Chatbots trace a genuinely direct lineage from 1966’s ELIZA to today’s LLMs — six decades of incremental progress worth understanding, since the same core question (can a machine convincingly converse?) has driven the field throughout.
ELIZA (1966): Pattern Matching as Illusion
Joseph Weizenbaum’s ELIZA simulated a Rogerian psychotherapist using genuinely simple pattern matching and scripted response templates — no actual understanding of language whatsoever, yet users famously formed genuine emotional attachments to it, a phenomenon now called the “ELIZA effect”: humans readily attributing real understanding to systems that are, underneath, following comparatively simple rules.
Rule-Based Systems (1970s-1990s)
Chatbots through this era remained fundamentally pattern-matching and scripted-response systems, gradually growing more sophisticated but never achieving genuine language understanding — ALICE (1995) won early chatbot competitions using an elaborate but still fundamentally rule-based approach, not real comprehension.
Statistical and Retrieval-Based Systems (2000s-2010s)
Early customer service chatbots and virtual assistants (Siri’s initial 2011 launch) primarily used intent classification and retrieval — matching a user’s genuine query against a predefined, curated set of known questions and canned responses, still fundamentally limited to genuinely anticipated inputs.
The LLM Era (2018-Present)
Transformer-based language models trained on truly massive amounts of text finally enabled genuinely open-ended, coherent conversation — not through explicit, hand-coded rules at all, but by learning statistical patterns of language at a scale and depth previous approaches genuinely couldn’t achieve, culminating in ChatGPT’s late-2022 mainstream breakthrough.