Surprising results from a test of the ideological biases of three leading AI models.