Do not panic
Those who read or watch a lot of science fiction may be shocked by all the recent reports about AI: artificial intelligence that can play chess, that makes art, understands text, can write by itself…. Let me reassure you: for the time being, AI cannot yet take over the world. My own experience with AI generated images makes it clear that AI does not give perfect results. The various text-to-image algorithms available remind me of a three-year-old child. It understands words, but the connection between the words quickly becomes diluted. If I ask for a black circle and a blue square, the chance of getting a blue circle is just as big as the chance of getting two black squares against a blue background. These AIs can't count either!
The special thing about this three-year-old child is that it has an incredible artistic talent: it has been trained with nothing but art (both painted and photographic) and will almost always return something atmospheric. In addition, it seems to have ten hands, which do not know of each other what they are doing. If I ask for a wolf and a deer by a stream, I will get one, two or three animals by a stream, but they will be one animal at the front and the other at the back - in an atmospheric way. Hand A had started with the head of animal X and hand E started with the tail of animal Y and on the way they met each other and they neatly drew the parts together. Note: usually the stream and an animal are not drawn together - usually! So there is an understanding in the trained neural networks about categories of things: what has legs and what doesn't. (You can click on the image to see an enlarged version of four examples.)
This ignorance leads to unintentional creativity, a bit like evolution is 'creative': the algorithm simply tries everything, and what works is kept by the client. You can therefore add the number of failed predecessors to every 'successful' image published in the media.
Another source of creativity - and this time intended - is the noise given to the algorithm at the start of image-generation. If no starting image is given, then there is only noise; if a starting image is provided, 50% noise is often added to it. With each repetition of the same assignment, the noise causes a different outcome, and with that 'creativity'.
You can also deliberately look for unexpected results, for example by entering a nonsense text. What I often do myself: I take a pop song as a starting point, usually 'title + artist', and then see what the AI makes of it. Sometimes the generated image fits the song completely, and sometimes not at all. Here is an example based on Iron Butterfly's 'In a gadda da vida'. The title of the song is drunken talk, so I only expected to see an iron butterfly, but that turned out to be too ambitious. Butterflies are just butterflies. None of the four are really existing species, but the one at the top right takes the crown in terms of creativity. The AI apparently did its best to understand and process the rest of the assignment as well.
The aforementioned mistakes of our three-year-old toddler often enough give unwanted creativity: mammals with five or three legs, people with three arms or only one, wrinkled faces or heads that fall out of frame. Undesired creativity in the example with the butterflies: extra borders in the image, as if the AI also thinks the frames are part of the image and wants to co-generate them. In addition, the butterfly at the top right has also fallen outside the picture at the top and bottom edge. Still, this is the prettiest of the four, so what does the sincere artist do: pick this one out, add missing parts, and improve on the white stripes. You can see the result in the Art section of this website.