1. Why AI can feel like it appeared out of nowhere
For most people, AI seemed to arrive suddenly.
One year, chatbots barely came up in conversation. The next, ChatGPT was being used to write emails, plan meals, and explain homework.
That sense of suddenness makes sense.
A product that most people had never touched became something millions of people used within weeks of launch.
But the history of AI is much longer than that.
The idea of a machine that could think, talk, or reason has been discussed, written about, and worked on for decades, long before any computer existed that could attempt it.
This guide traces that longer story.
It starts with the myths and science fiction that shaped what people expect AI to be, moves through the academic question that gave the field its name, covers the periods when AI research nearly stalled for good, and follows the breakthroughs that eventually led to the chatbots people use today.
For a plain-English answer to what artificial intelligence actually is today, see What Is Artificial Intelligence? The Complete Plain-English Guide.
Seen this way, ChatGPT is not really a sudden invention.
It is the latest chapter in a story that started with a wartime code-breaker asking a simple question: can machines think?
2. Long before computers: AI in myth and fiction
Mechanical men and ancient myths of artificial beings
The idea of an artificial being is far older than computing itself.
Ancient Greek myths described bronze figures brought to life to guard or serve their creators.
Jewish folklore tells of the golem, a being shaped from clay and animated to protect a community.
These stories were not about technology in any modern sense.
They were about a much older question: what would it mean to build something that acts like it is alive.
That question has never really gone away. It just changed shape as new tools became available to ask it with.
Science fiction's AI: from Frankenstein to HAL 9000 and Data
Mary Shelley wrote *Frankenstein* at eighteen, during a ghost-story writing contest with Lord Byron one wet summer by Lake Geneva. Published in 1818, it gave the idea a modern shape: Victor Frankenstein stitches a creature together from dead tissue, then bolts in horror the moment it opens its eyes.
That basic shape, creation followed by consequence, still runs through most fictional AI stories today.
Science fiction kept building on it through the 20th century.
HAL 9000, the ship's computer in Arthur C. Clarke and Stanley Kubrick's *2001: A Space Odyssey*, gave audiences a calm, capable AI that turns dangerous once its goals stop matching its creators'.
The Terminator brought a blunter version of the same fear: an AI that turns against humanity outright.
Not every fictional AI was a threat.
Data, the android from *Star Trek: The Next Generation*, offered a gentler version: a character who wanted to understand human behaviour rather than replace or destroy it.
Why fiction shaped what people expect AI to be
These stories matter because they arrived long before real AI did, and they set expectations that real AI has been measured against ever since.
Many people's first mental image of AI is not a spam filter or a recommendation system.
It is a talking robot, a rebellious computer, or a machine with its own hidden agenda.
That gap between fictional AI and real AI explains a lot of the confusion people still have today.
The AI in most homes and workplaces looks nothing like HAL or Data. It looks like a chat window that answers questions and sometimes gets things wrong.
3. The 1950s: a wartime code-breaker asks "can machines think?"
Alan Turing and the Turing Test (1950)
The academic story of AI usually starts with Alan Turing.
He was a British mathematician, best known at the time for breaking German codes during the Second World War.
In 1950, Turing published a paper called "Computing Machinery and Intelligence."
In it, he asked a deceptively simple question: can machines think (Britannica)?
Rather than trying to define "thinking" outright, Turing proposed a practical test instead.
A machine could be judged intelligent if its conversation was indistinguishable from a human's, at least to a person trying to tell the difference. This became known as the Turing Test.
The test did not attempt to explain what was happening inside the machine.
It judged intelligence purely by outcome.
Could the machine hold a conversation well enough that a person couldn't tell they were talking to something artificial?
People still judge AI mostly this way today: by what it produces, not by what it actually understands.
The 1956 Dartmouth workshop that named the field
The term "artificial intelligence" did not exist until a small group of researchers gave it a name.
In 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon wrote a proposal for a summer research project at Dartmouth College, which ran the following year.
That proposal coined the term "artificial intelligence" and is widely regarded as the founding moment of AI as an academic field (Britannica).
The workshop brought together a small number of researchers who believed that human learning and intelligence could, in principle, be described precisely enough that a machine could be made to simulate it.
That belief did more than any single invention to turn AI from a philosophical question into an actual research field. It gave the idea a name, a community of researchers, and money behind it.
4. Early optimism, an early chatbot, and the AI winters
ELIZA: the 1966 program that fooled people into thinking it understood them
In 1966, Joseph Weizenbaum, a computer scientist at MIT, built a program called ELIZA.
It is widely regarded as the first chatbot (MIT CSAIL; Smithsonian Magazine).
ELIZA simulated a Rogerian psychotherapist, a style of therapy that involves reflecting a patient's own statements back as questions.
If someone typed "I am feeling anxious about work," ELIZA might respond with something like "Why do you say you are feeling anxious about work?"
The program worked through pattern matching.
It looked for keywords in what someone typed and slotted them into pre-written response templates. It was not reasoning about what was said and had no understanding of the conversation at all.
Even so, some people using it responded as though they were talking to something that understood them, sharing personal feelings and engaging with it as they would with a real therapist.
That reaction, people attributing understanding to a system that had none, has followed chatbots ever since.
Why researchers thought human-level AI was decades away, not centuries
Early AI researchers were strikingly confident.
Progress in the 1950s and 1960s on tasks like game-playing and simple problem-solving convinced many in the field that human-level machine intelligence was close, a matter of years or a couple of decades, not centuries.
That optimism helped attract funding and talent into the field early on. It also set expectations that the technology of the time could not come close to meeting.
The AI winters: why funding and interest collapsed in the 1970s and 1990s
When the promised breakthroughs did not arrive on schedule, the reaction was sharp.
Funding and interest in AI dropped sharply from the 1970s into the 1980s, and again in the 1990s, a pattern now commonly called the "AI winter" (Britannica).
Each winter followed a similar shape.
Early results created excitement and ambitious predictions.
Those predictions ran ahead of what the computing power and methods of the time could actually deliver.
Funders and institutions lost patience, and money and attention moved elsewhere.
These were not brief dips.
Entire research groups shrank or closed, and "artificial intelligence" became a phrase some researchers avoided using in funding proposals, because it had come to sound overhyped.
The field did not disappear during these periods.
Progress simply slowed and moved further from public attention, waiting for the ideas, data, and computing power to catch up with the ambition.
5. AI proves itself at narrow tasks
Deep Blue beats world chess champion Garry Kasparov (1997)
In 1997, IBM's Deep Blue defeated reigning world chess champion Garry Kasparov (Britannica).
It was the first time a computer had beaten a sitting world champion in a full match under standard tournament conditions.
The result made headlines everywhere. For many people, it was the first time AI felt real rather than theoretical.
Deep Blue did not play chess the way a human does.
It worked by searching through millions of possible moves per second, calculating outcomes far ahead, and picking the strongest option.
This is often called brute-force search: raw computing power applied to a clearly defined problem with fixed rules.
Chess is a good match for this approach.
The board is small, the rules never change, and every move can be checked against every other.
That made it possible to solve with enough processing power, long before anything resembling general intelligence existed.
Why beating a chess champion did not mean AI could think
Deep Blue could not hold a conversation.
It could not recognise a face, plan a birthday party, or explain why it had made a particular move.
It was built for one task, and it did that one task extremely well.
This is the difference between narrow AI and general intelligence.
Narrow AI performs a specific, well-defined task, often better than any human.
General intelligence would mean reasoning, learning, and adapting across many different situations, the way a person does. See Different Types of AI Explained for Beginners for more on how these categories work.
Deep Blue was narrow AI. So was everything that came immediately after it for years.
Beating a grandmaster proved that a machine could outperform a human at a specific, rule-bound task.
It did not prove a machine could think, understand language, or reason about the world the way ELIZA's users once believed a chatbot could.
Fifteen years passed before the next real leap arrived.
6. The 2012 breakthrough that quietly changed everything
What deep learning actually is, in plain terms
Deep learning is a way of training a computer system to improve at a task by studying huge numbers of examples, rather than following rules a person has written by hand.
A traditional program follows instructions step by step: if this happens, do that.
A deep learning system works differently. It is shown large amounts of example data, such as thousands of photographs labelled "cat" or "not cat," and it gradually adjusts itself to get better at telling them apart.
For a fuller, plain-English walkthrough of how this works inside a modern AI tool, see How Does AI Actually Work? A Beginner's Explanation Without the Jargon.
The word "deep" refers to the layered structure inside the system, loosely inspired by how neurons in a brain are connected.
Each layer picks up on slightly more complex patterns than the one before it.
Early layers might notice edges and shapes. Later layers combine those into features like ears, whiskers, or fur texture.
Nobody writes rules for what a cat looks like. The system works it out from examples, through repeated adjustment.
The image recognition result that reignited the field
In 2012, a deep learning system produced a breakthrough result in image recognition, a turning point that helped trigger the current wave of AI investment (Britannica).
Before this point, deep learning was a known idea, but not one that had proven itself outside research papers.
The 2012 result changed that.
It showed the approach could outperform older techniques by a meaningful margin on a task researchers had struggled with for years: recognising what a photograph actually shows.
The effect on the field was immediate.
Researchers who had been sceptical of deep learning started adopting it.
Funding that had been hard to find during the leaner years after the AI winters began flowing back in.
This is the point where the modern AI boom actually starts, even though almost nobody outside research circles noticed it happening.
There was no product launch, no headline event most people would have seen. Just a research result that told the field it had found something that worked.
7. The invention that made today's chatbots possible
The 2017 paper that introduced the transformer
In June 2017, eight researchers at Google, Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin, published a paper called "Attention Is All You Need" (arXiv).
The paper introduced a new kind of neural network design called the transformer.
Earlier language systems had processed text mostly in order, one word after another, which made them slow and made it hard for them to keep track of information across long passages.
The transformer worked differently.
It used a mechanism called attention, which let the system weigh up how every word in a passage relates to every other word at the same time, rather than working through it in strict sequence.
This turned out to be both faster to train and far better at handling long stretches of text.
At the time, this was a research paper read mostly by other researchers.
It did not mention chatbots, and nobody reading it in 2017 was promised a tool that could hold a conversation.
Why this mattered more than any single chatbot
The transformer is the direct technical link between the "old AI" of chess-playing programs and expert systems, and the "new AI" of chatbots people use every day.
Virtually every major AI language model built since, including GPT, Claude, and Gemini, is built on the transformer design introduced in that 2017 paper (arXiv).
That is a five-year gap between the invention and anything the public could use.
Deep Blue's brute-force search could not have scaled into a chatbot. Early deep learning image recognition could not have scaled into one either.
The transformer was the piece that made it possible to train a system on enormous amounts of text and have it generate language that reads as fluent and coherent.
This is why a short history of AI has to pause on a paper almost no ordinary person has heard of.
Without it, the chatbots covered in the next section would not exist in their current form.
8. From research labs to everyday chatbots
ChatGPT's 2022 launch and why it went mainstream so fast
OpenAI released ChatGPT as a free research preview on 30 November 2022, describing it simply as "a model called ChatGPT which interacts in a conversational way" (OpenAI).
It reached 1 million users within five days of launch (OpenAI).
Nothing about the underlying technology was brand new.
The transformer had existed for five years by that point, and large language models had already been trained inside research labs for a while before this.
What changed was access.
ChatGPT put a transformer-based model behind a simple text box that anyone could open and type into, free, with no technical setup required.
People who had never heard of deep learning or the transformer paper could suddenly ask a computer a question in plain language and get a fluent, useful answer back. See Everything You Need to Know About ChatGPT for more on how it works today.
That combination, genuine capability plus a free and simple interface, is what turned a research advance into a mainstream event almost overnight.
Claude, Gemini, and the rest of today's chatbot landscape
ChatGPT was not the only chatbot to follow the transformer breakthrough, and it was not alone for long.
Anthropic launched Claude, alongside a faster version called Claude Instant, through its API on 14 March 2023, after testing with partners including Notion, Quora, and DuckDuckGo (Anthropic). See Everything You Need to Know About Claude for a fuller look at how it compares.
Google announced Gemini 1.0 on 6 December 2023, its first natively multimodal model family, released in three sizes: Ultra, Pro, and Nano (CNBC). See Everything You Need to Know About Gemini for more.
Other companies followed with their own chatbots built on the same transformer foundations.
Within roughly a year of ChatGPT's launch, what had been a research curiosity understood by a small community of specialists had become a crowded, competitive market with several major players, all descended from the same 2017 paper. For a direct comparison of the two biggest names, see ChatGPT vs Claude: Which Is Better for Everyday Use?.
9. What science fiction got right, and what it got wrong
Science fiction spent decades imagining artificial intelligence before it actually existed.
Some of that imagining turned out to be surprisingly close to what showed up. A lot of it did not.
Comparing the two is not just a fun exercise.
It helps explain why so many people meet a real AI chatbot with expectations that were shaped by a film rather than by the technology itself. Some of those mismatched expectations are exactly what Common AI Myths Explained works through in more depth.
The parts of the sci-fi version that came true
Fiction imagined machines that could hold a natural conversation, and that part has largely arrived.
Talking to a chatbot like ChatGPT or Claude and getting a fluent, relevant reply back is close to what writers pictured decades ago.
Fiction also pictured AI woven into everyday life instead of locked away in a lab, and that part came true as well.
It's in phones, search engines, email tools, and workplace software. Most people never stop to notice it's there.
There is also a genuine element of surprise capability.
Modern chatbots can write, summarise, translate, and explain things across an enormous range of topics with very little instruction.
Earlier generations of researchers did not expect a single system to be this broadly useful this quickly.
The parts that still have not, and probably will not soon
Where fiction overstated things is everything to do with a machine having its own mind. HAL 9000 had goals. Skynet had intent. Data from Star Trek had something resembling self-awareness.
Today's chatbots have none of that.
A large language model predicts the next likely word based on patterns learned from huge amounts of text.
It does not want anything, decide anything for itself, or understand a conversation the way a person does, even when its answers sound thoughtful.
General, human-level AI, a system that can reason, learn, and adapt across any task the way a person can, also does not exist yet.
Every AI tool available today is narrow. It is very good at specific kinds of tasks and has no general understanding beyond them.
Science fiction was right that machines would eventually talk like people. It was wrong that talking like a person means thinking like one.
Limitations: what the history of AI does not tell someone about AI today
This article explains how AI got here. It does not explain whether today's chatbot can be trusted with today's question.
Knowing that Turing asked "can machines think?" in 1950, or that ChatGPT launched in 2022, says nothing about whether a specific AI answer is accurate right now.
AI can still be wrong, and it often sounds just as confident when it is wrong as when it is right.
History explains the path. It does not vouch for the destination.
The pace of change is also worth flagging directly.
AI capabilities move fast, and a timeline like this one is a snapshot, not a live picture.
New models, new tools, and new limitations arrive constantly, and this history will need new entries added to it before long.
This history also says nothing about privacy or safety.
Whether it is sensible to share personal information with a given AI tool depends on how that specific product handles data, not on how the field developed since the 1950s.
The safest approach is to treat this article as background, not as a verdict on any current AI tool.
AI works best as a helper, not the final authority, and that applies just as much to a chatbot's factual claims as it does to its confident, well-written prose.
Final thoughts
The history matters because it puts today's headlines in perspective.
A lot of coverage about AI treats it as something that appeared overnight. It did not.
The version of AI people use today is the result of roughly seventy years of work, most of it slow, unglamorous, and largely ignored outside research circles.
The honest, specific truth is that progress was never one smooth line.
It was decades of steady effort interrupted by two long funding droughts, the AI winters, followed by a handful of pivotal breakthroughs that most people never noticed at the time.
Deep learning's return in 2012 and the transformer paper in 2017 are the clearest examples.
Neither made headlines outside AI research when they happened. Both quietly made everything that followed possible.
The story is also not finished.
The next chapter is being written now, and nobody outside the labs building these systems knows exactly what it contains.
Understanding where AI came from does not make anyone an expert in what it can do today.
What it does is give a sense of scale.
Sudden-looking breakthroughs are rarely as sudden as they appear, and long periods of quiet are not evidence that nothing is happening.
That context is a genuinely useful filter for judging AI claims more sensibly, whether they come from a headline, a company announcement, or a chatbot's own answer.
Frequently asked questions
Who invented artificial intelligence?
Nobody, really, if it's being honest about it. Turing gets the credit almost by default: his 1950 paper asking whether machines could think is the moment everyone points back to. The actual name came later, in 1956, when a handful of researchers at a Dartmouth College workshop settled on "artificial intelligence" as what to call it. Hundreds of people since then, working across decades, are who actually built it, not one person's lightning-bolt moment.
What is considered the first chatbot?
ELIZA is widely considered the first chatbot. Joseph Weizenbaum built it in 1966, using fairly simple pattern-matching rules to simulate a conversation. Even so, some people using it were convinced they were talking to something that understood them. That's really why people still remember it. ELIZA showed how easily people read understanding into a system that had none.
When did AI really start being useful to normal people?
Later than most people assume. The 2012 deep learning breakthrough and the 2017 transformer paper mattered enormously to researchers, but almost nobody outside AI circles noticed either one at the time. What actually changed daily life was ChatGPT's public launch on 30 November 2022, with Claude arriving in March 2023 and Gemini that December. Everything before ChatGPT lived inside research labs and specialist tools, out of reach for most people.
Is today's AI the same thing science fiction predicted?
Partly. Fiction correctly predicted natural-sounding conversation with a machine and AI becoming part of everyday life, both of which have happened. It significantly overstated the idea of AI having its own goals, desires, or self-awareness. Today's chatbots generate fluent language from learned patterns; they do not think or want things the way HAL 9000 or Skynet were depicted doing.
What caused the AI winters?
The AI winters in the 1970s and 1990s happened when AI research repeatedly failed to deliver on the ambitious promises made about it. Funding bodies and investors lost confidence when early systems could not scale up to real-world problems, and interest and money pulled back sharply both times. Each winter eventually ended when a new technical approach showed enough promise to rebuild confidence in the field.
Why did AI suddenly get so much better after 2012?
The 2012 breakthrough came from deep learning proving itself on an image recognition task, which reignited serious interest in the field after years of comparatively slow progress. That renewed momentum, combined with the 2017 transformer paper, which introduced a far more efficient way for AI systems to process language, gave researchers the tools needed to build the large, capable chatbots that arrived a few years later.
