A study confirms that no major language model is neutral and that its biases tend to be constant in the same direction