As mentioned at AutoML's preview back in May, Google is actually using "baby" neural networks to build these systems. It iterates the mini nets with reinforcement training and picks the best one from the bunch.
Not surprisingly, this costs money: you have to apply for access, and you'll be charged fees for both training the models and accessing them. You won't be using this to indulge in a hobby. However, this promises to make AI, and image recognition in particular, much more accessible. While there are already custom AI options (Microsoft can fine-tune its trained AI models for you), Google's approach is simple and hands-on enough that your favorite website or device maker could roll AI into their products with relatively little effort.
There are already some practical examples. Disney, for instance, is using Cloud AutoML to help you search for products on its store based on what they look like, not arbitrary tags. You can find that Buzz Lightyear toy even if it's been miscategorized. Conservationists at the Zoological Society of London, meanwhile, are hoping to automatically categorize animals that pass by wildlife cameras. While there will still be a need for advanced, manually programmed AI, it won't be as essential as it used to be.