ELIZA

And one of the things we, we learned making the series was that the first chatbot was created in 1966, I believe, at MIT.

It was called ELIZA, and it was, it was like very bad.

You know, you would sit and talk to it, and it would just mirror what you were saying back at you.

And the guy who created it, this professor there, Joseph Weizenbaum, the thing he found most unnerving was that he would sit people in front of ELIZA and have them talk to it, and he talks about how his secretary at one point asked him to leave the room, and he asked her why, and she said, "I wanna have a conversation with it." And he just like he, he ends up writing later, he couldn't believe how this, again, quite crap chatbot-

[laughs]

... could, I think his words were, "evoke powerful delusions in otherwise normal people".

It, it's a cr- like... [lips smack] The idea that we would offload our ability to, like, empathize with each other, to think about what this person needs to hear in this moment, to, like, lose the ability to do that, and in fact to get instructions on how to do that from, like, from a machine, I don't know.

I think y- you're kinda giving up something pretty essential.

They're always bugging us about something or other.

Can you think of a specific example? Eliza was programmed around 1964, on an amount of memory that couldn't run Mega Man 2.

Well, my boyfriend made me come here.

Your boyfriend made you come here.

You can read her entire code base in less time than it will take to listen to this episode, even if you scrub through the ads.

Yet even today, Eliza is very impressive.

He says I'm depressed much of the time.

I'm sorry to hear that you're depressed.

So he got interested in conversational programming, and the rest is very short history.

In only a year or two at MIT, he built the original ELIZA, the one we found.

ELIZA was the name Weizenbaum gave to the very first AI chatbot.

Now ELISA, as a basic program, gave the computer the ability to receive written sentences and to respond in turn.

But in order to function, ELISA needed a script to follow.

You don't think the psychiatrist doesn't know about sailing boats, right? You know that the psychiatrist wants you to tell him or her what you think about sailing boats.

It's pretty clever to give the computer the role of a therapist, constantly searching for a better understanding and only ever asking questions.

What's more human of an interaction than a conversation between a patient and a therapist? But by the time Weissenbaum has built ELIZA, he's still just testing the waters, trying to understand the human-computer relationship.

Then, in 1959, the term machine learning was coined when talking about a computer that learned how to play a mean game of checkers.

But then, in 1966, we got Eliza, a natural language conversation program created by Joseph Weizenbaum, the world's first chatbot.

She was an incredibly rudimentary form of AI compared to what we're used to now.

The user would input some line of dialogue, then Eliza would scan the input for key words.

From those key words, she would then pick out a response through pattern matching.

However, despite her being a pretty unintelligent form of AI, Weizenbaum was shocked at how emotionally attached people got to the thing.

This became known as the Eliza effect, the tendency for...The humans have to falsely give human attributes to AI systems.

And as these computer models have gotten better, this effect has only gotten worse.

And there are a few different touchstones in the history of these kind of chatbot interfaces.

You've got ELIZA, which we've talked about before.

You've got Eugene Goostman.

You've got Siri and Alexa and the kind of voice-

... assistant technology that came around, like, early 2010s.

Mm.

And then you've got the genesis of the large language models, which are based on this architecture called the transformer, and that's very much the path that we've trodden in the past, is go, "Oh, ELIZA and Eugene Goostman," and into large language models.

But the term artificial intelligence, it turns out, is also really overloaded, and it means many, many different things in many different contexts, which does mean that it can be confusing to use the word AI or artificial intelligence without understanding the context.

So it's hard to tell, and I think, you know, there are some business reasons why that, that's the case, right? I think these companies would really like us to assume, to think that these tools are sentient.

Um, they pick up on something that's called the Eliza effect, which is this, uh, is this, you know, a- an effect that was named after a chatbot actually back from the '70s, um, called Eliza.

Uh, no fancy machine learning, nothing like that, um, but it was really enthralling.

People would play with it, and they would think they were having these kind of, you know-

Mm

... in-depth conversations with it.

So the Eliza effect is when people, uh, project sentience, complexity, you know, emotions onto a chatbot.

Right.

But yes, I think that's a very natural tendency for us.

And there is, in fact, a cognitive bias called the Eliza effect that's been known since the 1960s.

That was when one of the, if not the first chatbot was invented.

And users interacted with it.