A conversational interface lets a person ask for something in ordinary language, typed or spoken, and get a reply in the same form. Two designs sit behind almost every one: fixed rules that match keywords, as in the 1966 program ELIZA, and statistical language models built on the Transformer architecture published in 2017, which power chatbots such as ChatGPT (released November 30, 2022). Both can sound as if they understand, and the operator of the interface stays responsible for what it says.
Key takeaways
- ELIZA, described by Joseph Weizenbaum in January 1966, chose replies by spotting key words in the input and filling in a template.
- Modern chatbots generate text from a trained model instead of a script, which makes them flexible but able to produce fluent, wrong answers.
- A person's sense of being understood comes largely from the person, a point Weizenbaum made in 1966 and that still holds.
- In Moffatt v. Air Canada (February 14, 2024), a British Columbia tribunal rejected the argument that an airline was not liable for what its website chatbot said.
Two ways to build a conversation
The table sets the two designs side by side. It simplifies, because real products often combine them: a bank might use a language model for free-form questions and fixed rules for anything involving a payment.
| Question | Rule-based design | Language-model design |
|---|---|---|
| How is a reply chosen? | A rule fires when a keyword or pattern appears, and a template builds the reply | A trained model produces the reply one piece of text at a time, based on the conversation so far |
| Historic example | ELIZA, 1966 | ChatGPT, introduced November 30, 2022 |
| Typical strength | Predictable; every possible answer was written by a person | Handles wording that no designer anticipated |
| Typical failure | Gives a stock or irrelevant answer when no rule matches | Gives a confident answer that is false |
| Who fixes a mistake? | A person edits the rule | The developer retrains, adjusts, or adds safeguards around the model |
ELIZA: keyword rules in 1966
Weizenbaum, then at the Massachusetts Institute of Technology, described ELIZA in a paper in the January 1966 issue of Communications of the ACM. Its input sentences were analyzed with decomposition rules triggered by key words, and replies were built with reassembly rules attached to those decompositions. The paper lists the technical problems the program had to solve, among them finding the key word, finding minimal context around it, and producing a reply when no key word appears.
The best-known script made ELIZA respond roughly as a Rogerian psychotherapist would. Weizenbaum chose that setting because the program needed no knowledge of the world to keep the exchange going. In the paper's sample, a typed remark that a boyfriend had made the speaker come to the session produced "YOUR BOYFRIEND MADE YOU COME HERE." The program had swapped pronouns and handed the words back.
The paper is also frank about the limits. Weizenbaum wrote that the human speaker supplies much of the plausibility and will not defend the illusion of being understood against all odds. The program was a demonstration of how little machinery could create that impression, not a claim of comprehension.
A rule, shown step by step
An illustrative example in the style of ELIZA, not a rule copied from the original script. Suppose one rule says: if the input contains the word "mother," reply "Tell me more about your family." The input "The mother called yesterday" matches, so the reply appears. A second rule says: if the input contains "always," reply "Can you think of a specific example?" The program has not represented what a mother or a habit is. It matched a word and returned a template. When the input is "Nothing works," no rule matches, and the program falls back on a neutral stock phrase, one of the cases the 1966 paper lists as a design problem.
What changed after 2017
In the paper "Attention Is All You Need," first posted to arXiv on June 12, 2017, eight authors introduced the Transformer, a design that relies on attention and drops recurrence and convolutions. It was tested on machine translation, where it scored 28.4 BLEU on the WMT 2014 English-to-German task, more than two points above the best earlier results. Later language models for dialogue built on the same family of architectures.
OpenAI introduced ChatGPT on November 30, 2022, as a free research preview. Its announcement says the model was fine-tuned from a GPT-3.5 series model and trained with reinforcement learning from human feedback, using conversations written by human trainers. The same post lists limitations: the model sometimes writes plausible-sounding but incorrect answers, and it can be sensitive to small changes in how a question is phrased. In other words, the shift from scripts to learned models widened what an interface can handle without removing the need to check what it says.
Who answers for the chatbot
In November 2022, after a death in the family, Jake Moffatt booked a flight with Air Canada and asked the airline's website chatbot about bereavement fares. The chatbot suggested a fare could be requested after travel. Air Canada's policy did not allow that, as Moffatt learned from airline staff.
The decision of the Civil Resolution Tribunal of British Columbia, issued February 14, 2024, treated the dispute as negligent misrepresentation. Air Canada argued it could not be held liable for information from its chatbot, which in effect treated the bot as a separate legal entity. The tribunal member called that submission remarkable, noting that the chatbot was part of the airline's website, and found the airline had not taken reasonable care to keep it accurate. Air Canada was ordered to pay a total of CAD 812.02: CAD 650.88 in damages, CAD 36.14 in interest, and CAD 125 in tribunal fees.
This was one small-claims decision. It shows how a tribunal reasoned about a specific website, not a settled rule for every chatbot case, and it is not legal advice.
Checklist before relying on a chat interface
- Does the interface say plainly that it is automated, and does it give a way to reach a person?
- Are answers about money, health, or legal rights confirmed against an official page or a staff member?
- Does the answer point to a source that can be opened and read, or only assert a result?
- Is the question phrased the same way twice, and does the answer stay consistent?
- For an organization deploying one: who reviews the transcripts, and who corrects wrong answers?
FAQ
Did ELIZA understand what people typed?
No. According to Weizenbaum's 1966 paper, it matched key words, applied decomposition and reassembly rules, and produced a reply from a script. Its apparent understanding came from how people read the output.
Is a modern chatbot looking up answers in a database?
Not in the usual sense. A language model generates text from patterns learned in training. OpenAI's own announcement of ChatGPT notes that it can write plausible-sounding but incorrect answers, so a fluent reply is not evidence that the facts are right.
Is a voice assistant different from a text chatbot?
A voice assistant adds speech recognition at the front and speech synthesis at the back. The step in the middle, working out what to answer, can use either design described above.
Can a company in Canada be held responsible for its chatbot's mistakes?
In one 2024 British Columbia tribunal decision, yes, for a misleading answer on the airline's own website. How other courts and tribunals treat similar cases is not settled by that single ruling.
Related reading
- AI and human-computer interaction
- Replika as a companion chatbot
- A timeline of AI history
- Computer vision in AI
More in AI Basics.
Last verified: October 2026
Sources
- Weizenbaum J., ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine, Communications of the ACM, Volume 9, Number 1, January 1966
- Vaswani A. et al., Attention Is All You Need, arXiv:1706.03762, June 12, 2017
- Introducing ChatGPT, OpenAI, November 30, 2022
- Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal of British Columbia, February 14, 2024