3 ms·

It's not just people purposefully injecting bias into a model, it's about the biases that get baked into a model completely by accident. If there is a lot of m

by bena 4y ago

It's not just people purposefully injecting bias into a model, it's about the biases that get baked into a model completely by accident. If there is a lot of material written about how short people are horrible, ChatGPT will hate short people. Without me making an explicit decision to make ChatGPT hate short people. And that's a whole side of the AI conversation very few people are actually having. Are we feeding these neural nets bad models? Who has actually vetted the data we're using to train?

foofoo4u 4y ago
This is an aspect I'd like to get a better understanding of. ChatGPT is excellent at writing out claims. I don't know how ChatGPT is implemented, but I am going to assume these are not its claims. But rather, these are claims that ChatGPT found from publications injected into its training models. If this is the case, then your example should be correct: if there is a lot of negative material about short people, then ChatGPT will hate short people. This is "fine", insofar as to the fact that we humans too can make such conclusions. But what I really want to know is can ChatGPT explain its reasoning. Why do you [ChatGPT] hate short people? What makes short people so terrible? Are taller people superior? If so, why? If you were short, would you still hold the same views? Is there a possibility that your position is wrong? What would it take to change your mind? Where do you draw the line to make the distinction between who is "short" and who is "tall"? What if someone was born short, but went through surgery to make themselves taller? And vice versa. This is what I would like to see from ChatGPT. I am going to assume ChatGPT cannot accomplish this. Yes, it can probably write explanations to each of these questions, but I will assume the responses will be inconsistent to its original claim. Why? Because as long as the claims are not derived from a chain of reasoning, then what ChatGPT is is a glorified — but rather impressive — pattern recognizer, constructor, and conveyor. Someone please correct me if my assumptions are wrong.

by bena 4y ago

foofoo4u 4y ago

This is an aspect I'd like to get a better understanding of. ChatGPT is excellent at writing out claims. I don't know how ChatGPT is implemented, but I am going to assume these are not its claims. But rather, these are claims that ChatGPT found from publications injected into its training models. If this is the case, then your example should be correct: if there is a lot of negative material about short people, then ChatGPT will hate short people. This is "fine", insofar as to the fact that we humans too can make such conclusions. But what I really want to know is can ChatGPT explain its reasoning. Why do you [ChatGPT] hate short people? What makes short people so terrible? Are taller people superior? If so, why? If you were short, would you still hold the same views? Is there a possibility that your position is wrong? What would it take to change your mind? Where do you draw the line to make the distinction between who is "short" and who is "tall"? What if someone was born short, but went through surgery to make themselves taller? And vice versa. This is what I would like to see from ChatGPT. I am going to assume ChatGPT cannot accomplish this. Yes, it can probably write explanations to each of these questions, but I will assume the responses will be inconsistent to its original claim. Why? Because as long as the claims are not derived from a chain of reasoning, then what ChatGPT is is a glorified — but rather impressive — pattern recognizer, constructor, and conveyor. Someone please correct me if my assumptions are wrong.