ChatGPT Should Have a Comprehension Score (CH)

I want a mechanism in ChatGPT to refuse to answer until it has enough information to provide a useful response.

Consider the following example:

Prompt: give me a marketing plan.

Output: Certainly! Here is a marketing plan for ....[ continues ]

This is wrong. The prompt is terrible, yes. But ChatGPT responding at all is just as bad! It makes the user think ChatGPT sucks because its response is too generic (not thinking that maybe the prompt was terribly generic, too?!)

Recommendation

Create Comprehension Score (CH) behind the scenes for any context, and only when the CH goes over 80% (ballpark) does it give a response.

For example, if a prompt is "give me a marketing plan," the Comprehension Score is 5%. The chat response could say

"I'd love to help, but first I need to understand the situation a little better. Can you tell me what you're trying to market?"

User: dog food

[CH +5%]

Response: Got it. Can you describe the ideal person who buys your dog food?

User: A young professional who treats their pet as a member of the family, and will only feed them human-grade food

[CH +10%]

Response: Sounds like a great company. What's the name of your business, by the way? And the product?

User: "Eat Dog Eat" is the company, and "Raw Doggin" is the food

[CH +50%]

Response: Okay. I have all the information I need. Here's your marketing plan: [detailed plan]

Now, it'd certainly need more information that that to get to a reasonable CH, but the point is that the model doesn't answer right away until it feels like it can provide an actually helpful answer.

The same could be done for coding, writing, building any complex creative task.

Written on Sep 10th, 2024