That's the problem. How can a logic system process non-logic data? This is already demonstrated with pictures and especially identifying objects inside of a picture. These constructs (images, movies, characters, emotions, sensations, feelings, etc.) are completely alien to transistor-based systems.
For example, to a transistor based system, the letter Z is known as 1011010 (0x5A) in the ASCII file format and if it is to be displayed, it has to look up the requested font (e.g. Courier New), find the TrueType Font file (e.g. cour.ttf), load the instructions on how to create the symbol, then send it through the display pipeline. The reverse is terribly inefficient and unreliable (as seen by OCR software) because it has to interpret symbols (often without the font to provide context) back into that binary it knows and binary is very inflexible. It is either right or it is wrong.
Just like a computer has extreme difficulty defining shapes in a picture, it has extreme difficulty picking up subtle expressions. Facial recognition, for example, can get a false negative simply because the individual being photographed is laughing at a joke. Following that line of thought, most people can pick up on whether a laugh is genuine or faked because a genuine laugh uses a lot more muscles than a fake laugh. It's literally a Pandora's Box of computing problems.
Not to mention, the human element of the problem. The Japanese made a robot that was an attempt to make it look human. People were revolted by it because they knew it was off just by looking at it. The human brain is incredibly skilled at identifying other people because we are a social species. It discovers a fraud just as easily as it discovers someone real. This phenomena is what leads people to pick out "human faces" from the mundane like toasted bread or the moon. There's a lot of easy ways to fool the brain (like special awareness) but this is not one of them because it was born out of survival instincts.