AI feels inevitable now — like it was always going to succeed. It wasn’t. Twice in living memory, artificial intelligence went from the most exciting technology in the world to a punchline. Funding dried up, researchers fled the field, and “AI” became a word you avoided on grant applications.
These collapses are called AI winters. Understanding them matters, because the same dynamics — hype, disappointment, retreat — are always lurking. Here’s the full story, and the honest question of whether it could happen again.
Table of Contents
- The Short Answer
- The Pattern: Hype, Disappointment, Winter
- The First AI Winter (1970s)
- The Second AI Winter (Late 1980s–1990s)
- What Actually Caused the Winters
- What Ended Each Winter
- The Survivors: What Kept Working in the Cold
- Could Another AI Winter Happen?
- Frequently Asked Questions
The Short Answer
An AI winter is a period when funding, interest, and progress in artificial intelligence collapse after a hype cycle fails to deliver. There have been two major ones: in the mid-1970s (after early promises about machine translation and general intelligence fell flat) and from the late 1980s through the 1990s (after expert systems proved brittle and expensive). Each winter lasted roughly a decade and set the field back years.
The Pattern: Hype, Disappointment, Winter
Every AI winter follows the same script:
- Breakthrough demo. Something impressive works in a lab. Researchers extrapolate wildly.
- Bold predictions. “Within a generation, machines will do any work a man can do” (actual 1965 quote from a Nobel laureate).
- Money flows in. Governments and companies fund AI labs generously.
- Reality arrives. The demos don’t scale. The hard problems — common sense, real-world messiness — prove much harder than expected.
- Funding collapses. Disappointed sponsors pull out. Researchers rebrand their work to avoid the toxic “AI” label.
- Slow thaw. Quiet progress continues with less money and less attention, eventually seeding the next boom.
Recognizing this pattern doesn’t make you a cynic. It makes you historically literate. And right now, in the middle of the biggest AI boom ever, historical literacy is valuable.
The First AI Winter (1970s)
The setup: In the 1950s and 60s, early AI produced genuinely exciting demos. Programs solved algebra word problems, proved math theorems, and played decent checkers. Pioneers made extraordinary claims — Marvin Minsky said in 1967 that “within a generation, the problem of creating artificial intelligence will substantially be solved.”
The trigger: Governments had funded ambitious projects, especially machine translation (the US wanted Russian translated automatically during the Cold War). The results were embarrassing. The famous (possibly apocryphal) story: “the spirit is willing but the flesh is weak” translated to Russian and back as “the vodka is good but the meat is rotten.”
In 1973, the UK’s Lighthill Report savaged AI research as failing to deliver on its “grandiose objectives,” leading to dismantled funding in Britain. In the US, DARPA cut academic AI funding sharply. The money didn’t just slow — it fell off a cliff.
The damage: AI became career poison. Researchers called their work “informatics” or “pattern recognition” — anything but AI. Progress continued in the shadows, but the field lost a decade of momentum and talent.
The Second AI Winter (Late 1980s–1990s)
The setup: AI came roaring back in the 1980s on the back of expert systems — programs that encoded human expert knowledge as rules. “If the patient has these symptoms and these test results, then consider this diagnosis.” Companies spent billions. Japan launched the Fifth Generation Computer project. Expert systems were going to revolutionize business.
The trigger: Expert systems turned out to be brittle and expensive in exactly the ways critics predicted:
- Knowledge acquisition bottleneck. Extracting expertise from humans and encoding it as rules was agonizingly slow. Every new domain meant starting over.
- Brittleness. The systems worked within their narrow domain and failed absurdly outside it. No common sense, no graceful degradation.
- Maintenance nightmare. When the world changed, someone had to manually update thousands of rules. The systems rotted.
- Cheaper alternatives. As computing got cheaper, simpler statistical approaches often outperformed expensive expert systems.
By the early 1990s, the expert system industry had collapsed. The AI hardware companies (Lisp machines) died when general workstations caught up. Funding dried up again. “AI” was once more a word you didn’t put on your grant proposal.
The damage: Another lost decade for the label, though crucially, the ideas survived. Neural networks — dismissed as a dead end — were quietly kept alive by a handful of researchers who turned out to be right.
What Actually Caused the Winters
Strip away the details and the causes rhyme:
Overpromising. Each boom featured claims that went far beyond the evidence. When you promise general intelligence and deliver a chess program, disappointment is inevitable — even if the chess program is genuinely impressive.
Demos that don’t scale. Lab demos work on toy problems. The real world is messier, noisier, and less cooperative. The gap between “works in the lab” and “works in production” killed more AI projects than any technical flaw.
Economics. AI research is expensive. When the returns don’t materialize on schedule, funders leave — and they’re slower to return than they were to arrive.
Narrow success mistaken for general progress. Each era mistook progress on specific tasks for progress toward general intelligence. Playing chess well didn’t mean understanding the world. Neither does writing fluent text — a point worth remembering today.
What Ended Each Winter
Winters don’t end because the hype returns. They end because of boring, unglamorous progress:
After the first winter: Cheaper computing, better algorithms, and the expert systems boom. Also, researchers simply kept working — the ideas didn’t die, just the funding.
After the second winter: Three things converged in the 2000s-2010s. Data — the internet created training datasets of unprecedented scale. Compute — GPUs, built for games, turned out to be perfect for neural networks. Algorithms — deep learning breakthroughs (backpropagation made practical, then transformers) finally made neural networks work at scale.
Notice what ended the winters: not better promises, but better fundamentals. Data, compute, and algorithms — the unsexy infrastructure. The way AI models are trained today is the direct descendant of that thaw.
The Survivors: What Kept Working in the Cold
Here’s the part people miss: AI winters weren’t total freezes. Enormous progress happened during them, just without the label or the spotlight:
- Machine learning for practical tasks. Spam filters, fraud detection, recommendation systems — all advanced during the “winter.”
- Robotics and control systems. Industrial robots transformed manufacturing while nobody called it AI.
- Speech and vision research. The foundations of today’s systems were laid by researchers working with tiny budgets.
- The neural network faithful. A small group kept believing in neural nets through decades of dismissal. They were vindicated spectacularly.
The lesson: winters kill hype and funding, not ideas. The best time to do foundational work might be when nobody’s watching.
Could Another AI Winter Happen?
The honest answer: a full winter like the 1990s is unlikely, but a correction is very possible. Here’s my assessment:
Why a true winter is unlikely:
– AI is now embedded in real products generating real revenue — not just research demos
– The infrastructure (cloud, chips, data pipelines) is permanent now
– Too many industries depend on it to let it fully collapse
Why a correction is plausible:
– Investment levels assume continued exponential progress, which isn’t guaranteed
– If the next capability jump takes longer than expected, funding could tighten sharply
– Regulatory shocks or a major incident could freeze deployment
– The “productivity miracle” might arrive slower than the valuations require
What a modern correction would look like: Not researchers fleeing the field, but startup die-offs, Big Tech AI budget cuts, consolidations, and a shift from “AI everywhere” to “AI where it provably works.” Painful for investors, mostly invisible to users of mature products.
My advice: enjoy the boom, build on the fundamentals, and keep the history in mind. And build with tools that survive hype cycles — the mature AI tools with real users are the ones that make it through any correction. The people who survived the winters were the ones focused on what actually worked, not what was hyped. That’s still the right strategy.
Frequently Asked Questions
An AI winter is a period when AI funding, interest, and progress collapse after a hype cycle fails to deliver on its promises. There were two major ones: in the mid-1970s (after machine translation and early AI overpromised) and from the late 1980s through the 1990s (after expert systems proved brittle). Each lasted roughly a decade.
The same pattern both times: overpromising (claims far beyond the evidence), demos that didn’t scale to real-world messiness, expensive research that didn’t deliver returns on schedule, and mistaking narrow task success for general progress. When funders got disappointed, money vanished and researchers avoided the “AI” label for years.
The second AI winter ran roughly from the late 1980s through the 1990s, ending in the 2000s as internet-scale data, GPU computing, and deep learning breakthroughs converged. The current boom began around 2012 (deep learning breakthroughs) and accelerated massively after 2022.
Not renewed hype, but better fundamentals: cheaper computing, internet-scale training data, and algorithmic breakthroughs (deep learning, then transformers). Boring infrastructure progress — data, compute, algorithms — is what thawed each winter, not better promises.
A full 1990s-style winter is unlikely because AI is now embedded in revenue-generating products and permanent infrastructure. But a correction is plausible: if capability progress slows, expect startup die-offs, budget cuts, and consolidation — a shift from “AI everywhere” to “AI where it provably works.”
No — that’s the overlooked part. Spam filters, fraud detection, industrial robotics, and speech research all advanced during the “winters,” just without the AI label or spotlight. Winters kill hype and funding, not ideas. Foundational work continued and eventually seeded the next boom.
