Dieser Chatbot aus den 1960er Jahren war ein Vorläufer der KI. Sein Schöpfer fürchtete sich davor: „Ich bin kein KI-Kritiker. Ich bin ein Kritiker der Gesellschaft.“

> Users opened up, shared intimate details about themselves and treated the program as if it were human. The response was so intense that even Weizenbaum’s secretary at the Massachusetts Institute of Technology (MIT) reportedly asked him to step out so she could speak with the program in private.

–

> ELIZA worked by scanning a user’s typed input for keywords and generating responses that resembled human conversation based on pattern matching. “ELIZA shows, if nothing else, how easy it is to create and maintain the illusion of understanding,” Weizenbaum wrote. He believed that once the program’s mechanisms were revealed, the “magic” would “crumble away.”

People used to be so stupid. We still are, but we used to be, too.

Been trying to tell people that LLMs are just fancy auto-complete, been repeating this ad nauseam, trying to reveal the magic in the simplest way.

Fuckers are still getting married to ChatGPT.

The book about this (Computer Power and Human Reason) is fascinating.

I repeated this in the mid-1970s, at CSUN. I wrote a version of ELIZA – same thing – simulated „therapist“ chatbot. At that time, the main computer system was a timeshared CDC mainframe, going out over serial lines to terminals at several locations on campus, and also servicing all other Cal State campuses at that time.

The software was accessed through the terminals, and I had business majors try it. We did have a chat app so people on various terminals wherever could communicate, so the students didn’t know it might not be a person. The idea of a chat bot was essentially unknown to almost everybody, then. After they were done, my question was what they thought about the interaction. And every one thought it was a real person. None implied any issue being understood (even though ELIZA is strictly a string pattern processor and has zero internal representation of meaning). So this very minimal chatbot passed the Turing test’s main point.

Because of this, I knew the Turing test itself was invalid way back then. Now, it’s recognized as not valid, since modern LLMs can regularly fool people, even though they clearly also don’t have real sentience. And that’s because, at least for this kind of interaction, people generally are very easy to fool. The bar is just far too low for the Turing test to be a proper way to evaluate AI.

Edit: Additional thoughts. I think what the Turing test really shows is whether an AI is so poorly convincing that it wouldn’t fool a person. It’s not measuring the sentience or intelligent behavior of the software so much as whether it meets the minimum mechanical interaction that doesn’t itself make the user doubt it might be a normal person.

And etch-a-sketch was a precursor to television