I am using Delphi to develop similar technologies in the area of odour measurements. Very often we find, that we are able to discriminate between different qualities of products or genuine and tampered products. However, statistics can not tell you the reason for the differences.

For that we need to correlate to reference information (e.g. human odour impression). The correlation is achieved by mathematical correlation of information on reference material (e.g. from human test sniffers) with the output of measurement systems. The main obstacle in using such methodology is the need of that reference information. Due to the vast amount of variance in real world objects, a very large amount of correctly classified reference material is necessary. Often this kind of reference material is not available (or extremely expensive to collect). Due to the variance we have to limit the systems to specific applications (say detect a certain impurity, general maturity of a certain food).

In the case of identifying faked masterpieces, the amount of reference information is probably limited. The algorithms may be able to identify differences within a picture or between different pictures, but you still need an expert to make use of this information.

Maybe one could develop a system that integrates the experts knowledge, but that would be much more expensive than to just use an expert.