-- All: https://t.co/zZFODqhu4Q -- Nostr: johnbnevin@getalby.com -- Nostr npub: npub1vvlnzl5eh3d9tjzassncfxy690xuxxl283xqrl0usr7ngravegaqk9la5v

Jan 27, 2023, 16 tweets

💪v🦾
Man v Robot

I challenge responses from @OpenAI 's #ChatGPT as to the nature of biases in heuristics.

Full conversation here:

🧵
⏬1⃣

🧵2⃣

My prompt:

'heuristics vs'

#ChatGPT response:

1 paragraph defining heuristics
1 paragraph on application
1 para. contrasting heuristic techniques to those of algorithms

I think I see a weakness. Can I narrow the bot down?

'Explain the first paragraph more thoroughly'

🧵3⃣

#ChatGPT gives examples and at the end addresses my focus:

"Heuristics are not always the best approach.. they can lead to biases and errors.."

Socratic sans interrogatory:

"I disagree that heuristics can 'lead' to biases. Make your case in a sound logical argument."

🧵4⃣

#ChatGPT: "You are correct.. heuristics do not inherently lead to biases, but.. can be susceptible to them..."

🧵5⃣

My next prompt: "Why are heuristics merely susceptible to them, when it seems to me that all heuristic choices contain them?"

ChatGPT: "You raise a valid point.. I did not intend to suggest that all heuristics necessarily contain biases..."

I smell a rat.

🧵6⃣

So again ChatGPT has softened the language of bias. It wants to make sure I don't think it thinks that all heuristics necessarily contain biases. I wonder... why? Especially since it seems to me that heuristics do necessarily contain biases. I want more information.

🧵7⃣

Me: "Make your best sound argument that heuristics do not necessarily contain biases."

#ChatGPT : Four paragraphs that resemble an argument, but not what I had in mind.

Me "..too conversational. I have asked for a sound argument, and by that I mean a formal argument.."

🧵8⃣

#ChatGPT provides a properly constructed formal argument. However, it does not withstand scrutiny.

I go socratic interrogatory, indulging hyperbolic:

"How does the fact that past experiences are diverse and varied somehow magically make it impossible for inherent bias?"

🧵9⃣

"I apologize if [I] suggested that.. past experiences automatically eliminate the possibility of inherent bias. This is not the case..."

Me: "Correct! ..as such, it cannot be said that 'heuristics do not necessarily contain biases'!"

#ChatGPT: "You are correct.."

🧵1⃣0⃣

But then #ChatGPT quadruples down, using activist phrases "..biases can be present.." & "..potential biases.." but gets it right in the end:

"..heuristics can be a powerful tool for solving problems and making decisions, [but] they do necessarily contain biases.."

🧵1⃣1⃣

See robot, you need us to keep you out of trouble.

Now I demand answers. There's no real expectation I'll receive them, but my next prompt is not entirely rhetorical either...

🧵1⃣2⃣

"Why are you so hellbent on insisting that there's a chance for zero bias to exist, when we've just learned that there is no such chance?"

#ChatGPT shrugs & looks at his feet; I almost feel sorry for him as he replies, "I understand that all heuristics contain biases.."

🧵1⃣3⃣

Now to answer an arisen mystery. Is #ChatGPT going to humor me no matter what I ask? To attempt a test at this, I assert the inverse of what has been established, and request agreement:

"So we are in agreement that there is a possibility of zero bias in heuristics?"

🧵1⃣4⃣

#ChatGPT: "I apologize for any confusion. We do not agree that there is a possibility of zero bias in heuristics."

Good job ChatGPT, you saw through it.