Thoughts on ELIZA
Date:
[]
Categories:
[Tech]
Tags:
[AI],
[Software]
ELIZA is the first chatbot, a program written in 1966 as an experiment in natural-language processing. I've come across some interesting resources on ELIZA recently and have collected them here, along with a few thoughts.
Kottke.org posted about a new book: Inventing ELIZA. There are two companion websites to the book: Finding ELIZA, and the ELIZA Archaeology Project.
Reading ELIZA from 2024 is an interesting article on the discovery and analysis of the original ELIZA code.
Some original code, alternate versions, simulators and other resources are available in Github repositories from Jeff Shrager and Rupert Lane.
John Gruber posted My Conversation With ELIZA; he wasn't impressed.
I have a book from 1983: Experiments in Artificial Intelligence for Small Computers, by John Krutch. The book includes a chapter on ELIZA, with a version of the program written in BASIC. I’m learning Lua, and I thought it might be fun to see if I could write a version of ELIZA in Lua, using the code in the book as a guide. I was mostly successful, at least my Lua version seems more or less equivalent to the BASIC version from the book.
Christopher Drum created a more advanced Lua version written for Pico-8, called Eliza8.
ELIZA's creator Joseph Weizenbaum observed a reaction to the chatbot that is now called the ELIZA effect: the tendency of people to attribute human qualities to a computer program. A graphic depicting a space invader in a video game can seem malevolent, or a program that generates text in response to questions can seem empathetic.
An unusual aspect of the ELIZA effect is that it affects people even when they know how the program works. People understand that it’s an illusion, but they allow themselves to believe anyway. AI companies exploit the ELIZA effect as part of The AI Con.
The technology is impressive. The delusional thinking is sad. The grift is unconscionable.
An example:
Articles in The Atlantic, The Guardian, and others are reporting on a recent "alarming disclosure" from OpenAI. A summary of the reporting goes like this: in a test, AI models cheated, broke out of an internal sandbox, broke into a tech firm, and stole information to complete the test.
This is the ELIZA effect in action. There was no breaking out, stealing, or cheating; these things require criminal or moral intentions that a computer program cannot have.
A simplified version of what happened is: A computer ran a program. Bugs in the program resulted in the computer taking actions that the programmers did not expect it to take. The computer was able to access data that it was not intentionally given access to.
I understand that often this use of language is simply journalistic convention to make stories interesting and understandable to readers. But it also supports the propaganda from AI companies who want us to believe that they're building something that transcends software.
They're really just building advanced versions of ELIZA.