didn't rely on censoring the AI's "thinking" process?