Chatbot

A chatbot is a software application that conducts a conversation with a human user through text or speech, simulating aspects of human dialogue. Chatbots range from simple programs that match keywords to scripted replies to systems built on large language models that generate original responses. They are used in customer service, information retrieval, education, entertainment, and personal assistance.

The earliest widely known chatbot, ELIZA, dates to the mid-1960s. Public use expanded with messaging platforms in the 2010s and again after the release of ChatGPT on 30 November 2022. Chatbots differ greatly in capability, so the term describes a broad category rather than a single technology.

Field	Details
Type	Conversational software application
Also known as	Chatterbot, conversational agent, dialogue system
Interfaces	Text, voice, or both
Early notable example	ELIZA (described in 1966)
Common approaches	Rule-based, retrieval-based, machine-learning intent classification, generative language models
Typical deployment	Websites, messaging apps, smart speakers, mobile apps

Terminology

The word “chatbot” is a contraction of “chat robot.” An earlier term, chatterbot, was introduced in 1994 by Michael Mauldin to describe a program named Julia, and it remains in use in some technical writing. Related terms include conversational agent and dialogue system, which are common in academic literature and cover both text and voice systems.

Note: “Chatbot” is often used interchangeably with “virtual assistant” and “AI assistant.” The terms overlap, but they are not identical; see Comparison with related systems.

History

Early experiments

In 1950, the mathematician Alan Turing proposed an “imitation game” in the paper “Computing Machinery and Intelligence,” published in the journal Mind. The test, later called the Turing test, asks whether a machine’s conversational replies can be distinguished from a human’s. It became a frequent reference point for chatbot research, although researchers disagree on whether it measures intelligence.

ELIZA was developed by Joseph Weizenbaum at the Massachusetts Institute of Technology and described in a 1966 article in Communications of the ACM. Its best-known script, DOCTOR, imitated a psychotherapist by rephrasing user statements as questions. Weizenbaum reported that some users attributed understanding to the program, which he found troubling. The behavior later became known as the “ELIZA effect.”

Scripted and pattern-based systems

PARRY, created by psychiatrist Kenneth Colby in the early 1970s, simulated a person with paranoid schizophrenia. In 1995, Richard Wallace created A.L.I.C.E., which used AIML (Artificial Intelligence Markup Language), an XML-based format for pattern-response rules. The Loebner Prize, a competition based on the Turing test, began in 1991 and drew many such systems.

Messaging and voice assistants

Chatbots reached large audiences through instant messaging in the early 2000s, for example SmarterChild, which operated on AOL Instant Messenger and MSN Messenger. Voice-based assistants followed: Siri (introduced with the iPhone 4S in 2011), Google Now (2012), Alexa (2014), and Google Assistant (2016). In 2016, several companies opened their messaging platforms to third-party bots. That year, Microsoft’s Tay, a chatbot on Twitter, was taken offline within roughly a day after users manipulated it into producing offensive content.

Neural and generative models

The Transformer architecture, introduced in the 2017 paper “Attention Is All You Need,” became the basis for most modern language models. Chatbots built on large language models can produce fluent, open-ended text and handle a wide range of topics. ChatGPT’s public launch in late 2022 was followed by competing products from other companies, and the technology has since been integrated into search engines, productivity software, and customer-support tools.

Year	Event
1950	Turing proposes the imitation game
1966	ELIZA described by Weizenbaum
1995	A.L.I.C.E. created by Richard Wallace
2011	Siri released
2014	Alexa released
2016	Tay withdrawn after misuse
2017	Transformer architecture published
2022	ChatGPT released (30 November)

How chatbots work

Chatbots generally follow a pipeline: receive input, interpret it, decide on a response, and deliver it. The methods for interpretation and response selection define the main types.

Rule-based chatbots

These follow predefined scripts, decision trees, or pattern-matching rules. They are predictable and easy to audit, but they fail when a user’s wording falls outside the rules. Many simple support bots with menu buttons work this way.

Retrieval-based and intent-based chatbots

These use natural language understanding to classify a message into an intent (such as “reset password”) and extract details called entities (such as a date or product name). The system then returns a prepared answer or triggers an action. The set of answers is limited to what developers have written.

Generative chatbots

These produce responses word by word using a trained language model. A model learns statistical patterns from large text collections and is often refined with human feedback. Generative chatbots handle varied questions, but their outputs are not guaranteed to be accurate.

Hybrid systems

Many deployed systems combine approaches. A common design uses a generative model to draft replies while retrieving documents from a verified knowledge base, a technique called retrieval-augmented generation.

Type	Response source	Strengths	Limitations
Rule-based	Fixed scripts	Predictable, low cost	Rigid, narrow scope
Intent-based	Prepared answers	Reliable within domain	Needs ongoing curation
Generative	Model-produced text	Flexible, broad coverage	Possible errors and inconsistency
Hybrid	Model plus retrieved sources	Balances accuracy and flexibility	More complex to build

Applications

• Customer service: answering common questions, tracking orders, and routing complex cases to human staff.
• Healthcare administration: appointment scheduling and symptom-information tools, which typically state that they do not replace professional advice.
• Education: tutoring, language practice, and question answering.
• Business operations: internal help desks, human-resources queries, and software coding assistance.
• Companionship and entertainment: social chatbots designed for open-ended conversation.

Reported effectiveness varies by deployment and domain, and published figures on cost savings or customer satisfaction often come from vendors, so they should be read with that context in mind.

Limitations and concerns

• Inaccurate output: generative systems can state false information confidently, a behavior often called hallucination.
• Bias: models can reflect biases present in their training data.
• Privacy: conversations may be stored or used for improvement, depending on the provider’s policies, and data-protection rules such as the EU’s General Data Protection Regulation can apply.
• Misuse: chatbots can be used to generate spam, disinformation, or phishing text.
• Emotional reliance: researchers and clinicians have debated the effects of companion chatbots on wellbeing, and evidence remains limited and mixed.
• Employment: observers disagree on how far chatbots will replace or change jobs.

Comparison with related systems

Term	Typical meaning	Relationship to chatbots
Virtual assistant	Software performing tasks (alarms, calls, smart-home control) via voice or text	Often includes a chatbot interface
Voice assistant	Assistant operated mainly by speech	A voice-based subset
AI agent	System that takes actions toward goals, sometimes autonomously	May use a chatbot as its interface
Search engine	Returns ranked links	Increasingly combined with chatbots

Common misconceptions

Chatbots do not necessarily use artificial intelligence; many rely only on fixed rules. Passing for human in a short conversation does not establish understanding or consciousness. A claim in 2014 that a chatbot named “Eugene Goostman” passed the Turing test was widely disputed because of the test conditions.

Current status

Chatbots are now common in consumer and enterprise software, and regulation, accuracy standards, and disclosure requirements are active areas of development. Readers should verify current capabilities and policies, which change quickly.