In 1966 a program named ELIZA convinced people it understood their problems. It didn't. It just turned their own sentences back into questions, so if you said you felt sad, it asked why you felt sad. Almost sixty years later we're still refining that same trick, except the machines have gotten much better at hiding the seams.
Knowing how we got here is more than trivia. The mistakes each generation of chatbot made are the same mistakes a badly set-up bot still makes on business websites right now.
The bot that pretended to listen
ELIZA was written by Joseph Weizenbaum at MIT. Its best-known script imitated a talk therapist, reflecting your words back at you. Type "I'm worried about my job" and it might respond "Why are you worried about your job?" Simple stuff. Yet Weizenbaum was unsettled by how fast people opened up to it. The story goes that his own secretary asked him to leave the room so she could talk to the program in private.
That reaction is the whole reason this history matters. People will talk to software readily, and even trust it, as long as it answers in a way that feels personal. The technology has changed enormously since then. That basic human instinct has not budged.
Rules, patterns, and a long middle age
For decades, chatbots ran on rules a person wrote out by hand. PARRY, built in 1972, simulated someone experiencing paranoia and was convincing enough that psychiatrists struggled to tell it from a real patient. Decades later, ALICE used a pattern-matching setup where designers mapped thousands of possible inputs to canned replies.
These bots could hold a decent exchange as long as you stayed on their track. Step off it and they collapsed, repeating themselves, missing the point, or looping you back to the start.
This is the era most people remember with a sigh. The phone menu that makes you press 1 for billing. The website widget that knew four questions and answered everything else with "I'm sorry, I didn't understand that." The bot wasn't stupid, exactly. It only knew what someone had thought to script ahead of time, and nobody can script every question a real customer will ask.
Businesses felt the other half of that pain too. Every new product, policy, or promotion meant someone had to go back in and write more rules by hand. The bot was never really finished, and the moment maintenance slipped, it started giving stale or wrong answers. Grow the number of possible questions and the whole approach buckles under its own upkeep.
When bots learned from data instead of rules
The next real shift moved the work away from hand-written rules. Instead of a person mapping every input to an output, statistical models and then neural networks learned patterns from large amounts of text. The bot started to generalize, handling a phrasing it had never seen exactly because it had seen thousands of similar ones.
This made replies feel less brittle. It also introduced a problem that's still with us: once a system guesses based on patterns rather than fixed rules, it can guess wrong with total confidence. A rule-based bot fails obviously. A learning bot can fail smoothly, which is much harder to catch.
This matters for anyone putting a bot on a business site. A smooth, wrong answer about your shipping policy or your prices doesn't just fail to help. It actively misleads a customer, and they'll hold you responsible for it, not the software. The lesson from this era is that fluency and accuracy are two different things, and a bot can have plenty of the first with none of the second.
The language-model era, and the fix for its worst habit
In the early 2020s, large language models changed what a chatbot could do the moment you switched it on. They wrote in a natural voice, tracked context across several messages, and answered questions nobody had explicitly programmed. For the first time a bot could take "do you ship to Canada, and how long does it take around the holidays" as one fluid thought.
The catch is that a raw language model will invent an answer when it doesn't have one. Ask about your return window and it might state a confident, completely made-up "30 days" because that sounds right for a store like yours.
The practical fix is retrieval. Rather than trusting only what the model absorbed during training, a business bot pulls answers from a specific source you control: your help docs, product pages, and policies. Now when someone asks about returns, the bot quotes your real 14-day window because it's reading your actual page. This is the approach modern tools take, SpideyChat included. You train the bot on your own content, and it answers from that instead of guessing.
| Era | Roughly | How it worked | Where it broke |
|---|---|---|---|
| ELIZA / PARRY | 1960s–70s | Reflected and matched simple patterns | No real understanding |
| Scripted bots | 1990s–2010s | Hand-written rules and menus | Fell apart off-script |
| Learning models | 2010s | Learned patterns from data | Confidently wrong |
| LLM + retrieval | 2020s | Generates language, answers from your docs | Needs good source content |
What sixty years of bots means for your storefront
Consider Maple & Thread, a small online fabric shop run by two people. A few years back their only option was a scripted widget with a dozen canned answers. It handled "where's my order" and nothing else, so most chats ended with a customer emailing anyway. The bot made work instead of saving it.
With a retrieval-based setup, the same shop trains a bot on its shipping page, fabric care guides, and return policy. Now a customer asking "will this linen shrink if I wash it warm" gets the answer straight from the care guide, at 11pm, without sending an email. When the question is genuinely outside the docs, like "can you custom-dye 40 yards," the bot passes it to a human instead of bluffing.
The two owners noticed something else. Because the bot answered the routine questions overnight, the emails waiting each morning were the interesting ones, the custom orders and the wholesale inquiries, not the tenth "do you ship to the UK" of the week. The history of chatbots is partly a history of lifting that boring, repetitive layer off human shoulders, and only recently has the technology been good enough to do it without annoying the customer in the process.
That handoff is the lesson ELIZA taught, applied properly. The goal was never to fake understanding. It's to be genuinely useful on the questions you can answer and honest about the ones you can't.
Where this leaves you
The technology finally caught up to an expectation people have carried since 1966: talk to it like a person and get a real answer back. The difference between a bot that helps and one that annoys isn't the model anymore. It's whether you've given it good source material and a clean way to pass hard questions to a human. Get those two things right and the long arc of chatbot history ends, for your customers, in something that simply works.