The way the algorithm works is by gathering large samples of data and then calculating the probability of a match. So in the example below (taken from tensorflow.org/tutorials/image_recognition) you can see how the mite has a small chance of being a starfish.
From the youtube video, you can find frames where there's some funny mis-identifications, such as the guy's hand as a dog.
My SO ran some of my comic panels through YOLO with the threshold lowered to 2% (which means even if there's only 2% of probability, it'll still be identified as such) and we got some funny results.
This potential for mis-identification will be a plot point for the next chapter of the story. Stay tuned! :D