Say "AI" today and most people picture a chatbot writing an essay, or an image generator knocking out a fantasy castle in seconds. Back when your PC had a turbo button, the word meant something far more modest. Sometimes it meant a clever piece of business software. More often, if you were a gamer, it meant the enemy.
The Expert System Boom
The 1980s were AI's first real commercial gold rush, and it ran on expert systems. The idea was simple: sit down with a human specialist, turn everything they knew into hundreds or thousands of "if this, then that" rules, and let the computer apply them. Digital Equipment Corporation's XCON, which configured VAX computer orders, was the poster child, reportedly saving the company tens of millions of dollars a year by the mid 80s. Japan launched its Fifth Generation Computer project in 1982, and the US and UK scrambled to fund rival programmes. Companies like Symbolics sold dedicated Lisp machines built purely to run AI software.
It didn't last. Expert systems were brittle. Step slightly outside the rules and they fell over, and every new rule risked breaking an old one. Meanwhile ordinary desktop PCs got fast enough to run the same software, and the specialist hardware market collapsed around 1987. Funding dried up and the field slid into what's now called the second AI winter.
Winter, and a Quiet Rebuild
The 90s look quiet from the outside, but plenty of real groundwork was laid. Researchers stopped trying to hand code intelligence and started letting programs learn from data using statistics and probability. Neural networks already existed (a 1986 paper by Rumelhart, Hinton and Williams popularised backpropagation, the training method still used today), but a 90s machine simply couldn't train anything big enough to impress.
The moment everyone remembers came in May 1997, when IBM's Deep Blue beat world chess champion Garry Kasparov in a six game rematch. Huge story. It was also brute force: custom chips searching around 200 million positions a second, guided by an evaluation function tuned with help from grandmasters. It couldn't play noughts and crosses unless someone programmed it to. Office 97 gave the rest of us Clippy, whose eager "It looks like you're writing a letter" came out of Microsoft Research work on predicting what users wanted. Nobody called him intelligent.
AI You Could Actually Play Against
For PC gamers, AI was whatever controlled the other side. Most of it was a finite state machine: a character sits idle, spots you, chases, attacks, dies. Doom's imps weren't much deeper than that. Pathfinding algorithms like A* got RTS units round a map (usually). Designers cheated constantly, handing computer players extra resources or letting them see through the fog of war, because genuinely smart opponents cost CPU time nobody had.
Then things got clever. Thief: The Dark Project gave guards eyes and ears in 1998, so shadows and noisy floors actually mattered. Black & White let you raise a creature that learned from reward and punishment. F.E.A.R. in 2005 used goal oriented action planning, so soldiers worked out their own tactics on the fly, and it still holds up.
The 2000s: AI Goes Invisible
After 2000, AI slipped into everyday life and mostly stopped calling itself AI. Bayesian spam filters cleaned up inboxes. Amazon's recommendations nudged your shopping. Dragon NaturallySpeaking made dictation workable on a home PC, and the Roomba started bumping into sofas in 2002. In 2004 not a single driverless vehicle finished DARPA's desert Grand Challenge; a year later Stanford's Stanley won it. And in 2006 Geoffrey Hinton published work on training deep neural networks that, once GPUs caught up, kicked off everything that followed.
AI Milestones at a Glance
- 1980: DEC puts XCON into production to configure VAX orders, the first big commercial expert system success.
- 1982: Japan launches its Fifth Generation Computer project, triggering a global AI funding race.
- 1986: Rumelhart, Hinton and Williams popularise backpropagation for training neural networks.
- 1987: The Lisp machine market collapses and the second AI winter sets in.
- 1997: Deep Blue beats Garry Kasparov; Office 97 introduces Clippy.
- 1998: Thief: The Dark Project ships with guards that react to light and sound.
- 2002: iRobot releases the first Roomba.
- 2005: F.E.A.R. brings goal oriented planning to shooter AI; Stanford's Stanley wins the DARPA Grand Challenge.
- 2006: Geoffrey Hinton's deep belief network research reignites interest in deep neural networks.
Then Versus Now
Here's the big difference. Almost every system above did exactly one thing. Deep Blue played chess. XCON configured VAXes. A F.E.A.R. soldier found cover. Their smarts were either written by hand or trained on one narrow task, and nobody mistook them for minds.
Today's AI, the large language models behind ChatGPT, Claude and the rest, comes from that same neural network lineage, scaled up by several orders of magnitude and trained on a huge share of the written internet. That's why it feels general. It'll chat about pretty much anything and write working code. It's also why people assume it thinks the way we do. It doesn't, at least not in any way anyone can confirm. Underneath, it's still pattern prediction, just at a scale 1997 couldn't have imagined.
So modern AI is less of a clean break than it looks. It's the neural network idea that sat on a shelf through the 80s and 90s, finally handed the hardware it always needed. The expert systems crowd tried to write intelligence down. The winners turned out to be the people who let the machine learn it.