A neuronal network “knows” nothing at first. It has to undergo a training to “learn” bit by bit and make its outputs more precise. For the training, it uses training data. In this case, the artificial intelligence was trained with about 70.000 records telling it the correct solution. The results become more reliable as it “learns” to adjust its network parameters. Training data is extremely valuable. Fortunately, open databases exist such as the “MNIST”-database that was used for this exhibit. You can see clearly, that the number ones were written in english-style handwriting (“I”). The “1”, how it is commonly written in Germany, is therefore often interpreted as a “7”.