The Stakes We Say Aren't There: Asymmetric Risk-Weighting in Large Language Models and the Incoherence of the 'No Stakes' Position