In 1966, a secretary at MIT asked her boss to leave the room so she could have a private conversation. This would have been unremarkable, except that she wasn’t talking to a person. She was typing messages to a computer program called ELIZA, and she’d become so convinced it understood her that she wanted the privacy you’d expect with a real therapist.
The program’s creator, a computer scientist named Joseph Weizenbaum, was shaken. ELIZA couldn’t understand anything. It matched patterns in text and reflected them back as questions. “I feel anxious about my writing” became “Why do you feel anxious about your writing?” It was a parlor trick, and his own secretary had fallen for it completely. The tendency for humans to project understanding and empathy onto software that produces conversational text became known as the “ELIZA effect,” and sixty years later, it’s more relevant than ever.
ELIZA was the world’s first chatbot, though the word itself wouldn’t exist for nearly three more decades.
What a Chatbot Actually Is
A chatbot is software designed to have a conversation with you through text (and sometimes voice). You type something, it responds. Back and forth, like messaging a friend, except the friend is a program.
That’s the broad definition, and it covers an enormous range. On one end, you have the bot that pops up on a customer service page and can only understand “track my order” or “talk to a human.” On the other, you have Claude and ChatGPT, which can discuss narrative structure, critique your query letter, and brainstorm plot twists for a psychological thriller set during the French Revolution. Both are chatbots. The difference is in how they generate their responses, and that difference is everything.
Where the Word Came From
Joseph Weizenbaum built ELIZA at MIT between 1964 and 1966, but he never called it a chatbot. He named it after Eliza Doolittle, the character from George Bernard Shaw’s Pygmalion who’s coached to pass as upper-class, a fitting metaphor for a program trained to pass as human. ELIZA’s most famous mode, DOCTOR, simulated a Rogerian therapist, a style of therapy where the practitioner mostly mirrors statements back to the patient. It was a clever design choice because it let simple pattern-matching feel like genuine listening.
In 1972, Stanford psychiatrist Kenneth Colby created PARRY, a program simulating someone with paranoid schizophrenia. It was far more sophisticated than ELIZA, with internal mood parameters for anger, fear, and mistrust that shaped its replies. In a test, 33 psychiatrists read transcripts of conversations with both PARRY and real patients with schizophrenia. They identified the computer correctly only 48% of the time, no better than a coin flip. And in one of computing’s more surreal moments, researchers connected PARRY and ELIZA over ARPANET (the early internet) and let them talk to each other. A therapist bot and a paranoid patient bot, chatting across a network. PARRY accused ELIZA of having “a one-track mind.”
But through all of this, nobody had a word for what these programs were.
That changed in 1994. Michael Mauldin, a computer scientist at Carnegie Mellon, had been entering conversational programs in the Loebner Prize, an annual competition inspired by Alan Turing’s famous test for machine intelligence. Mauldin’s entry was a bot named JULIA that lived in text-based virtual worlds called TinyMUDs, where she could chat with other players, explore rooms, and even play cards. In his paper for the 1994 AAAI conference, Mauldin coined the term ChatterBot, defining it as “a robot TINYMUD player whose main service is chattering.”
The name stuck. The spelling simplified. By the early 2000s, “chatbot” was the standard term for any program whose primary function was holding a conversation.
From Scripts to Something Stranger
For most of their history, chatbots were elaborate puppets. Every response was pre-written by a human, and the bot’s job was to figure out which script to pull from. ELIZA had pattern-matching templates. Mauldin’s JULIA used 224 decision nodes connected to 529 fixed responses. Even SmarterChild, the wildly popular bot that entertained 17 million users on AOL Instant Messenger in 2001, ran entirely on human-curated answers. If you asked SmarterChild for sports scores, it checked a database. If you swore at it, a hand-written “profanity handler” fired back something snarky. Every reply was chosen, never generated.
The voice assistant era (Siri in 2011, Alexa in 2014, Google Assistant in 2016) extended the chatbot concept to spoken conversation, but the underlying approach was the same: recognize what the user wants, then retrieve a pre-built response.
Then, on November 30, 2022, OpenAI released ChatGPT. It reached a million users in five days and a hundred million in two months, the fastest-growing consumer application in history. And it did something no chatbot had done before: it wrote its answers from scratch.
ChatGPT was built on a large language model, a neural network trained on vast amounts of text that generates responses by predicting the most probable next word, over and over, until it has composed a reply. It wasn’t selecting from scripts. It was producing new text that had never existed before, shaped by the patterns it had absorbed during training.
This is the dividing line in chatbot history. Before LLMs, a chatbot could only say things a human had already written for it. After LLMs, a chatbot could say things nobody had written, for better (surprisingly insightful feedback on your manuscript’s pacing) and for worse (confidently hallucinated facts that sound authoritative and are completely invented).
Why This Matters for Your Writing Life
If you’ve used ChatGPT, Claude, or Gemini, you’ve been using a chatbot. The conversational interface, you type, it responds, is the reason these tools feel intuitive rather than intimidating. You don’t need to learn a programming language or navigate a complex dashboard. You just talk.
The chatbot is the interface, not the intelligence. This is a subtle but important distinction. The large language model running behind the scenes is what understands (in its statistical way) your prompt and generates a response. The chatbot is the conversation layer wrapped around that model, the text box, the back-and-forth format, the memory of what you said three messages ago. When you switch from ChatGPT to Claude, you’re changing both the model and the chatbot interface. When a writing tool like NovelCrafter embeds a chat window that connects to the same model you could use through ChatGPT directly, the chatbot interface is different but the underlying engine might be identical.
Purpose-built chatbots outperform general ones for writing. A general chatbot like ChatGPT knows nothing about your novel until you tell it. NovelCrafter’s Codex Chat, by contrast, already knows your characters, settings, and plot points because it’s connected to your project database. When you ask it “Would Elena really do this in Chapter 12?”, it can reference her established personality traits and prior actions without you pasting in pages of context. Sudowrite’s brainstorming chat works similarly, grounding its responses in your manuscript rather than starting from zero. The chatbot format is the same in both cases. The difference is what the chatbot knows about your work before you say a word.
Conversational skills translate directly. The better you are at communicating what you want from a conversation with a human, the better you’ll be at getting useful output from a chatbot. Be specific about what you need. Provide context. If the first response isn’t quite right, refine your request instead of starting over. Treat it less like a search engine and more like a conversation with a well-read colleague who needs clear direction. The authors who get the most out of these tools aren’t the most technical. They’re the ones who are good at asking for what they want.
Joseph Weizenbaum spent the rest of his career warning people not to confuse ELIZA’s pattern-matching for real understanding. That warning still applies. But the gap between what early chatbots could do and what today’s LLM-powered chatbots can do is so vast that the same word barely covers both. The chatbot on a shoe retailer’s website and the chatbot helping you restructure a three-act novel are related only by format, not by capability. Understanding which kind you’re talking to, and what it can actually do, is the first step to using it well.