History of Chatbots: From ELIZA to AI Agents

In 1966 Joseph Weizenbaum at MIT created ELIZA, a program that simulated a Rogerian psychotherapist by reflecting users' statements back as questions using simple pattern matching. It understood nothing — yet users became emotionally attached and insisted it truly understood them, a phenomenon Weizenbaum found deeply disturbing.

The 'ELIZA effect' — the human tendency to attribute understanding to machines that converse — became the founding observation of chatbot history. It presaged modern debates about AI companions and chatbot attachment by nearly six decades.

In 1968 Terry Winograd's SHRDLU could hold a genuine dialogue about a simulated world of blocks — moving objects on command, answering questions, resolving pronouns. Within its tiny domain it was startlingly capable, and it convinced many researchers that scaled-up rule-based language understanding was within reach.

It wasn't. SHRDLU's competence collapsed outside the blocks world, and the gap it exposed — between narrow scripted dialogue and open-ended understanding — defined why chatbots stagnated for the next four decades.

Apple's Siri (2011) and Amazon's Alexa (2014) put conversational interfaces in hundreds of millions of pockets and homes. They normalized talking to machines and built the always-listening hardware ecosystem — but under the hood they remained mostly intent-classification systems routing to scripted skills, and users quickly learned their rigid limits.

The voice assistants proved massive demand for conversational computing existed years before the technology could satisfy it. That unmet expectation is part of why ChatGPT's fluency landed with such force.

On November 30, 2022, OpenAI released ChatGPT, a conversational interface over GPT-3.5 fine-tuned with human feedback. It reached 100 million users in two months — the fastest-growing consumer application in history — because for the first time a chatbot could actually hold up its end of nearly any conversation.

Anthropic's Claude followed in March 2023, trained with Constitutional AI to follow explicit principles. Within a year the chatbot went from a punchline to the primary interface through which the public experiences artificial intelligence.

GPT-4o (2024) made conversation multimodal and near-instant — voice, vision, and text in real time. But the deeper shift was from conversation to action: by 2025-2026, systems like ChatGPT Work carried out whole jobs across applications and files rather than merely answering questions.

The trajectory that began with ELIZA reflecting statements back now points at conversational systems that observe, plan, and act. The chatbot was never really the destination — it was the interface through which AI agents arrived.