The neural net objective doesn't add or subtract points. Instead, the "show" option uses red, white, and blue to indicate how well the current protein matches its AlphaFold prediction. Areas shown in red are far from where they were predicted to be. Areas shown in blue are close to the prediction. Areas shown in white are somewhere in between.
Mutating the areas highlighted in red or white may improve AlphaFold confidence in a new prediction. Changing the blue area is less likely to help.
The protein's current amino acid sequence (primary structure) must have already been processed by AlphaFold for the coloring to work. The neural net option shows "no data" if the protein hasn't been submitted to AlphaFold, and the "show" option will show the entire protein in white.
When an AlphaFold prediction is loaded, the neural net "show" option may color the protein mainly blue. Some areas may appear in white or even red, however. The coloring relies on a two-dimensional prediction, which is then translated into a three-dimensional model that be be loaded in Foldit. The final position in three-dimensions can be a bit different from the two-dimension prediction.
The Foldit blog post The Neural Net Objective describes how things work in more detail.