First AI Chatbot: ELIZA, Jabberwacky or ChatGPT? Why the Answer Depends on What You Count as AI (Free ‘Which Was First?’ Tool)
First AI chatbot is one of those questions with three or four correct answers. If any program that chats counts, it’s ELIZA in 1966. If the program has to learn from conversations rather than follow rules, it’s Jabberwacky in the 1990s. If it must use a neural network and be open to the public, the answer moves to 2020 or 2022. This guide lines up the eight main contenders, explains exactly what each one could and couldn’t do, and lets you pick your own definition in a free tool that names the winner.
⚡ In short
- ELIZA (1966) is the first chatbot, but it followed hand-written rules and understood nothing.
- Jabberwacky, online from 1997, learned replies from what people typed: an early claim to the first chatbot that learned.
- Google’s 2015 research model was the first widely known chatbot built purely from a neural network.
- ChatGPT (2022) wasn’t the first neural chatbot, but it was the first used by the general public at huge scale.

Why it’s disputed
The question hides a second question: what counts as AI?
Ask ten people what the first AI chatbot was and you’ll hear ELIZA, ChatGPT, Siri, Cleverbot and possibly a few others. They’re not wrong; they’re answering different questions.
The word “AI” has meant different things in different decades. In the 1960s, any program that did something people thought needed intelligence, such as playing chess or holding a conversation, was called artificial intelligence, whatever was going on inside. By the 1990s, many researchers reserved the term for systems that learned from data. Today most people hear “AI chatbot” and picture a large language model that can discuss almost anything.
So before naming a winner you need a rule. Here are the four that matter most:
- Does it have to learn? Or is a program that follows hand-written rules allowed to count?
- Does it need a neural network? That’s the technology behind modern AI.
- Did the public have to be able to use it? Or do research systems count?
- Must it handle almost any topic? Or is a single role, such as a therapist or a shop assistant, enough?
Each rule changes the answer. The rest of this post walks through the eight best-known contenders in order, then the free tool lets you set the rules and see the result. If you’re curious how the *word* chatbot itself changed over the same years, our history of chatbots’ meaning covers that side.
Timeline
Eight contenders, in order
Alan Turing’s paper “Computing Machinery and Intelligence” proposes the imitation game: can a machine chat well enough to pass as a person? Not a chatbot, but the starting gun.
Joseph Weizenbaum at MIT publishes ELIZA. It reflects your words back as questions using keyword rules. The first chatbot.
Kenneth Colby’s program simulates a person with paranoid beliefs, with internal emotional states. It and ELIZA were connected for a famous computer-to-computer chat in 1972.
Bundled with Sound Blaster cards, it speaks ELIZA-style replies aloud on home PCs.
Richard Wallace’s rule-based bot with thousands of AIML patterns, chatting on the early web.
Rollo Carpenter’s bot goes online and learns replies from earlier visitors. Becomes Cleverbot in 2008.
Google researchers train a neural network on conversations; it replies with no hand-written rules.
Meta releases BlenderBot openly in 2020; OpenAI launches ChatGPT to the public on 30 November 2022.
Contender 1
ELIZA: the first chatbot, and the case against calling it AI
ELIZA is the safe answer, and in most quizzes it’s the expected one. Joseph Weizenbaum wrote it at MIT and published a paper about it in January 1966. Its most famous script, DOCTOR, imitated a psychotherapist who answers questions with questions.
Type “I am worried about my exams” and ELIZA might reply “How long have you been worried about your exams?” It found the keyword pattern “I am …”, swapped “my” for “your”, and slotted your words into a template. If nothing matched, it fell back on lines like “Please go on.”
That’s the case against calling it AI in today’s sense. ELIZA had no memory of the conversation beyond a few stored phrases, learned nothing, and understood nothing. Weizenbaum himself was alarmed that people, including his own secretary, confided in it as if it understood them. He spent much of his later career warning against exactly that.
The case *for* is historical. In 1966 ELIZA was an artificial intelligence project, built at a leading AI lab, and it was the first program to hold an open text conversation at all. If your rule is “the first program designed to chat with people”, ELIZA wins without argument.
For business owners, ELIZA’s lesson has aged well. A simple keyword bot can feel surprisingly helpful when its rules match what people actually type, and frustrating when they don’t. That’s still how free keyword bots work; our rule tester guide shows how to write rules that don’t misfire.
Contenders 2 to 4
PARRY, Dr. Sbaitso and A.L.I.C.E.: smarter rules, still rules
The next three contenders all improved on ELIZA without changing its basic method: people wrote the rules.
PARRY (1972) is the most interesting. Kenneth Colby, a psychiatrist at Stanford, built it to simulate a person with paranoid schizophrenia. Unlike ELIZA, PARRY kept internal variables for fear, anger and mistrust, and those levels changed how it replied as the chat went on. Psychiatrists reading transcripts reportedly struggled to tell PARRY from real patients, an early informal Turing-style test. Still, every behaviour was programmed by hand.
Dr. Sbaitso (1991) came free with Creative Labs’ Sound Blaster sound cards to show off text-to-speech. It asked your name, then chatted in an ELIZA-like way, out loud. For many people in the early 1990s, this was the first chatbot they ever met at home, which gives it a claim to “first AI chatbot the public could use”, if rule-based programs count.
A.L.I.C.E. (1995), by Richard Wallace, used a markup language called AIML containing thousands of pattern-and-response rules. It ran on the web, anyone could chat with it, and it won the Loebner Prize, a yearly contest for human-like chatbots, three times in the early 2000s. AIML is still used by hobbyists today.
If your definition requires learning, all three are out. If it only requires a public, conversational program, Dr. Sbaitso (bundled with PCs) and A.L.I.C.E. (on the open web) are the earliest strong candidates after the lab-only ELIZA and PARRY.
Free tool
Choose your definition, find your first
Tick the rules a program must meet to count as an AI chatbot. The tool names the earliest well-known program that meets them all, and shows which earlier ones your rules exclude.
What has to be true for a program to count? Tick the rules you think a real AI chatbot must meet. The tool finds the earliest well-known program that meets all of them.
Runs only in your browser. Dates are the commonly cited year each program appeared or became available; sources differ by a year or two for some.
Contender 5
Jabberwacky: the first chatbot that learned from people
If your rule is "it must learn", the story moves to Jabberwacky. Rollo Carpenter, a British programmer, began working on the idea in the 1980s. It went online in 1997.
Jabberwacky's method was simple and radical. Instead of hand-writing replies, it stored what people typed and reused those human replies in later conversations when the context looked similar. The more people chatted with it, the more it had to say. Its descendant, Cleverbot, launched in 2008 and grew to hold a very large collection of stored exchanges.
That counts as learning in a meaningful sense: its behaviour came from data, not from a programmer predicting every question. But it wasn't a neural network, and it didn't understand meaning. It picked a plausible stored reply, which is why Cleverbot conversations swing between uncanny and nonsensical.
Two lessons carry over to modern business bots. First, learning from real conversations is powerful; it's why a bot trained on your own customers' questions beats one built from imagination. Second, learning without supervision is risky. When Microsoft's Tay learned from public tweets in 2016, it picked up offensive language within a day and was taken offline.
That's why today's business bots, including LiveChatBot's AI Agent, answer only from knowledge you provide instead of learning freely from every visitor. Our AI training guide explains how that works in practice.
Contenders 6 to 8
The neural era: Google's model, BlenderBot and ChatGPT
The final rule, "it must use a neural network", moves the answer into the 2010s.
Google's neural conversational model (2015). Researchers Oriol Vinyals and Quoc Le published a paper showing a neural network that learned to converse from examples alone: movie subtitles and IT helpdesk chats. No rules, no templates. It could answer simple general questions and troubleshoot a VPN in transcripts. But it was a research demonstration, not something the public could use.
BlenderBot (2020). Facebook AI Research, now Meta, released BlenderBot as an open model anyone could download and run, trained to chat engagingly about almost any topic. A public web demo, BlenderBot 3, followed in August 2022, in the US only. If your rules are "neural, public, any topic", BlenderBot has a genuine claim, especially for developers.
ChatGPT (2022). OpenAI launched ChatGPT on 30 November 2022 as a free web app built on a large language model. It wasn't technically first at anything in particular, but it was the first neural chatbot that the general public used in enormous numbers, and it changed what "AI chatbot" means to most people.
So the honest answer to "which was the first AI chatbot?" is usually a sentence, not a name: *ELIZA was the first chatbot; ChatGPT made AI chatbots mainstream; in between, Jabberwacky learned from people and Google and Meta built the first neural ones.*
At a glance
What each contender could do
| Year | Program | Learns? | Neural? | Public? | Any topic? |
|---|---|---|---|---|---|
| 1966 | ELIZA | No | No | No (lab) | No |
| 1972 | PARRY | No | No | No (lab) | No |
| 1991 | Dr. Sbaitso | No | No | Yes | No |
| 1995 | A.L.I.C.E. | No | No | Yes | No |
| 1997 | Jabberwacky | Yes | No | Yes | No |
| 2015 | Google neural model | Yes | Yes | No (research) | Yes |
| 2020 | BlenderBot | Yes | Yes | Yes (open model) | Yes |
| 2022 | ChatGPT | Yes | Yes | Yes | Yes |
Why it matters now
What 60 years of chatbot history teaches a small business
History is fun, but there's a practical point too. Every lesson from ELIZA onward still applies to the bot on your website.
People trust bots more than they should. Weizenbaum's ELIZA showed in 1966 that people confide in programs that seem to listen. Your bot must never pretend to know what it doesn't. Make it admit uncertainty and hand over to a person.
Rules work when they match real words. ELIZA, PARRY and A.L.I.C.E. all worked well within their scripts and fell apart outside them. A keyword bot built from your customers' real questions is useful; one built from guesses isn't.
Learning needs limits. Jabberwacky showed the power of learning from people; Tay showed the danger. Let your AI answer from knowledge you control.
Fluent doesn't mean correct. Modern neural bots write beautifully even when they're wrong. Check answers against your facts before visitors see them.
LiveChatBot combines the old and new approaches: an AI Agent that answers from the knowledge you give it, with a small free monthly allowance per website, and a free keyword bot that answers from your saved replies when AI isn't available, all handing over to your team in one inbox. See what a chatbot is today, the AI Agent, or start at livechatbot.in.
Frequently asked questions
What was the first AI chatbot?
ELIZA, written by Joseph Weizenbaum at MIT and published in 1966, is generally called the first chatbot. If you require the program to learn, the answer is Jabberwacky (online 1997); if it must use a neural network and be public, BlenderBot (2020) or ChatGPT (2022).
Was ELIZA really artificial intelligence?
It was built as an AI research project, but it only matched keywords and filled templates. It didn't learn or understand, so by today's standards most people wouldn't call it AI.
Did ChatGPT come first?
No. Neural chatbots existed before it, including Google's 2015 research model and Meta's BlenderBot. ChatGPT was the first to be used by the general public at massive scale.
What was the first chatbot to pass the Turing test?
No chatbot has passed a rigorous, widely accepted Turing test. Some contests and short experiments have claimed passes, but they're disputed.
Is the 'Which was first?' tool accurate?
It uses commonly cited dates and simplified yes/no features for well-known programs. Sources differ by a year for some, and other lesser-known programs existed, so treat it as a guide to how definitions change the answer.
Keep reading
Related guides
Features of Chatbot Tools, Verified: How to Test Each Claimed Feature During a Free Trial Before You Trust It (Free Feature Test Sheet)
Examples of Conversational AI Beyond the Website Chatbot: 10 Places People Already Talk to AI, and Which Ones a Small Business Can Use (Free Skill Spotter)
Educational Chatbot Examples for Schools, Coaching Centres and Colleges: 12 Conversations Sorted by Who Should Answer (Free Question Router)
Sixty years later, put a chatbot on your site
An AI Agent that answers from your knowledge and hands over to your team. Start free on livechatbot.in.