History of Chatbots: How Old AI Chatbots Became AI Agents

When did chatbots start?

Chatbots started in 1966 with ELIZA, a program built by Joseph Weizenbaum at MIT that simulated conversation by matching keywords and replying from a fixed script. Every old AI chatbot worked that way: ELIZA, then PARRY in 1972, ALICE in 1995 and SmarterChild in 2001. The modern kind, trained on a company's own content and able to answer a question nobody wrote a script for, only arrived after 2022.

This page traces the whole line, from ELIZA to the AI agents businesses put on their websites today, and ends with what that history tells you to do if you want one answering questions on your own site.

Key takeaways

• Chatbots started in 1966 with ELIZA, the first program to hold a convincing text conversation.
• Old AI chatbots were rule based. A developer wrote the keywords and the replies by hand.
• PARRY, ALICE and SmarterChild refined that approach, and Siri and Alexa took it mainstream.
• Large language models removed the scripts after 2022, so an agent can be trained on a business's own material instead.
• What took a research lab months in 1966 now takes under 60 seconds.

The history of chatbots: a timeline from 1966 to 2026

Six moments account for most of the change.

• 1966, ELIZA. Joseph Weizenbaum at MIT builds the first program to hold a human-like text conversation, using pattern matching and substitution.
• 1972, PARRY. The psychiatrist Kenneth Colby models a person with paranoid schizophrenia, adding an internal model of mood and belief on top of the pattern matching.
• 1995, ALICE. Richard Wallace's Artificial Linguistic Internet Computer Entity runs on a far larger hand-written rule set and wins the Loebner Prize three times.
• 2001, SmarterChild. The first chatbot most people actually met, on AOL Instant Messenger and MSN Messenger, fetching news, weather and sports scores on request.
• 2010s, voice assistants. Siri, Alexa and Google Assistant put conversational interfaces on phones and kitchen counters, and raised what people expect from a machine that talks.
• 2022 to 2026, large language models. Open-ended conversation stops being scripted. By 2026 a business can train an AI agent on its own website and documents and have it live the same day.

ELIZA and the first AI chatbot

In 1966, Joseph Weizenbaum at MIT released ELIZA. It looked for keywords in what you typed and answered with a pre-set phrase built around them. The technique was simple. The effect was not.

ELIZA's best-known script, DOCTOR, played a therapist, reflecting your own words back at you as questions. People confided in it, and some refused to believe there was no one there. That gap between what the program did and what users believed it understood is still called the ELIZA effect, and it is worth remembering every time a chatbot sounds more capable than it is.

ELIZA proved a computer could hold the shape of a conversation. Everything since, including the AI agent answering questions on a website today, has been an attempt to fill that shape with real understanding.

The role of natural language processing

The step from ELIZA to later chatbots was mostly more rules, written better. Natural language processing is what let programs start working out intent instead of matching strings: reading the structure of a sentence, resolving what "it" refers to, tolerating a typo. Machine learning then let a system improve from data rather than from a developer's next edit. Both had to arrive before a chatbot could answer a question its builders had never anticipated.

What were old AI chatbots, and why did they break?

An old AI chatbot was a rule engine. A developer wrote keywords, decision trees and canned replies, and the program matched your message against those rules and picked one. That is why old chatbots felt like a phone menu in text form. Ask what the script anticipated and you got an answer. Phrase it slightly differently and you got "Sorry, I don't understand."

Three limits defined the entire generation:

• No real language understanding. Keyword matching breaks on typos, synonyms and any follow-up that depends on what was said a moment earlier.
• Every answer hand-written. The chatbot knew exactly what its builders typed into it and nothing more, so coverage was always partial and the maintenance never ended.
• Dead ends by design. When the script ran out, the conversation stopped. That is where the reputation for infuriating chatbots came from, and it is still what many people expect when they see a chat window open.

Old chatbots vs modern AI agents: what actually changed

The practical difference is where the knowledge lives. You no longer write the answers. You point the agent at material you already have and it grounds what it says in that. Dante AI works exactly this way: point it at a website, it learns that site's content, and one script tag puts an agent on the page that answers visitors from the site's own material and captures leads. Our guide on how to train a chatbot on your own data covers what to feed it.

What the history means for your website today

Every limit on that list has already been solved by someone else. You do not write scripts, you do not maintain a decision tree, and you do not need a developer. What is left is deciding what your agent should read.

When we speak with new users, the hesitation is almost never technical. It is working out which pages and documents genuinely represent the business. The history gives you the shortcut. ELIZA failed because it knew nothing, so start by giving your agent the material that makes it right the first time.

Three steps, all of them in the dashboard:

1. Point it at your website. Create a free Dante AI account, paste your URL, and let it read the pages you already publish.
2. Add the documents your team keeps re-sending. Price lists, policies, FAQs, onboarding guides. The answers usually exist already, buried in email.
3. Paste one script tag into your site. The agent goes live, answers from your own material and collects the visitor details worth following up.

Then read the conversation logs for a week. What people actually ask is the most honest content brief you will ever get, and it is the one thing no chatbot before 2022 could hand you.

What to expect once your agent is live

• Visitors get an answer at the moment they ask, not the next working day
• Repetitive questions stop reaching your inbox
• You can see, in writing, what people came to the site to find out
• Anything the agent cannot answer goes to a person instead of dead-ending
• Leads are captured inside the conversation rather than in a separate form

Give it one honest test before you decide anything. Write down the five questions your inbox answers most often this month, put them to the agent, and compare each reply with what you would have typed yourself. Weizenbaum's users were fooled because they could not see the script. You can: every answer traces back to a page or a document you supplied, so when one comes back thin, you already know what to add.

You can build and test the whole thing on the free plan, which includes 100 message credits a month plus up to 700 more for completing onboarding, and paid plans start at $40 per month when you need more volume.

Further reading

Keep going with these guides from the Dante AI library:

• How to Create an AI Chatbot for Your Website in 60 Seconds
• What Is an AI Chatbot?
• How Do You Implement an AI Chatbot From Scratch?