The Universe of Discourse

Mon, 22 Apr 2024

Talking Dog > Stochastic Parrot

I've recently needed to explain to nontechnical people, such as my chiropractor, why the recent ⸢AI⸣ hype is mostly hype and not actual intelligence. I think I've found the magic phrase that communicates the most understanding in the fewest words: talking dog.

These systems are like a talking dog. It's amazing that anyone could train a dog to talk, and even more amazing that it can talk so well. But you mustn't believe anything it says about chiropractics, because it's just a dog and it doesn't know anything about medicine, or anatomy, or anything else.

For example, the lawyers in Mata v. Avianca got in a lot of trouble when they took ChatGPT's legal analysis, including its citations to fictitious precendents, and submitted them to the court.

“Is Varghese a real case,” he typed, according to a copy of the exchange that he submitted to the judge.

“Yes,” the chatbot replied, offering a citation and adding that it “is a real case.”

Mr. Schwartz dug deeper.

“What is your source,” he wrote, according to the filing.

“I apologize for the confusion earlier,” ChatGPT responded, offering a legal citation.

“Are the other cases you provided fake,” Mr. Schwartz asked.

ChatGPT responded, “No, the other cases I provided are real and can be found in reputable legal databases.”

It might have saved this guy some suffering if someone had explained to him that he was talking to a dog.

The phrase “stochastic parrot” has been offered in the past. This is completely useless, not least because of the ostentatious word “stochastic”. I'm not averse to using obscure words, but as far as I can tell there's never any reason to prefer “stochastic” to “random”.

I do kinda wonder: is there a topic on which GPT can be trusted, a non-canine analog of butthole sniffing?


I did not make up the talking dog idea myself; I got it from someone else. I don't remember who.

Addendum 20240517

Other people with the same idea:

if your dog could summarize news articles accurately 30% of the time you'd be fucking amazed that it could do that, but you still wouldn't trust it

and other phrasings 1 2.

[Other articles in category /tech/gpt] permanent link