The AI Hype Cycle: How to Tell Signal from Noise

If you work in or around technology right now, the skill that matters most is not technical literacy. It is knowing how to tell a durable shift from a passing frenzy. Every major technology of the past half-century has gone through the same arc: wild expectations, a painful correction, then years of quieter, productive adoption. AI is following the same script, only faster and with more money behind it.

Consider the numbers. Global venture capital poured $202 billion into AI in 2025, roughly half of all VC funding worldwide. The four largest cloud providers committed over $380 billion in AI capital expenditure for the same year. Yet McKinsey's 2025 survey found that only 7% of organisations have fully scaled AI across their operations, and Goldman Sachs' chief economist noted that AI investment contributed "basically zero" to US GDP growth in 2025.

That gap between investment and measurable impact is where clear thinking matters most. This article offers a framework for it, drawing on academic theory, historical evidence, and current data.

The intellectual toolkit for understanding technology hype

Several complementary frameworks help explain why societies repeatedly overestimate new technologies in the short term and underestimate them in the long term. None is perfect on its own, but together they offer a useful lens for evaluating what is happening with AI right now.

The Gartner Hype Cycle: useful metaphor, unreliable map

The most widely recognised framework is the Gartner Hype Cycle, created by analyst Jackie Fenn in 1995. Its five phases (Innovation Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, Plateau of Productivity) provide an intuitive narrative for how expectations evolve. Gartner now produces over 100 hype cycles annually across different domains. In 2023, it placed generative AI at the Peak of Inflated Expectations; by 2025, it had officially begun its slide into the Trough.

Yet the model's predictive value is questionable. Martin Steinert and Larry Leifer of Stanford analysed 46 technologies across Gartner's 2003–2009 cycles and found methodological flaws and procedural inconsistencies. Technology placements frequently diverged from actual market visibility as measured by news coverage and search interest.

Michael Mullany, a venture partner at Icon Ventures, tracked every Hype Cycle for Emerging Technologies from 2000 to 2016, cataloguing over 200 technologies. His conclusion was blunt: the median technology does not obey the Hype Cycle. Only a handful (cloud computing, 3D printing, natural-language search) followed the predicted path. More than 50 technologies appeared only once and vanished. Many of the technologies that actually changed things, including x86 virtualisation, NoSQL databases, and smartphones, never appeared on the cycle at all, or arrived at the wrong phase.

The Hype Cycle is best understood as a useful metaphor rather than an empirical model. It tells a good story about technologies, but it is not a scientific prediction.

Amara's Law: the most reliable pattern in technology forecasting

More robust than the Hype Cycle is a simpler observation. Roy Amara, president of the Institute for the Future, noted sometime in the 1960s or 1970s: "We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run."

The dot-com crash of 2000 seemed to confirm the internet was overhyped; today's internet economy vastly exceeds even the most ambitious 1990s predictions. Paul Krugman's 1998 assertion that "by 2005 or so, it will become clear that the internet's impact on the economy has been no greater than the fax machine's" is a monument to long-term underestimation. The mechanism, as complexity theorist Stuart Kauffman observed, involves the "adjacent possible": each technological advance opens doors to further advances that were previously invisible. The IEEE has noted that Amara's Law remains the most durable insight in technology forecasting.

For professionals, the practical implication is straightforward: whatever AI achieves in the next two years will probably disappoint relative to current expectations. Whatever it achieves over the next fifteen will probably exceed them.

Carlota Pérez: technology revolutions unfold in two phases

The Venezuelan-British economist Carlota Pérez goes deeper. In Technological Revolutions and Financial Capital (2002), she identifies five great technological revolutions since the 1770s, each unfolding over 40–60 years in two phases separated by a financial crisis.

The Installation Period, driven by speculative financial capital, features an Irruption Phase followed by a Frenzy Phase where asset bubbles inflate. A Turning Point (typically a crash and regulatory reform) then gives way to the Deployment Period, where the technology spreads widely through a Synergy Phase and eventual Maturity Phase.

Writing in Project Syndicate in March 2024, Pérez cautioned that AI must be understood as belonging to a larger, more mature technological revolution that began a half-century ago with the microprocessor. It is a major development within the ongoing ICT revolution, not necessarily the start of a new one. That distinction matters, because it tempers the claims about AI being the biggest thing since the printing press.

For professionals evaluating AI investments, Pérez's framework is clarifying. If we are in a late Installation Period, the current spending boom is not irrational; it is how technological revolutions have always been financed. But the Turning Point, when speculative excess gets corrected and the real productive work begins, is likely ahead of us rather than behind us.

Rogers and Moore: the adoption gap that hype obscures

Everett Rogers' Diffusion of Innovation theory (1962) provides the adoption lens. His bell curve (innovators at 2.5%, early adopters at 13.5%, early majority at 34%, late majority at 34%, laggards at 16%) helps explain why technologies that seem universally adopted in Silicon Valley have barely reached the pragmatic mainstream.

Geoffrey Moore's extension in Crossing the Chasm (1991) identified the critical gap between visionary early adopters and pragmatic early-majority buyers. The Hype Cycle's Peak of Inflated Expectations corresponds roughly to adoption by innovators and early adopters; its Trough maps onto Moore's Chasm.

Rogers himself identified pro-innovation bias (the assumption that innovation is inherently good and should be adopted by all) as a weakness in how researchers and practitioners evaluate new technology. This bias is everywhere in the current AI conversation, where questioning whether a particular AI application is actually useful is often treated as a failure of imagination rather than a reasonable exercise of judgement.

Why smart people consistently get technology predictions wrong

Understanding these frameworks is necessary but not sufficient. You also need to understand why intelligent, well-informed people repeatedly fail to apply them. The answer lies in a set of cognitive biases that novel technology is particularly good at triggering.

Kahneman and Tversky's availability heuristic means vivid success stories (the next Google, the next Uber) are recalled more easily than the silent majority of failures, inflating our sense of a technology's likely trajectory. The bandwagon effect drives companies and investors to adopt technologies they do not fully understand, because everyone else appears to be doing so. Anchoring bias means that early, speculative market-size projections ("AI will be worth $X trillion by 2030") become reference points that shape investment decisions long after the projections have lost their basis. And Nassim Taleb's narrative fallacy, described in The Black Swan (2007), captures our limited ability to look at sequences of facts without weaving an explanation into them. We construct post-hoc stories that make chaotic technological developments seem inevitable.

Philip Tetlock's 20-year study provides the quantitative backing. His research, tracking 284 experts making roughly 28,000 predictions and published in Expert Political Judgment (2005), found that the average expert performed about as well as random guessing, his famous "dart-throwing chimpanzees." The split that mattered: "hedgehogs" (experts with a single big idea) performed worst, while "foxes" who drew on diverse perspectives did significantly better. This is directly relevant to AI forecasting, where the loudest voices tend to be hedgehogs with singular narratives about imminent superintelligence or inevitable doom.

The consulting industry's track record reinforces the point. In the 1980s, McKinsey predicted there would be only 900,000 mobile phone subscribers by 2000; the actual figure was 109 million, off by a factor of 120. AT&T reduced cell tower investment based on this advice and later paid $12.6 billion to acquire McCaw Cellular to recover lost ground. IMF economist Prakash Loungani found that economists failed to predict 148 of 150 recessions. The pattern is consistent: forecasters systematically underestimate how technologies improve and overestimate how quickly improvement translates into adoption.

What history actually tells us

Theory is useful. Evidence is better. The following case studies show a consistent pattern across decades of technology hype, and they say something about where AI sits today.

The dot-com bubble: the technology was real, the timeline was wrong

The NASDAQ peaked at 5,132 on 10 March 2000, then crashed 78% to roughly 1,114 by October 2002, destroying over $5 trillion in market capitalisation. Pets.com raised $82.5 million in its IPO, reached a market capitalisation exceeding $300 million, and declared bankruptcy nine months later. Webvan burned through $1.2 billion building logistics infrastructure for online groceries before collapsing.

Yet the underlying thesis, that the internet would change commerce, media, and communication, proved correct in ways that even the optimists underestimated. Amazon's stock fell 94% to roughly $7 after the crash. It did not regain its 1999 peak until late 2009. Today, a $1,000 investment at the dot-com peak would be worth approximately $15,500. The NASDAQ itself did not recover to its 2000 level until April 2015. Fifteen years.

The technology was real. The timing was wrong. And the infrastructure laid during the bubble (fibre optic cables, early e-commerce platforms, data centres) enabled the productive era that followed. This is Pérez's framework in action: Installation Period excess funding Deployment Period infrastructure.

The AI winters: right ideas, wrong century

The AI winters provide the most direct precedent for the current moment. At the 1956 Dartmouth conference, the field's founders believed human-level intelligence could be achieved within a generation. Herbert Simon predicted in 1965 that machines would be capable, within twenty years, of doing any work a man could do.

Reality intervened. Sir James Lighthill's 1973 report to the British Science Research Council criticised AI for failing to achieve its grandiose objectives, and the British government gutted AI funding at all but two universities. The second winter, triggered by the collapse of the expert systems market in 1987, saw an entire industry worth half a billion dollars evaporate in a single year when desktop computers became powerful enough to run LISP software at a fraction of the cost. The Japanese Fifth Generation Computer Project spent roughly ¥50 billion over a decade without meeting its goals.

Yet the ideas explored during these periods (neural networks, natural-language processing, computer vision) were largely correct. They were simply 30 to 50 years ahead of the available compute and data. The researchers who kept working through the winters laid the intellectual groundwork for everything happening today.

Blockchain: a solution looking for a problem

Blockchain and crypto offer a cautionary tale about what happens when enthusiasm outpaces utility. Bitcoin rose 1,900% in 2017. Long Island Iced Tea Corporation renamed itself Long Blockchain Corporation and saw its stock rise 400% overnight. IBM called itself "the blockchain leader for business" in 2017; by 2021, its blockchain unit had missed revenue targets for two consecutive years and most staff had departed. TradeLens, IBM and Maersk's shipping blockchain, shut down in November 2022 citing lack of commercial traction.

The pattern: extraordinary investment, corporate rebranding to capture buzz, then a painful reckoning when the technology failed to solve problems better than existing alternatives. The question that separated blockchain from the internet, and the one that should be applied to every AI use case, is simple: does this solve a real, measurable problem that people are willing to pay for?

Self-driving cars: the moving-goalposts problem

Elon Musk predicted full self-driving capability in December 2015 ("approximately two years"), April 2017 ("November or December of this year"), 2019 ("this year"), and 2021 ("over 1 million robo-taxis" by year-end). None materialised on schedule. This pattern of annually repeating the same prediction, always "next year," always deferred, is one of the most reliable red flags in technology forecasting.

Waymo, by contrast, spent 16 years developing its technology before reaching meaningful commercial deployment: 450,000 weekly rides across five cities by late 2025, with safety data showing 85% fewer crashes with serious injuries than human drivers.

Real technological deployment follows Waymo's patient, decade-long trajectory, not Musk's perpetual "next year" promises. When evaluating AI claims, ask which timeline a company's roadmap more closely resembles.

Meta's metaverse and the cost of premature commitment

Meta's metaverse pivot consumed roughly $80–90 billion in cumulative Reality Labs operating losses, with only $2.2 billion in revenue for all of 2025. When Meta announced budget cuts to VR, its stock rose 4–5%. Investors viewed the metaverse as a drag on the business, not a driver of it.

Cloud computing, by contrast, moved from Larry Ellison publicly mocking it in 2008 to becoming a $723-billion market by 2025. The smartphone went from Steve Ballmer's dismissal in 2007 ("There's no chance that the iPhone is going to get any significant market share") to 7 billion devices within 15 years. The difference between the metaverse and the smartphone was not ambition. It was that one solved an immediate, tangible problem for hundreds of millions of people from day one.

The pattern across all of these cases

The technology is usually real; the timeline is usually wrong. The hype-to-productivity gap is typically 10–15 years. Infrastructure investment survives busts. And the companies that find sustainable business models (Amazon, Google, Waymo) win precisely because they outlast the hype.

The current AI moment: what the evidence actually shows

With this historical context, we can look at the current AI moment more clearly. The empirical picture in early 2026 shows a technology of genuine capability operating well below its most inflated expectations.

Benchmarks are impressive but misleading

On standard tests, top models score above 90% on MMLU (the broad knowledge test) and above 96% on HumanEval (Python coding tasks). These scores are impressive but they flatter the technology.

On "Humanity's Last Exam," a rigorous multi-domain academic test, the best system scored just 8.8%. On FrontierMath, AI solved only 2% of problems. On BigCodeBench, which tests real-world programming with library calls and external dependencies, top models achieved 35.5% versus a 97% human baseline. The Stanford AI Index 2025 concluded that even with chain-of-thought reasoning improvements, these systems still cannot reliably solve problems for which provably correct solutions can be found using logical reasoning.

Benchmark performance is not real-world reliability. Benchmarks are saturating faster than real-world capability is advancing, and the gap between the two is where costly mistakes get made.

Adoption is wide but shallow

The adoption picture shows a technology that is everywhere and yet somehow still peripheral. McKinsey's 2025 survey found 88% of organisations use AI in at least one function. But only 7% have fully scaled AI across their organisations. An MIT study found 95% of generative AI pilots failed to yield meaningful results. IBM reported that less than half of IT leaders said their AI projects were profitable in 2024.

The pattern is experimentation without follow-through: widespread tinkering, limited embedding. This is consistent with the early stages of a general-purpose technology deployment, but it is a long way from the "AI is transforming every business" narrative that dominates headlines.

Productivity gains are real but uneven

This is where the data deserves the closest attention.

Brynjolfsson, Li, and Raymond's study of 5,172 customer support agents found a 14–15% average productivity increase, with novice workers improving by 34–35%. The BCG/Harvard experiment with 758 consultants showed those using GPT-4 completed 12.2% more tasks, 25.1% faster, with 40% higher quality. But only for tasks within AI's capability frontier. Outside that frontier, AI users were 19 percentage points less likely to produce correct solutions than those working without it.

GitHub Copilot studies show developers completing specific tasks 55.8% faster, but an independent study by Uplevel Data Labs found significantly higher bug rates with no throughput improvement. A METR study of experienced open-source developers found AI tools actually slowed them down. And an MIT Sloan analysis of over 100 experimental studies found that on average, human-AI collaborations underperform both the AI alone and the best human decision-makers.

The concept that emerges from this research, the "jagged technological frontier," matters more than any single number. It is not obvious, even to skilled workers, which tasks AI handles well and which it handles poorly. The frontier is irregular and unpredictable. Working effectively with AI means developing an intuition for where that frontier lies in your specific domain, and that intuition can only come from careful experimentation, not from reading about someone else's experience.

The investment-to-revenue gap

On the financial side, the numbers are hard to ignore. Sequoia Capital's David Cahn identified a $500–600 billion revenue gap: infrastructure investment massively outpacing demand. Goldman Sachs asked: "What $1 trillion problem will AI solve?" OpenAI's projected 2025 revenue of $15 billion sits against $150 billion in investment. Goldman warned that hyperscalers would need annual profits of over $1 trillion, more than double the 2026 consensus estimate, to justify current spending.

The comparison to the dot-com era is instructive but not straightforward. Current AI spending sits at roughly 0.8% of GDP, versus 1.5–2% during the late-1990s telecom boom. And the spending is funded by the cash flows of the world's most profitable companies (Apple, Microsoft, Google, Amazon) rather than by speculative debt. Goldman Sachs projects over $500 billion in AI investment for 2026. The question is not whether this money is being spent but whether it will generate returns before patience runs out.

Scaling laws and hallucinations: two structural uncertainties

Two technical developments add complexity to the picture.

The power-law relationships between model performance and compute, identified by Kaplan et al. in 2020, appeared to promise continuous improvement through sheer scale. But as HEC Paris analysis noted, frontier models appear to have reached a ceiling, a poorly kept secret across the AI industry. Ilya Sutskever, OpenAI co-founder, stated publicly that "we have achieved peak data." Multiple labs delayed or disappointed with next-generation models. Pre-training scaling may be plateauing, though new paradigms (test-time compute scaling, post-training reinforcement learning, agentic architectures) offer different pathways forward.

Hallucination remains structurally unsolved. Even the best models hallucinate on roughly 0.7–3% of prompts, a figure that sounds small until you consider the volume of queries. Legal AI tools hallucinate in 17–34% of cases, and 47% of enterprise users report having made at least one major decision based on hallucinated content. This is not a bug that will be patched in the next release. It is a property of how current generative AI systems work, and any professional deployment needs to account for it.

Where the experts actually stand

The people who know the most about AI disagree about something more basic than timelines. They disagree about what current systems actually are.

Yann LeCun, Meta's former chief AI scientist and Turing Award laureate, represents the strongest internal critique. He departed Meta in late 2025 to found AMI (Advanced Machine Intelligence), calling LLMs a "dead end" and arguing that they lack true understanding of the physical world, persistent memory, genuine reasoning, and complex planning. His core point: LLMs operate in "System 1" mode (reactive pattern matching) not "System 2" deliberate reasoning. Whether or not you agree with LeCun's conclusion, his argument is worth understanding because it identifies the specific limitations that current systems have yet to overcome.

Gary Marcus, NYU professor emeritus of cognitive science, has been the most consistent public sceptic. His 2024 predictions were largely vindicated: GPT-5 never shipped, scaling laws showed diminishing returns, and corporate adoption remained limited. His 2025 predictions included that AI agents would be endlessly hyped but far from reliable, profits from AI models would remain modest or nonexistent, and less than 10% of the workforce would be replaced by AI.

Dario Amodei, Anthropic's CEO, offers cautious optimism. His October 2024 essay "Machines of Loving Grace" argued that most people are underestimating both how radical the upside of AI could be and how bad the risks could be. He suggested AI could compress a century of progress in biology into 5–10 years, a claim he hedged with extensive caveats about uncertainty and risk.

Demis Hassabis, DeepMind CEO and 2024 Nobel laureate, placed AGI 5–10 years away but noted this requires one or two more genuine breakthroughs in world models, continual learning, and hierarchical planning. He described current systems as exhibiting "jagged intelligence": genius in some areas, worse than ordinary humans in others.

Among economists, the disagreements are just as sharp. Daron Acemoglu, MIT professor and 2024 Nobel laureate, estimated AI would boost US productivity by roughly 0.05% per year, a total GDP increase of just 1.1% over a decade. His reasoning: AI currently automates only a narrow slice of tasks, and automation creates displacement costs that offset some of the gains. Erik Brynjolfsson of Stanford countered with his "productivity J-curve" theory: general-purpose technologies require years of complementary investment before benefits appear, and he argued that recent productivity gains signal the beginning of a harvest phase. David Autor of MIT offered a different framing entirely: how AI is designed is a choice. It can be an automation tool that eliminates expertise or a collaboration tool that democratises it. The outcomes depend less on the technology itself than on the institutional and design choices made in the coming years.

The upshot: people with deep expertise disagree substantially about AI's trajectory. Anyone who tells you the answer is obvious, in either direction, is selling something.

A practical framework: eight tests for separating signal from noise

Historical patterns and current evidence point to specific indicators that distinguish technologies delivering on their promises from those that will not. Drawing on Stanford HAI's validity-centred framework for evaluating AI claims, multiple analyst assessments, and the historical record, these eight tests apply to any AI development, vendor claim, or investment decision.

1. Specificity. Are claims precise, with metrics, trade-offs, and acknowledged limitations, or vague, employing words like "revolutionary" and "transformative" without quantification? Genuine capability tends to be described in measurable terms. Hype tends to be described in adjectives.

2. Problem-solution fit. Does the technology solve a real, measurable problem better than existing alternatives? This is the question that separated the internet (which solved immediate problems for hundreds of millions of people) from the metaverse (which solved problems most people did not have). Apply it ruthlessly.

3. Practitioner versus influencer. Are engineers and operators quietly adopting the technology and integrating it into their workflows? Or is the buzz driven primarily by marketing departments, venture capitalists, and social media? When practitioner adoption leads, as it did with cloud computing and smartphones, that is a strong positive signal.

4. Revenue. Does the technology generate revenue from end users, not merely attract investment? When a technology's primary "customers" are other companies in the same ecosystem (AI companies buying chips from Nvidia, funded by investors who also invest in Nvidia) that is a circular financing pattern worth noting.

5. Benchmark-to-reality. Does laboratory performance translate to real-world deployment? The gap between MMLU scores above 90% and Humanity's Last Exam scores below 9% shows why this question matters. Ask for deployment data, not benchmark data.

6. Post-announcement. What happens after launch? Sustained engagement and deployment, or a spike of attention followed by silence? Track a technology's trajectory over 6–12 months, not 6–12 hours.

7. Independent verification. Has any third party validated the claims? Stanford HAI emphasises that AI capabilities should be evaluated through independent, reproducible assessments rather than vendor demos. If the only evidence for a claim comes from the company making it, apply extra scepticism.

8. Accountability. Would someone accountable for real outcomes, a CFO, a surgeon, a structural engineer, stake their professional reputation on deploying this technology in a high-stakes context today? This test cuts through theoretical potential and gets at practical readiness.

Red flags that recur across historical bubbles

Several warning signs keep appearing across technology hype cycles.

The first is circular financing. When investors fund companies that then buy products from the investors or their portfolio companies, growth metrics can look impressive while masking a lack of genuine end-user demand. Yale's Jeffrey Sonnenfeld observed this pattern in the current AI ecosystem, where Nvidia invests in OpenAI, OpenAI uses that capital to buy Nvidia chips.

Then there are moving goalposts. When the timeline for delivering on a promise shifts repeatedly (Musk's annual self-driving predictions, repeated "next year" AGI claims) that signals a gap between aspiration and capability that is not closing at the rate promised.

AI washing is another. Companies rebranding existing products as "AI-powered" echoes the ".com" suffix mania of 1999 and the "blockchain" renaming of 2017. The US SEC has already fined investment advisers $400,000 for misleading AI claims.

Finally, watch for investment massively outpacing revenue. When hyperscaler capex consumes 94% of operating cash flows, up 18 percentage points from the prior year, and revenue from the technology remains a fraction of the investment, history counsels caution.

Positive signals that the current moment is not merely hype

It is equally important to recognise what distinguishes the current AI moment from past bubbles that fizzled entirely.

Costs are falling fast. AI model costs dropped 99% in 18 months, faster than solar energy achieved over 40 years. Rapidly falling costs are one of the strongest historical indicators that a technology will find widespread adoption.

Developers are adopting at scale. GitHub Copilot reached over 15 million users by early 2025, a fourfold increase in one year. When the people who build things adopt a tool voluntarily, that is worth paying attention to.

The productivity gains, while uneven and domain-specific, are consistently real. A 14% average productivity increase in customer support, a 25% speed improvement for consultants on appropriate tasks. These are not fabricated numbers. They are modest, honest, and they separate genuine utility from vaporware.

And the funding base is different this time. Unlike the dot-com era, current AI spending is funded primarily by the cash flows of the world's most profitable companies, not by speculative debt or venture capital. This does not make the investment wise, but it does make the financing more resilient to sentiment shifts.

Where AI stands now, and what comes next

Where does AI sit in early 2026? It is a general-purpose technology in the early stages of its installation period. The capabilities are real but circumscribed. The productivity gains are measurable but domain-specific. The investment is enormous but not yet irrational by historical standards. And the persistent gaps, between benchmarks and real-world performance, between pilot adoption and enterprise-wide scaling, between capital expenditure and revenue, suggest we are closer to the peak of inflated expectations than to the plateau of productivity.

Benedict Evans captured the product challenge well: 800 million people have access to powerful AI tools, but most don't know what to do with them more than once a week. That is not a technology problem. It is a product problem. And product problems, historically, take years to solve.

For professionals making decisions today, the historical record offers three principles worth holding onto.

First, apply Amara's Law with discipline. Whatever AI achieves in the next two years will probably disappoint relative to current expectations; whatever it achieves over the next fifteen will probably exceed them. Make decisions accordingly. Avoid both panic buying and dismissive inaction.

Second, watch deployment, not demos. The difference between a compelling product launch and a technology that changes how organisations actually operate is typically a decade of unglamorous integration work: workflow redesign, staff training, process adaptation, error handling. Look for evidence of this work happening, not for flashy demonstrations.

Third, be a fox, not a hedgehog. Tetlock's research consistently shows that those who draw on diverse perspectives and update their beliefs incrementally make far better predictions than those committed to a single narrative, whether that narrative is imminent superintelligence or inevitable bust. Read the sceptics and the optimists. Seek out the people who changed their minds, not the ones who haven't.

There is real signal in the current moment. There is also a lot of noise. Telling them apart requires resisting both the breathless enthusiasm of the installation period and the premature dismissal that comes with the trough. The quieter indicators of productive adoption are there if you watch for them, but you have to be willing to look past the headlines.

Key takeaways

Every major technology follows a pattern of inflated expectations, painful correction, and eventual productive adoption. The hype-to-productivity gap is typically 10–15 years. For AI, the underlying capabilities are real and measurable, but the timeline for widespread deployment is almost certainly longer than current narratives suggest. The Gartner Hype Cycle is a useful metaphor but an unreliable predictive tool; Amara's Law and Pérez's two-phase framework are more dependable guides. Current AI productivity gains are genuine but domain-specific. The "jagged frontier" means it is not obvious which tasks benefit and which do not, even to experts. The eight-test framework in this article provides a practical method for evaluating any AI claim, product, or investment decision you encounter. The most reliable approach is to be a fox, not a hedgehog: draw on diverse perspectives, update beliefs incrementally, and invest based on evidence of deployment rather than the excitement of demos.

Further reading

Foundational texts referenced in this article include Carlota Pérez, Technological Revolutions and Financial Capital (2002); Philip Tetlock, Expert Political Judgment (2005); Geoffrey Moore, Crossing the Chasm (1991); and Nassim Taleb, The Black Swan (2007).

Key research papers include Brynjolfsson, Li, and Raymond, "Generative AI at Work" (NBER, 2023); Dell'Acqua et al., "Navigating the Jagged Technological Frontier" (Harvard Business School, 2023); and the Stanford AI Index 2025 annual report.

For analyst and expert perspectives, see Dario Amodei, "Machines of Loving Grace" (October 2024); Michael Mullany, "8 Lessons from 20 Years of Hype Cycles" (LinkedIn, 2016); and McKinsey's State of AI 2025 report.

On AI investment, see Goldman Sachs on AI capex projections and Crunchbase's AI funding analysis for 2025.