What makes you think computers are less likely to be biased? It's far easier for an algorithm to "learn" an association between probability of guilt and the colour of pixels in the moving bit of the video screen or dialect words used in the defendant's transcript than it is for the algorithm to actually evaluate the testimony itself.