--- title: "Chatbot: definition and example – Carl-Uwe Würfel" description: "Chatbot explained: how chatbots evolved from ELIZA to language models, how they differ from AI agents, and what disclosure the EU AI Act requires." url: https://carl-wuerfel.de/en/glossary/chatbot author: Carl-Uwe Würfel --- 1. [Home](https://carl-wuerfel.de/en) 2. [Knowledge](https://carl-wuerfel.de/en/knowledge) 3. [Glossary](https://carl-wuerfel.de/en/glossary) 4. Chatbot # Chatbot A chatbot is a program that takes input in natural language, typed or spoken, and responds in a dialogue, either by following fixed rules or by using a language model. ## On this page 1. How a chatbot works 2. Three designs emerged one after another 3. Chatbots, AI assistants, and AI agents are different things 4. What the EU AI Act requires of chatbots 5. The company is liable for the answers 6. What good implementation looks like 7. Where the term stops working 8. Frequently asked questions 9. Sources A chatbot knows only what its rules and sources provide. _Figure: A loop using the example of a municipal utility that runs a chatbot on its website for questions about monthly installments, meter readings and moving. The chatbot states at the outset that an AI system is responding. It draws only on the approved rate and service pages and cites the source with each answer. Questions without a matching source go on a list the team uses to expand the service pages. Questions about a specific bill go to customer service with the conversation history._ **Example** A municipal utility runs a chatbot on its website that answers questions about monthly installments, meter readings, and moving. It draws only on the approved rate and service pages, cites the source with each answer, and states at the outset that an AI system is responding. Questions about a specific bill go to customer service with the conversation history; questions without a matching source go on a list the team uses to expand the service pages. A chatbot holds a conversation: a person types or speaks, the program answers, and both sides can build on what was said before. The word combines “chat,” casual talk in a text window, with “bot,” a program that runs on its own. A chatbot can match patterns, map input to predefined requests and return stored answers, or write every answer from scratch with a [large language model](https://carl-wuerfel.de/en/glossary/llm) (LLM). In 1966, Joseph Weizenbaum described ELIZA, a program that answered input using keywords and transformation rules. Later platforms used [machine learning](https://carl-wuerfel.de/en/glossary/machine-learning) to recognize the request behind a message, then played a scripted dialogue. Since OpenAI released ChatGPT in November 2022, most people use “chatbot” to mean an LLM chatbot that writes its own answers. For executives, a chatbot speaks to customers or employees on the company’s behalf, and what it says reads like a company statement. Since August 2, 2026, Article 50(1) of the EU [AI Act](https://carl-wuerfel.de/en/ai-act) also requires that people be told when they are interacting with an AI system, unless that is obvious. Answer quality, ownership, and disclosure belong on the leadership agenda. ## How a chatbot works. The chatbot takes the input, typed or spoken and converted to text, then works out what the person wants. If needed, it pulls data from connected sources such as a knowledge base or customer account. Then it answers and keeps the history, so a follow-up like “And on Saturday?” makes sense. In an LLM chatbot, a fixed instruction placed before every conversation, the [system prompt](https://carl-wuerfel.de/en/glossary/system-prompt), sets the role, tone, and limits. For company-specific knowledge, many implementations use [retrieval-augmented generation](https://carl-wuerfel.de/en/glossary/rag) (RAG): the system searches approved documents for relevant passages and hands them to the model along with the question. The model then writes an answer that, ideally, points back to those passages. Production adds two components. A handoff to a human kicks in when the chatbot is stuck or the person asks for one. A log records which questions came in and which answers went out. Without it, nobody spots errors or sees which sources are missing or outdated. ## Three designs emerged one after another. In his 1950 paper “Computing Machinery and Intelligence,” Alan Turing proposed an imitation game in which a person, in written conversation, tries to tell whether the other party is a machine. The first design matches patterns. ELIZA, published by Joseph Weizenbaum in 1966, scanned input for keywords and built a reply from them according to rules. Its best-known script imitated a psychotherapist. Weizenbaum noted at the time that ELIZA showed how easy it is to create and maintain the illusion of understanding. The second design maps input to intents. Google’s Dialogflow, for example, defines an intent as the user’s goal for one conversational turn: the team enters example phrases, a machine learning model also catches other wording, and answers and flows are fixed in advance. These chatbots are predictable but fail at anything nobody built an intent for. The third design writes freely with a language model. When OpenAI launched ChatGPT in 2022, it stated that the system sometimes writes plausible-sounding but incorrect or nonsensical answers. Many current solutions combine both: a language model for open questions, fixed flows for sensitive steps such as cancellations or payments. ## Chatbots, AI assistants, and AI agents are different things. All three often sit behind a chat window. What separates them is what they do. A chatbot answers. Its output is information in a dialogue, and its main risk is a wrong answer. An AI agent pursues a goal across several steps, calls tools on its own, and changes data in other systems, such as an appointment, an order, or a CRM record. Its main risk is a wrong action. “[AI assistant](https://carl-wuerfel.de/en/glossary/ki-assistent)” has no fixed definition. Vendors usually mean a system that helps a person write, summarize, or search, often inside existing software. The lines are blurry. The language model itself is not a chatbot either. The model is the engine; the chatbot is the application around it, with interface, instructions, sources, rules, and logging. One question settles most cases: what can the system change? Once a chatbot books appointments, modifies contracts, or issues credits, it behaves like an agent and needs tightly scoped permissions and approvals. ## What the EU AI Act requires of chatbots. Article 50(1) of the EU AI Act, Regulation (EU) 2024/1689, requires providers to design AI systems intended to interact directly with natural persons so that those persons are informed they are interacting with an AI system. The obligation does not apply where this is obvious to a reasonably well-informed, observant, and circumspect person, given the context. Under Article 50(5), the information must be clear and distinguishable, provided at the latest at the first interaction, and meet accessibility requirements. Article 50 has applied since August 2, 2026, the general application date under Article 113. Amending Regulation (EU) 2026/1744 added a transition period only for the marking obligation in paragraph 2; paragraph 1 was left unchanged (as of November 2026). The European Commission reads the exception narrowly in its Q&A on Article 50. The obligation falls on the provider. Under Article 3(3), a company that has an agency build a chatbot and puts it into service under its own name can be the provider. Violations of Article 50 can draw fines of up to 15 million euros or 3 percent of worldwide annual turnover under Article 99. Individual cases are for legal counsel. ## The company is liable for the answers. A Canadian case shows that a chatbot is no separate legal actor. In February 2024, the Civil Resolution Tribunal of British Columbia decided Moffatt v. Air Canada. The airline’s chatbot had told a customer he could apply for a reduced bereavement fare after booking. Another page on the website said the opposite. Air Canada argued it was not liable for information from its agents or representatives, including a chatbot. The tribunal called this a remarkable submission: the chatbot was part of the website, and the airline was responsible for all the information on its website, whether it came from a static page or a chatbot. Air Canada had to refund the fare difference. The ruling rests on Canadian law and binds no court elsewhere. Its logic is hard to argue with: customers may rely on what a company’s website tells them. In operation, the chatbot therefore answers only from approved sources and says so when it lacks an answer. Prices, contract terms, and legal or medical statements are either fixed in advance or routed to people. Pre-launch tests include users trying to coax the chatbot into commitments. ## What good implementation looks like. Good implementation starts with a narrowly defined purpose: which questions the chatbot answers and which it does not. Behind it sit reviewed sources that someone keeps current. A clear notice at the start says that an AI system is answering. A handoff to a human is always available, and the history goes along so nobody has to repeat themselves. Before launch and after every change to model, instructions, or sources, a fixed set of real questions with known answers runs through the system. In production, questions without a good answer are collected and reviewed regularly; they show which content is missing. The data flow is documented: what the system stores, where it goes, and how long logs are kept. Data protection law applies alongside the AI Act. On cost: a chatbot with fixed flows costs little per conversation but takes effort to set up; an LLM chatbot costs money per call, depending on model and text volume, and needs ongoing source upkeep. If the model can be swapped later, the company keeps leverage on price and quality. ## Where the term stops working A chatbot is a conversational interface and knows only what its rules and sources provide; fluent answers prove neither accuracy nor understanding. It differs from the language model that powers many chatbots, from the AI agent that takes actions in other systems, and from search, which returns results instead of a written answer. “AI assistant” has no fixed definition and overlaps with both chatbot and agent. ## Frequently asked questions What is a chatbot? A chatbot is a program that takes questions in natural language, typed or spoken, and answers them in a dialogue. Older chatbots follow fixed rules and stored answers, while newer ones write their answers with a large language model. What is the difference between a chatbot and an AI chatbot? The difference is the design. A traditional chatbot maps input to predefined requests and returns stored answers, while an AI chatbot writes each answer with a language model. The traditional kind is more predictable; the AI chatbot handles open-ended questions but can make up wrong answers. What is the difference between a chatbot and an AI agent? A chatbot answers, an AI agent acts. The agent pursues a goal across several steps, calls tools, and changes data in other systems. That is why an agent needs tightly scoped permissions and human approval for consequential steps. Does a chatbot have to disclose that it is AI? Yes, under Article 50(1) of the EU AI Act, people must be told they are interacting with an AI system unless that is obvious. The obligation has applied since August 2, 2026, and falls on the provider of the system. The notice must be clear at the latest at the first interaction. Is a company liable for wrong answers from its chatbot? As a rule, the company answers for what its chatbot says. In the 2024 Canadian case Moffatt v. Air Canada, a tribunal rejected the argument that the chatbot was responsible for its own statements. Elsewhere, it depends on local law and the customer relationship. How much does a chatbot cost? There is no reliable flat price. Cost depends on the design, the number of conversations, the chosen model, and the effort to connect and maintain knowledge sources. LLM chatbots cost money per call; rule-based ones mainly take setup effort. What was the first chatbot? ELIZA, developed by Joseph Weizenbaum at MIT and described in 1966, is generally considered the first well-known chatbot. It used keywords and transformation rules without understanding the content. Its best-known script imitated a psychotherapist. ## Further reading * [**AI glossary**](https://carl-wuerfel.de/en/glossary) * [**AI Act: classifying how you use AI**](https://carl-wuerfel.de/en/ai-act) * [**AI labeling and disclosure**](https://carl-wuerfel.de/en/glossary/ki-kennzeichnung) * [**AI agents**](https://carl-wuerfel.de/en/glossary/ki-agenten) * [**Hallucination**](https://carl-wuerfel.de/en/glossary/hallucination) * [**LLM: Large Language Model** Term](https://carl-wuerfel.de/en/glossary/llm) * [**RAG: Retrieval-Augmented Generation** Term](https://carl-wuerfel.de/en/glossary/rag) ## Sources 1. [Turing: Computing Machinery and Intelligence, Mind 59 (236), 1950](https://doi.org/10.1093/mind/LIX.236.433) 2. [Weizenbaum: ELIZA, a computer program for the study of natural language communication between man and machine, Communications of the ACM 9 (1), 1966](https://doi.org/10.1145/365153.365168) 3. [Google Cloud: Dialogflow ES, Intents, documentation, as of November 2026](https://docs.cloud.google.com/dialogflow/es/docs/intents-overview) 4. [OpenAI: Introducing ChatGPT, November 30, 2022](https://openai.com/index/chatgpt/) 5. [European Union: Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 3, 50, 99, and 113, 2024](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689) 6. [European Union: Regulation (EU) 2026/1744 amending the AI Act (Digital Omnibus on AI), 2026](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng) 7. [European Commission: Transparency obligations under Article 50 of the AI Act, Q&A, updated July 24, 2026](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act) 8. [Civil Resolution Tribunal of British Columbia: Moffatt v. Air Canada, 2024 BCCRT 149, February 14, 2024](https://decisions.civilresolutionbc.ca/crt/crtd/en/item/525448/index.do)