INFOGRAPHIC
From early iterations in the 1960s and 1970s to the complex creations of the 2020s, AI chatbots have come a long way from their rule-based beginnings.
Depending on how you define the term “chatbot,” there are likely millions of them in existence today being used by more than a billion people, by some estimates.
According to the Pew Research Center, about half of all adults under 50 interact with AI on a daily basis. Meanwhile, the share of U.S. workers who use AI to accomplish some of their work grew from 16 percent in 2024 to 21 percent in 2025, although about 65 percent don’t use AI much if at all to perform their work.
If we’re talking about general-purpose AI chatbots like ChatGPT, although there are fewer out there, the number of these platforms is growing as well. And their market share is evolving as well, as new platforms continue to go mainstream. In December, Similarweb pinned the market share of ChatGPT at 68 percent and Gemini at 18.2 percent, with Claude coming in at 2 percent and Copilot at 1.2 percent. ChatGPT had an 87 percent share in December 2024, by comparison.
But every tool has a beginning. Here’s a look back at a selection of the many systems and software that made it all possible. (We even tested out doing some of the research for this list using an AI chatbot, which seemed quite fitting.)
1966
ELIZA // RULE-BASED SYSTEMS
Created by Joseph Weizenbaum, MIT
The first chatbot that could simulate human conversation was ELIZA, which utilized rule-based logic in the form of pattern matching and substitution methodology. This created the illusion of human conversation by passing words entered into a computer and matching those to scripted responses from a script. It’s where the phenomenon called the ELIZA effect, which is the projection of human traits onto programs and machines, originates. Weizenbaum developed the program between 1964 and 1967, and published his research in January 1966.
1972
PARRY // RULE-BASED SYSTEMS
Created by Kenneth Colby, Stanford University
Rather than relying on pattern matching, PARRY took a step beyond ELIZA with its use of internal state modeling. This incorporated a system of beliefs and emotional responses into the program’s answers, which were also more conversational, making interactions with it much more believable. Colby, a psychiatrist and computer scientist, designed PARRY to simulate an individual with paranoid schizophrenia. It was the first chatbot to pass a variation of the Turing Test, which determines a program or machine’s ability to mimic human conversation.
1995
ALICE // RULE-BASED SYSTEMS
Created by Richard Wallace
The Artificial Linguistic Internet Computer Entity, or ALICE, used heuristic pattern matching to simulate conversations. Activated in November 1995, the program was the first Artificial Intelligence Markup Language (AIML)-based personality program. Richard Wallace created AIML as an open-source XML-based framework, which allowed users worldwide to contribute toward its development. ALICE also became a widely used chatbot development standard for years after its creation.
2001
SMARTER CHILD // HUMAN-CURATED RETRIEVAL SYSTEM
Created by ActiveBuddy, Inc.
Individuals of a certain age will remember the days of instant messaging, from AOL to MSN and Yahoo. SmarterChild launched on all three platforms and was the first bot to integrate real-time data feeds into a conversational interface at scale. It was the first mainstream consumer chatbot as well. Using natural language queries, SmarterChild’s responses were human-curated, so it was unable to generate anything on its own. This bot had more than 30 million users over its lifetime, peaking at 17 million concurrent users. The program was eventually renamed Colloquis, then decommissioned after Microsoft acquired it in 2007.
2011
WATSON // PROBABILISTIC QUESTION-ANSWERING SYSTEM
Created by IBM Research (led by David Ferrucci)
What better reason to build a chatbot than to defeat a Jeopardy! champion? That’s what IBM did when it created Watson more than a decade ago. The chatbot was powered by DeepQA, a massively parallel software architecture with the ability to examine natural language content. A team of researchers spent three years developing Watson, which used 16 terabytes of RAM and ran at 80 teraFLOPS. The system generated and scored hundreds of hypotheses probabilistically in almost real time. Watson beat two Jeopardy! champions, Ken Jennings and Brad Rutter, in 2011, using simply its pre-loaded knowledge base.
2011
SIRI // TASK-ORIENTED VOICE ASSISTANTS
Created by SRI International/Siri, Inc.
The tech that many of us carry in our pockets today began life as the CALO (Cognitive Assistant that Learns and Organizes) project, under DARPA's Personalized Assistant that Learns (PAL) program. Siri merges speech recognition, neutral language processing, and machine learning to perform tasks and answer questions. SRI led what it calls the largest AI project in U.S. history, with a total investment of about $150 million across five years. SRI launched Siri, Inc. in 2007 and created a standalone iPhone app in February 2010. Apple then acquired Siri, Inc. in April 2010, and it became an integrated feature starting with the with the iPhone 4S in October 2011. It was the first of a series of voice assistants, from Amazon Alexa to Google Assistant and Samsung Bixby.
2022
ChatGPT // GENERATIVE AI SYSTEMS
Created by OpenAI
At the heart of today’s generative AI language models is the Transformer neural network architecture, first introduced in a 2017 Google paper, “Attention Is All You Need.” ChatGPT, which stands for Chat Generative Pre-trained Transformer, combines the Transformer architecture with massive-scale internet text pretraining, and reinforcement learning from human feedback (RLHF). The result is the first AI chatbot capable of generalizing about any subject. The application reached 100 million users within two months of launch.
2023
CLAUDE // GENERATIVE AI SYSTEMS
Created by Anthropic
Named after Claude Shannon, considered the father of information theory, Claude was built using a training methodology called Constitutional AI (described in “Constitutional AI: Harmlessness from AI Feedback.”) Instead of relying on human feedback like RLHF does, CAI trains the Claude model to critique and revise its own outputs using a written set of principles. This feedback helps reinforce helpful behavior.
2023
GOOGLE GEMINI // GENERATIVE AI SYSTEMS
Created by Google DeepMind
First known as Google Bard when it was created in 2023, this chatbot was renamed Gemini in 2024. It was multimodal from the beginning, as a single model handling text, images, audio, and video. Gemini’s million-token context window enables long-context reasoning, so entire document libraries can be analyzed in a single session. Gemini 1.5 introduced a 1,000,000-token context window, which was far larger than GPT-4’s 32,000 token window and Claude’s 100,000 token window.
© 2026 The American Society of Mechanical Engineers. All rights reserved.