The Man Who Built the First Chatbot, Then Warned Us

Author: Better Stack · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: The Man Who Built the First Chatbot, Then Warned Us · Channel: Better Stack · 09:18 · Watch on YouTube

Topics: AI agents, AI governance and regulation, AI models, AI safety

Summary

The video revisits Eliza, Joseph Weizenbaum’s 1966 chatbot, running its recovered original code and demonstrates how its keyword rules, scripts, and quirks work. It shows how inputs such as numbers and question marks can crash or disrupt the program, and explains how researchers restored the code after finding it in an MIT archive. Eliza’s simple mirroring behavior led people to treat it as more understanding than it was, prompting Weizenbaum to warn against assigning computers tasks that require wisdom. The video connects those concerns to emotional attachment to today’s chatbots.

Key points

Eliza was created by MIT professor Joseph Weizenbaum and published in a 1966 paper.
The Doctor script makes Eliza respond by matching keywords, changing pronouns, and inserting the user’s words into templates.
When it finds no keyword, Eliza may give a vague response or reuse something said earlier.
The original program can crash on number input, while pressing the question mark on the old system deletes the line and can trigger a script dump.
Eliza is an engine that can be given different scripts, and users could teach it new rules while it was running.
Researchers recovered the original code from an MIT archive box and restored it; the version running today is 96% exactly what they found.
Weizenbaum was alarmed that people—including his secretary and some psychiatrists—treated the program as if it understood them.
The video links Weizenbaum’s warnings about harmful attachments to concerns about emotional attachment to modern AI chatbots.

Chapters

Start	Chapter	Summary
00:00	Eliza runs again	The video introduces Eliza’s recovered 1960s code and explains its origins, hardware, and therapist-style conversation.
02:34	How Eliza works—and how it breaks	The video explains Eliza’s fallback replies and demonstrates how numbers and the question mark can disrupt the program.
05:40	Recovering the original program	Researchers reconstruct and run the archived code, then trace how Eliza’s clones spread while the original was forgotten.
07:44	Weizenbaum’s warning	The video describes how people attributed understanding to Eliza and connects Weizenbaum’s concerns to modern chatbot attachment.

Quotes

This is Eliza, the first chatbot ever made running its original 1960s code — 00:00

Eliza is basically just a mirror with some rules. — 01:31

So, you literally can’t ask the first chatbot a question — 03:38

we ought not now to give computers tasks that demand wisdom. — 08:16

Knowing how the trick works didn’t work on a secretary and I don’t think it stops it working on any of us either. — 09:18

Concepts and entities

Eliza, Joseph Weizenbaum, MIT, SLIP, IBM 7094, ARPANET, BBN, Eliza effect, Inventing Eliza, OpenAI, ChatGPT

Full transcript (original language)

00:00 This is Eliza, the first chatbot ever made running its original 1960s code, and it still answers exactly the same way it did in 1966. And if I type a number, it still crashes. This exact code sat unseen in a box at MIT for over 50 years, and now it runs again on my Mac in a program that pretends to be a 1960s mainframe. I’m going to replay its most famous conversations and break it in a couple different ways, and then we’ll get to why its creator spent his life warning people about it, because what he was warning about is the exact problem we’re

00:30 seeing with AI chatbots today. >> [music] >> So, quick explainer. Eliza was built by Joseph Weizenbaum, a professor at MIT, and he published it in a paper in 1966. It ran on a computer the size of a room, an IBM 7094, a machine that cost $2.9 million, which is around $23 million today, and it had less memory than a single photo on your phone. There was no screen, either. You had to type on a typewriter wired up to the computer, and it typed

01:01 its replies back on the same paper. There’s also one key that you weren’t allowed to press, which is the question mark. And remember that, because of course we’re going to press that later on. Now, Weizenbaum printed one full conversation with Eliza in his paper, so the real test is whether the code from that box actually talks the way his paper says it does. To find out, I’ve loaded the doctor script from his paper, where Eliza plays a therapist, and I’m typing in the exact lines he published. Men are all alike. >> In what way? >> They’re always bugging us about something or other.

01:31 >> Can you think of a specific example? >> That is word-for-word what Weizenbaum printed in 1966. And the next line shows us some of how this works. Well, my boyfriend made me come here. Comes back as >> Your boyfriend made you come here? >> Eliza is basically just a mirror with some rules. It spots a keyword and swaps my for your and me for you, and slots your own words into a template reply. If it does spot more than one keyword, the higher-ranked one wins. It has no idea what a boyfriend actually is. And that’s why Weizenbaum made this a therapist.

02:02 His example was if you tell a therapist you went on a long boat ride, and they say tell me about boats, you don’t think they’re clueless, you assume they’re going somewhere with it. But this system is also not without its faults. If you tell it you were like my father in some ways, it’ll just reply with equals DIT. That’s one of its own rules leaking out. It’s an instruction to use its DIT replies, which this older code can’t follow there. So it prints E instead. If we keep this conversation going the way Weizenbaum did, we get to the last line, bullies. >> Does that have anything [clears throat] to do with the fact that your boyfriend

02:34 made you come here? >> And this one is revealing another trick. When it finds no keyword, it either says something vague like please go on, or it brings back something you said earlier. So when you talk about my something, it can save a copy for later, so it looks like it’s been listening the whole time. You can run this all yourself by the way, it’s all open source, and I’ll leave the links in the description down below. So now that we know it can hold some kind of conversation, let’s try and break it by telling it you are 999 today. When you do this, the whole program dies, putting me straight back to a 1960s command line. To explain this

03:06 crash, if you think of your sentence as a chain of sticky notes, where each note says where to find the next one, Eliza’s library for handling list of words, called SLIP, works a bit like that. When you type a number, it mistakes that number for one of those directions, and then readdress, so it goes looking in some random spot in memory, and falls over. The team this code only found this out by actually running it, and they left this in on purpose, and we’ll get back to why later. But first, a number isn’t the only way to break this. Remember the key that you weren’t allowed to press earlier? Let’s ask it

03:38 how are you with a question mark. On MIT system, the question mark was the delete this line key, so it wipes my whole line, and Eliza thinks I’ve sent it nothing. A blank line is its signal to save its script, so it asks me for a number, then prints its entire script onto the screen and quits. So, you literally can’t ask the first chatbot a question, and if you try, it dumps every rule it has out onto its screen. This is technically the first chatbot system prompt leak, but it also reveals the final piece of how Eliza works. The

04:08 script it just dumped is its entire personality. Eliza itself is just an engine, and the script on top tells it who to be. Pretty similar to how chatbots you use today have one model with a system prompt on top. A new book from MIT Press this July called Inventing Eliza makes that same comparison. This model was never just a toy therapist, it was the first chatbot platform. You could even rewrite its personality while it was running. If you type a plus sign, Eliza stops chatting and says, “Please instruct me.” So, here I’ve added a rule where if you say hi,

04:39 it replies, “Hello from” and then whatever came after the hi. But, if I say hi to all my friends, it ignores my rule and just says, “Your friends.” because my ranks higher. So, I go back in, rank my rule above it, and now it says, “Hello from to all your friends.” This ability to teach is actually where its name comes from. In the play Pygmalion, which later became My Fair Lady, a professor teaches a flower girl called Eliza Doolittle to speak like an upper-class lady. Weizenbaum named his program after her because it could be taught, too. And like her, it could also

05:09 play a part. It even starts by asking, “Which script do you wish to play?” The therapist was just one of those scripts. The Inventing Eliza team found others in the MIT’s archive, like one that only makes small talk about New England. In one conversation, they show someone types, “I like Missouri.” and it replies, “Missouri is not a New England state.” Now, getting any of this running nearly didn’t happen. The code is about 2,600 lines and mostly undocumented. Text recognition struggled on 60-year-old printer ink, so two researchers typed most of it back by

05:40 hand. One function wasn’t in the archive at all or any manual, and in their words, “Google was not helpful.” So, they worked out what it must do and wrote it themselves. When they finally ran it, they tried sending men are all alike and it crashed. The cause was one typo deep in the code, four zeros where there should have been five. After they fixed that, on December 21st, 2024, Eliza held a conversation for the first time in over 60 years. So, how does the first chatbot go missing for that long? Well, it kind of got beaten by its own

06:12 clones. Shortly after the paper came out, a programmer at BBN wrote his own version in Lisp and BBN was building the ARPANET, which was the network that became the internet. So, that copy spread to university and Weizenbaum’s original sat on an IBM machine that was never connected to it. On top of this, his paper printed the doctor script, which looked exactly like Lisp, so people assumed for about 50 years that Eliza was a Lisp program. Then, in 1977, a magazine published a version in BASIC that was only a couple of pages long,

06:43 right as home computers took off. That one got copied into every language [music] that you can think of and the researchers call Eliza perhaps the most knocked off program in history. Funnily enough, the person who wrote that BASIC version is the same person who found the original. In 2021, he and an MIT archivist pulled out a box labeled computer conversations box eight and inside was the printout of Eliza’s code plus conversations nobody had ever seen before and a lot of them hand edited by Weizenbaum. Once they found this code, they wanted to restore it to its

07:13 original and the code running today is 96% exactly what was found in that box. Going back to the bugs, they said that fixing the number bug would mean changing its core and in their words, finding bugs only adds to the authenticity. They even named that part of the paper after Ada Lovelace, who wrote back in the 1840s that computers could work on other things besides numbers. It’s just that Eliza took that a little bit too literally. So, that’s how Eliza works and how it went missing. But, now let’s talk about the darker part, why its creator spent the rest of

07:44 his life warning people about AI. Under the hood, Eliza is just a mirror handing your own words back to you, but the trouble is people didn’t treat it like one. Weizenbaum’s own secretary knew full well it was just a program that he’d written, but after only a few lines of chatting with it, she asked him to leave the room so she could talk to it in private. On top of this, psychiatrists started to take it seriously as well. Three of them wrote in 1966 that a computer system like this could handle several hundred patients an hour. Weizenbaum was so alarmed he wrote that simple program like this could

08:16 induce powerful delusional thinking in quite normal people. And this is now called the Eliza effect. So, all of this turned the man who built the first chatbot into one of AI’s earliest critics. In his 1976 book, he wrote that since we have no way of making computers wise, we ought not now to give computers tasks that demand wisdom. He worried that people would form harmful attachments to machines and he kept warning about this until he died in 2008. And this is kind of the exact problem that we’re seeing with AI chatbots today. A year ago, OpenAI said

08:46 around 0.15% of ChatGPT’s users showed heightened levels of emotional attachment to it. At the time, over 800 million people use ChatGPT every week, so that’s more than a million people. And the pattern that OpenAI says it’s worried about is attachment at the expense of real-world relationships. Weizenbaum actually opened his 1966 paper with the line, It is said that to explain is to explain away. Once you see how a program works, the magic crumbles and you think, I could have written that. Overall, Eliza is just a mirror with a few rules and people still

09:18 confided in it. Today’s chatbots are wildly more capable than this, which makes his warning more relevant and not less. Knowing how the trick works didn’t work on a secretary and I don’t think it stops it working on any of us either. Let me know what you think about all of this in the comments down below. While you’re there, subscribe and as always, see you in the next one. >> [music] [music]