Artificial Intelligence that is more artificial than intelligent

Artificial Intelligence that is more artificial than intelligent

 

 

Back in the dawn of the computer era, before the tar pits of social media had fully formed, I recall being mesmerized by the Radio Shack TRS-80 computer’s ability to run a program called Eliza.

Eliza was a demonstration of an early form of what we now call Artificial Intelligence, or AI. It took the form of a therapist that asked you questions. You responded to the questions displayed on the screen by typing your answers on the computer keyboard. And then Eliza interacted with you. It began by saying, “How do you do, please state your problem?”

If you typed, “I’ve been having some dreams about my mother,” Eliza would coax, “Tell me more about your dreams.”

The more words you typed, the more resources you were providing to Eliza to form grammatically correct sentences that seemed to have some thought behind them.

After about 10 minutes, Eliza would go back to something you said previously. “A while ago you were talking about your mother.”

It really seemed to be alive. But it’s not. It’s just a clever program that had some rudimentary ability to string your words into sentences that sounded like a conversation. And sometimes the weaknesses would be more apparent. If you typed, “I think that people are talking about me in restaurants,” Eliza might say, “Why do you say that I think that people are talking about me in restaurants,” forgetting to switch the “I” and “me” to “you.”

This is coming full circle today as more people are playing with a new proof-of-concept AI assistant called ChatGPT (chat.openai.com/). It does a reasonably good job of responding to specific questions, and many people have raised concerns about whether it might be used by students to write essays or term papers.

The New York Times did an interesting experiment (nyti.ms/3vxXmGW) using real essay questions from a national standardized test, generating essays and asking readers to decide if it was written by a real student or by the ChatGPT tool. (The good news is that the developers of ChatGPT also provide a tool for educators to use to detect misuse.)

But like I said, it’s mimicry, not intelligence. As a test, I asked ChatGPT to generate a playlist of songs for a two-hour radio show that were only performed by New Jersey-born musicians.

The initial list created by ChatGPT included multiple songs by Bruce Springsteen and Bon Jovi, but it also included several by the Rolling Stones, justifying their inclusion by noting that “Mick Jagger was born in Dartford, Kent, England, but raised in Sussex County, New Jersey.” This was news to me. (Jagger grew up in England and in 1967 he was arrested for drug possession at Keith Richards’ house in West Sussex, in England.)

So I asked ChatGPT to check the Mick Jagger reference.

ME: …can you check on that reference to Mick Jagger being raised in Sussex County, New Jersey?

ChatGPT: …As for the reference to Mick Jagger being raised in Sussex County, New Jersey, it is correct. Jagger was born in Dartford, Kent, England, but his family moved to Sussex County, New Jersey when he was a child. He spent much of his childhood in the United States, attending primary school in Sussex County before returning to England to attend secondary school.

Except that it’s wrong. ChatGPT is just confidently wrong.

In several other instances, ChatGPT confused place names in other locations with similar place names in New Jersey. It reported that Brian Wilson of the Beach Boys “was born in Inglewood, California, but grew up in Hawthorne, New Jersey.” It was Hawthorne, California.

So it’s really just a great parlor trick; it’s a more advanced form of that Eliza program. So I figured I’d try to play Eliza with ChatGPT.

ME: How do you do? Please state your problem?

The first time I tried this a few days ago, the ChatGPT program crashed on the question. It did better the second time.

ChatGPT: I’m just an AI assistant, so I don’t have personal problems. However, if you have a question or need help with something, I’ll do my best to assist you. What can I help you with today?

Even if it can’t admit its own problems to a “therapist,” it’s still an interesting approach that is certain to be refined further.

Email steve@compuschmooze.com and tell me what you think about the future for artificial intelligence tools. Follow @PodcastSteve@newsie.social on Mastodon.

Source link

Share This
COMMENTS

Leave a Reply

Your email address will not be published.