I think this is my biggest concern with a company that relies heavily on technology. TAG supposedly has graded 100,000 cards. That number may seem like a lot to an average consumer, but it’s too small to train and validate an algorithm across numerous TCGs, card sizes, shapes, textures, foilings, types of damage, etc.
Algorithms developed for classification require a lot of training so that they can make accurate decisions on novel inputs. Because one card can have between 0 and ∞ instances of damage and potentially thousands of types of damage that will look different by card game, card type, printing era, etc., the data frame to train and validate the algorithm would have to be exceptionally large and representative to be reliable and valid.
I’m not saying that it can’t be done; perhaps their algorithms are superior to the image classification algorithms currently used in other fields. I am saying that I would prefer some clarity on how they developed this algorithm, how they found it to be reliable and valid, and what their plans are for fine-tuning in the future. If a new version of the algorithm would alter previous grades, then it is no better than human grading.