Training for Chat AI

In order to access its intelligence sometimes you have to talk to it as if it were an idiot. Or perhaps an idiot savant. It’s a bit like the character played by Dustin Hoffman in Rain Man.

Anyone can use ChatGPT. Just go to chat.openai.com and you can access the most recent version, GPT-4o for free (there are some premium features available for an extra $20/month).

When you go to ChatGPT you find essentially a blank screen and the implicit question “How can I help you today?”

You can ask it questions. Better still, you can upload a large PDF and ask questions of the file. You can also upload images, which it can describe, or a scanned page: it can recognize the text, even if handwritten.

What has emerged as a gating issue for the successful use of Chat AI is learning how to ‘speak’ with it (which you can literally do in the latest mobile app version). In Chat AI-ese is called creating “prompts” or “prompting.” Users have discovered that the more precise and detailed their prompts are, the better the responses they receive from ChatGPT. Further, prompts are not just one-off. Chat AI can continue a conversation for quite a while (though not indefinitely), and if you don’t get the answer you’re looking for you can revise and refine your prompts. This takes a lot of getting used to, and has spawned a series of how-tos, written and online, to train users on how to get the most out of prompting.

Try asking any of the Chat AI software to explain a concept like developmental editing. Then ask it to craft an explanation that a 12-year-old could understand. The results are dramatically different. Amusingly, Chat AI seems also to respond to emotional pleas. Adding “this is very important to my career” to a prompt can coax more useful responses.

(This is as good a time as any to add a parenthetical: It’s both reassuring and deeply troubling that the top scientists working on language-based AI are unable to explain why things like this occur.)

Of the various co-opted terminology seized upon during the AI gold-rush, “prompts” and “prompting” are as good as it gets. You don’t really “talk” to a Chat AI. You don’t “ask” it questions (or, you shouldn’t—that’s not how it works). You prompt these engines for responses, not for simple answers.

At first I thought an analogy might be theatrical prompts. The prompter gooses an actor who loses track of a line, causing them to jump back into action. But this isn’t accurate because an AI prompt doesn’t loosen a chunk of memorized text—that’s not how generative AI works. Give an AI engine a single prompt repeatedly, and the answers will always be at least a little bit different.

Just now I asked ChatGPT to “list the key elements of effective prompting.” I asked a second time and the answers were thematically consistent, but quite different in the specifics. Topics that were highlighted included:

• Be specific

• Avoid ambiguity

• Provide background

• Use clear instructions

In a sense it’s quite odd: AI has a good grasp of language—shouldn’t it have a sense of what you mean without pedantic prompting? But it’s precisely that mastery of language that makes prompting most powerful. Tell it exactly what you mean, not approximately.

Talk to it like the helpful assistant, the learned counselor, that you want it to be, not like a kid off the street.

The takeaway here is that, at least for the time being, the effective use of LLM-based AI tools requires at least conversancy with prompting best practices. A prompting mastery will be appropriately rewarded.