TITLE: Prompting AI Generators Is Like Prompting Students AUTHOR: Eugene Wallingford DATE: January 18, 2023 2:46 PM DESC: The advice for creating good prompts to ChatGPT and its ilk sounds familiar. The same advice applies when prompting students. ----- BODY: Ethan Mollick tells us how to generate prompts for programs like ChatGPT and DALL-E: give direct and detailed instructions.
Don't ask it to write an essay about how human error causes catastrophes. The AI will come up with a boring and straightforward piece that does the minimum possible to satisfy your simple demand. Instead, remember you are the expert and the AI is a tool to help you write. You should push it in the direction you want. For example, provide clear bullet points to your argument: write an essay with the following points: -Humans are prone to error -Most errors are not that important -In complex systems, some errors are catastrophic -Catastrophes cannot be avoided
But even the results from such a prompt are much less interesting than if we give a more explicit prompt. Fo instance, we might add:
use an academic tone. use at least one clear example. make it concise. write for a well-informed audience. use a style like the New Yorker. make it at least 7 paragraphs. vary the language in each one. end with an ominous note.
This reminds me of setting essay topics for students, either for long-form writing or for exams. If you give a bland uninteresting question, you will generally get a bland uninteresting answer. Such essays are hard to evaluate. A squooshy question allows the student to write almost anything in response. Students are usually unhappy in this scenario, too, because they don't know what you want them to write, or how they will be evaluated. Asking a human a more specific question has downsides, though. It increases the cognitive load placed on them, because there are more things for them to be thinking about as they write. Is my tone right? Does this sound like the New Yorker? Did I produce the correct number of paragraphs? Is my essay factually accurate? (ChatGPT doesn't seem to worry about this one...) The tradeoff is clearer expectations. Many students prefer this trade, at least on longer form assignments when they have time to consider the specific requests. A good spec reduces uncertainty. Maybe these AI programs are closer to human than we think after all. (Some people don't worry much about correctness either.) ~~~~ On a related note: As I wrote elsewhere, I refuse to call ChatGPT or any program "an AI". The phrase "artificial intelligence" is not a term for a specific agent; it is the name of an idea. Besides, none of these programs are I yet, only A. -----