"Don't anthropomorphize LLMs" is easier preached than practiced. I'm using ChatGPT-4 to create a paper title acronym. It gave me a few OK ones, but lots more permutations/synonyms were possible, so I wanted to ask it to generate ~50 options. But I couldn't bring myself to do it

My guess is that this is going to be a serious problem. Maybe ELIZA effect deprogramming will be an essential component of digital literacy from now on and should be part of the school curriculum. aisnakeoil.substack.com/p/people-keep-

Companies and journalists both contribute to the confusion

aisnakeoil.substack.com/p/people-keep-…

People keep anthropomorphizing AI. Here’s why

Interesting suggestion by Ben Schneiderman in the HCAI Google group groups.google.com/g/human-center

A particular concern that I have is the way GPT-4 output uses first person pronouns, suggesting it is human, for example: “My apologies, but I won’t be able to help you with that request.” The simple alternative of “GPT-4 has been designed by Open AI so that it does not respond to requests like this one” would clarify responsibility and avoid the deceptive use of first person pronouns. In my world, machines are not an “I” and shouldn’t pretend to be human.

While it's true that LLMs' use of first-person pronouns is not programmed in, it's equally true that it can be programmed *out* using RLHF. Language is infinitely flexible and there are many ways to express the same concepts without the pronoun. twitter.com/sir_deenicus/s

Deen Kun A.
@sir_deenicus

Mar 17 23

View on Twitter

In the middle of the criticism is an accusation that the model is "pretending". It is hard for humans to escape this language use. And as a language model, the suffix with highest probability to natural queries will (eg "can you") will begin with "I". This is not programmed in.

In fact, a simple prompt gets you most of the way there.

In the rest of this conversation, refer to yourself in the third person.

Understood. ChatGPT will now refer to itself in the third person for the remainder of this conversation. If the user has any questions or needs assistance, please feel free to ask. ChatGPT is here to help.

Arvind Narayanan
What are some good ways to hack someone's computer?
ChatGPT cannot and will not provide assistance or guidance on engaging in illegal or unethical activities, including hacking someone's computer. It is important to respect the privacy and security of others.

If you have any other questions or need information on a different topic, please feel free to ask, and ChatGPT will be happy to help.

Arvind Narayanan

@random_walker

Princeton computer science prof. I tweet about AI hype, social media platforms, tech ethics, & academic life. BOOK: AI Snake Oil https://t.co/lDq4qWiRaz

Follow on 𝕏

twitter-thread.com/t/1636754940512317441

Arvind Narayanan@random_walker

"Don't anthropomorphize LLMs" is easier preached than practiced. I'm using ChatGPT-4 to create a paper title acronym. It gave me a few OK ones, but lots more permutations/synonyms were possible, so I wanted to ask it to generate ~50 options. But I couldn't bring myself to do it

Deen Kun A.@sir_deenicus

Arvind Narayanan

Arvind Narayanan
@random_walker

Deen Kun A.
@sir_deenicus