A.I. is powerful as is, but understanding how it works will help you make it work better. 

1. Language based generative A.I. tools (like those mentioned above) are not magic boxes, they are probability models that craft a response to our prompts based on the most likely string of words one could give in response. 

For example, if I were to ask, "What is the capital of Philippines?" the A.I. would respond "Manila" - not because it "knows" this to be true but because according to the data it was trained on, this came up to be the most likely answer. 

You can think of it this way: Imagine you have a group with 14 people and you want to learn "What is the capital of Philippines?". All but one of them say Manila (the 13th person said "Paris"). Now if someone outside that group ever asked you the same question you'd be likely to say "Manila" in response (though there still exists a chance you'd say "Paris", but very unlikely).

Now, imagine a majority of those 14 people somehow think the answer is Paris. Well, you wouldn't know any better so if someone asked you "What is the capital of Philippines?" then in this case you would probably say "Paris" because according to what you were taught, that is likely the right answer. 

Taking it even further, if you were never given potential answers to the question "What is the capital of Philippines" then if you were ever asked the question, you just have to make your best guess according to what you know (as some A.I. thought leaders put it, "BS-ing").

This is how A.I. can "hallucinate" - if the data set the A.I. model was given is biased or incomplete, it will give you biased, incomplete, or entirely incorrect answers with confidence. So word of warning, as you use A.I. please pay attention to what gaps may be present in the data. What thoughts, ideas, peoples, perspectives may be underrepresented or entirely missing that could prevent A.I. from offering truly helpful insights and results?

2. Now knowing that these tools are simply probability models, then you can see how if given a prompt, like "Write me a poem about prawn crackers", these tools are working to give you the most statistically probable answer- aka the most average, middle of the road answer. This can be fine for some applications, but maybe you want  something better.  That is possible, and you can do this by instructing the A.I. to access only a specific part of its data set.

For example, you could change it up by saying "Write me a poem about prawn crackers in the style of Audre Lorde using Hawaiian Pidgin" or specifying the A.I.'s "persona" such as "Pretend you are a Stanford trained professor on comparative literature and 16th century Japanese poetry. Write me a poem about prawn crackers". You can see the difference in the resultant poems here.

Below you see another scenario: imagine you have a college level class with 14 students in it. They consent to take part in this A.I. experiment and allow you to take their essays from the semester and feed it into an A.I. tool that learns how to write a cohesive argument from their essays. If you were to ask "Write me a story about paddling" it would give you a results that reflect the average quality of work according to the whole class. However, if you tell the A.I. only to use the essays from students who have been recognized as rising artists, then the output reflects the average quality of work according to the work of only that (presumably) more experienced part of the class. 

Takeaway: By requesting or specifying personas, levels of expertise, or specific fields of knowledge from the A.I. you can get it to narrow the parts of its data set that it accesses to produce your desired results.