Explain Like I'm Five

It's all about smart averages. If you ask for a politician's wife riding a six seater bike with their face covered in bike grease from one that doesn't ban actual people, it'll generate the pieces and do it's best to combine them in a way that makes sense to it. It'll mash together pictures of multi people bikes, possibly extending it to make it longer, it'll mash together greasy faces with the politicians wife's face, and likely have them riding the bike.

Now, the bike may have six seats, but not enough pedals, or even seat stems, because it might not know that's important. It all depends on the days it was trained with.

If you are five and asking to see a gang bang with politicians' wives, I'm telling your mom.

But seriously, if you want to understand things, here's a boil down:
There are many types of "AI." The two you, as a non-researcher, are likely to come into contact with are LLMs (Large Language Models) and genAI. (generative AI)

LLMs work like autocomplete. In a similar way you can type 'C-H-O-C' into your phone and it can predict you probably want to type 'chocolate' or 'chocks' because that sequence of letters usually leads into one of those words, an LLM can take the prompt 'how to I center a div?' and predict the next piece of text will begin with 'The' because the pattern is similar to things seen on many websites that were in the data used to create the model. The system does not think or understand, just predicts the next piece of text.

The other type of AI, which you reference indirectly, is genAI. GenAI is a system that has been trained on many large data sets where there is a picture and a description of that picture so when shown a picture of a dog, it can return a probability the shown image is a dog. When the math is run backwards, it can turn the word 'dog' into a potential picture of a dog, and repeatedly tweak it to create something that more and more reads to the system as 'probably a dog.'

As to what you can do with these systems, that depends. Every implementation of the idea has different rules and limitations. Some you might be able to put that prompt into. Some sites filter your prompts. The general rule at this point though is that it's kind of pointless. You aren't likely to get anything amazing. It's all going to just be mediocre at best unless you have the resources to burn on making something bespoke, but at that point you could probably spend less money finding a bunch of look-alikes for those women and hiring them to be in a porno like people did with that Sarah Palin porno in the days before AI was talked about so much.

It's all probability. what is the probability that this word comes after that word?