The AI Hype Cycle

In January 2023, a single stat escaped the lab and ate the internet: ChatGPT had reached 100 million users in two months. The fastest consumer product adoption in history. Every newsletter, every LinkedIn feed, every dinner table got a version of it. CEOs called emergency meetings. Journalists raced to file think-pieces. Investors rewrote their theses overnight.

Then, roughly six months later, the think-pieces took a turn. "AI Hype Is Over," one read. "The ChatGPT Bubble Is Bursting," said another. The technology hadn't changed. What changed was the story.

This is the hype cycle. Not a metaphor — a documented pattern that Gartner Research has tracked across technology sectors since the mid-1990s. Five phases: Technology Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, Plateau of Productivity. We have been through this with AI before. Twice, actually.

The first AI winter hit around 1974. The field had been riding a wave of optimism since the 1956 Dartmouth Conference, where researchers predicted human-level machine intelligence within a generation. DARPA poured money in. Universities expanded programs. Then a UK government report — the Lighthill Report, commissioned in 1972 — concluded that AI had failed to meet virtually every promise made in the previous decade. Funding collapsed. The field entered a decade-long drought.

The second winter came in the late 1980s. Expert systems — AI programs encoding human expertise as rules — had attracted a fresh wave of investment. By 1987, the market for AI hardware had reached $400 million. By 1988, it was collapsing. The systems were too brittle, too expensive, and too limited to the specific domains they were built for. DARPA cut funding. Japan's Fifth Generation Computer Project, a decade-long national effort to build AI hardware, quietly wound down. Funding evaporated again.

Neither winter meant AI was wrong. They meant the gap between what people believed was imminent and what the technology could actually deliver had become unsustainable. Hype is not about capability. It's about expectation management — and who benefits from managing those expectations.

The gap between what AI can do in a controlled demo and what it does reliably in the real world has funded more venture rounds than any benchmark ever has.

What makes the current moment different from previous cycles is that the technology has genuinely crossed thresholds it never had before. Large language models do useful things at scale that 2010-era systems could not touch. But "genuinely different" does not mean "immune to the cycle." It means the cycle plays out at higher stakes and higher velocity. A research paper posts to arXiv at 9 a.m. A tech blog summarizes it by noon. Mainstream outlets run the headline by 3 p.m. It's trending by dinner. The reality check — narrower conditions, unreplicated results, limitations buried in appendices — arrives days or weeks later, after the next announcement has already taken over.

Watch the language in AI announcements and you'll start to see the machinery. A lab publishes a paper — usually measured in tone, heavy on caveats. The tech press rewrites the abstract as a headline. The mainstream press rewrites the headline. Social media argues about the rewrite. By the time the panic or euphoria reaches you, the actual paper has been through four layers of telephone. The hedges are gone. The conditions are gone. What remains is the shape of a claim, stripped of everything that made it defensible.

The specific words to watch: "can now" (means in a controlled setting, sometimes), "researchers say" (which researchers — pre-print or peer-reviewed, funded by whom?), "could" and "might" (doing enormous weight-bearing with no structural support), and percentage accuracy with no baseline (95% sounds impressive until you know the baseline was 94%).

Try this. Find an AI headline from this week that uses a strong verb — "masters," "beats," "surpasses." Find the original source. Check whether it's a pre-print or peer-reviewed. Find the limitations section, which in academic papers is typically at the end and almost never quoted by journalists. Then check who funded the research.

What we'd notice

The limitations section usually tells a different story than the title. Benchmarks are often narrow, conditions controlled, performance uneven across subgroups. Funding sources — particularly when a lab publishes results about its own model — create a structural incentive that no disclosure footnote fully resolves. This doesn't make the research false. It makes the gap between "paper finding" and "headline claim" almost always larger than it appears. And that gap is exactly where the hype lives.

None of this means you should become a dismisser. Cynicism and skepticism aren't the same thing. The dismissers in 1987 who called neural networks a dead end were wrong. The boosters in 1974 who called machine intelligence imminent were also wrong. The useful skill isn't knowing which camp to join. It's knowing how to sit in the uncertainty without needing to resolve it into a story.

The AI cycle has one more cruel feature: real progress often happens in the trough. The useful products ship quietly, months or years after the headline. The boring, incremental deployment that actually works gets no coverage because "AI does useful thing moderately well in production" doesn't move engagement metrics. The Gartner plateau — the phase where a technology becomes part of the infrastructure — is invisible precisely because it's working.

You're reading this on a platform built because the tools got useful enough to build it. The hype got us here faster. The cynicism might slow it down. What gets you through both is the habit of asking: who's making this claim, under what conditions, and what do they need you to believe?

The Deeper Question

If the hype cycle has been documented since the 1990s and we've already lived through two AI winters, why does each new cycle feel so different from the inside — and does that feeling tell us something about the technology, or only about us?

The limitations section of a paper is usually more informative than the abstract, but almost no one reads it. Is that a failure of science communication, a failure of journalism, or a feature of how humans process uncertain information about high-stakes topics?

Gartner's model implies that the "Plateau of Productivity" is the destination — the phase where useful, boring deployment takes over. But what if AI keeps triggering new hype peaks before any plateau fully forms? Does the cycle model still hold, or does continuous rapid improvement break the pattern?

Cynicism and skepticism often get conflated in conversations about AI claims. Is there a meaningful distinction — and if so, what does each one look like in practice when you're evaluating a specific headline?

The two AI winters dried up funding and slowed progress for years. If the current hype cycle inflates expectations unsustainably, what's the realistic downside — and is overcorrection more dangerous than the hype itself?