Why We Overestimate the Next Two Years and Underestimate the Next Twenty
AI hype is usually wrong about timing—not transformation
“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.” — Roy Amara
Key Takeaway
Amara’s Law explains one of the central paradoxes of artificial intelligence: society repeatedly expects AI to transform everything immediately, becomes disappointed when reality moves more slowly, and then fails to notice the deeper transformation taking place over decades.
AI is simultaneously overhyped and underestimated. Its immediate capabilities are exaggerated, while its cumulative effects on work, education, science, power, institutions, and human behavior are still poorly understood.
That is not a contradiction. It is Amara’s Law in action.
What Is Amara’s Law?
What Is Amara’s Law?
Amara’s Law describes a recurring mistake in the way people interpret technological change: we exaggerate what a new technology will achieve in the near future while underestimating how profoundly it may reshape society over the long term.
It is commonly summarized in Roy Amara’s famous observation:
“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”
This is not a mathematical law, nor does it predict when a technology will mature. It is a warning about expectations.
A dramatic demonstration creates the impression that mass transformation is imminent. The slower work of building infrastructure, reducing costs, improving reliability, creating regulations, and changing institutions is largely ignored.
When the promised revolution fails to arrive on schedule, enthusiasm collapses. Yet the technology continues developing beneath the surface, gradually becoming cheaper, more capable, and more deeply embedded in everyday life.
Artificial intelligence may be the clearest modern example of this pattern.
The graph below provides a simple visual model of how Amara’s Law works.
The expected-impact curve rises rapidly during the first wave of excitement, then declines as limitations, costs, and failed promises become visible.
The actual-impact curve grows more slowly because real transformation requires investment, infrastructure, regulation, skills, and social adaptation.
Over time, however, actual impact may surpass the original expectations as the technology becomes mature, widespread, and increasingly difficult to separate from ordinary life.
The distance between these two curves is where hype, disappointment, and genuine transformation occur. To understand how this insight emerged, we must now turn to the futurist whose name it carries.
But to understand why this simple observation became so influential, it helps to look at the futurist behind it.
Who Was Roy Amara?
Roy Charles Amara was an American engineer, researcher, and futurist whose career placed him at the intersection of computing, technological change, and long-term forecasting. After serving in the United States Navy during the Second World War, he studied business and engineering administration at MIT, earned a master’s degree in science teaching from Harvard, and later completed a doctorate in electrical engineering at Stanford University.
Amara did not arrive at his famous law through a single experiment or sudden moment of inspiration. It appears to have emerged gradually from decades spent comparing technological promises with what actually happened after companies, governments, and ordinary people began using those technologies.
At the Stanford Research Institute, Amara worked on early computing, interactive systems, and decision analysis. He later became president and CEO of the Institute for the Future, where he established its annual ten-year forecast and led research into emerging technologies and their social consequences.
The institute repeatedly encountered the same forecasting problem. New technologies attracted enormous attention because their demonstrations were immediate and visible. Their practical limitations, institutional barriers, and slow adoption were less exciting. Yet once the hype faded, those technologies often continued spreading until they quietly transformed everyday life.
Early computer networking offered a striking example. ARPANET had been developed mainly for sharing information among research and defense institutions, but scientists soon began using it for ordinary communication. Email was initially dismissed by some as a misuse of expensive computing resources; decades later, it became one of the basic infrastructures of modern society.
Amara’s own approach to forecasting reflected this experience. In 1977, he argued that simply extending current trends was not enough: futurists had to examine available choices and the consequences of those choices. In a later paper, he reviewed the Institute for the Future’s forecasting “hits and misses” and warned against rigid point predictions, emphasizing instead the forces that shape change and the need to revise forecasts as new information appears.
The precise first published appearance of the famous quotation is difficult to establish, although it is commonly associated with Amara’s work in the late 1970s. It is therefore better understood not as a formal law discovered on a particular date, but as the condensed lesson of a professional lifetime spent observing how technological change actually unfolds.
Amara’s Law survives because it describes more than technological hype. It captures the mismatch between the speed of human expectations and the much slower process through which inventions acquire infrastructure, users, institutions, and social power. Artificial intelligence is only the latest—and perhaps most dramatic—example.
The History of AI Is Amara’s Law in Motion
Artificial intelligence did not begin with ChatGPT.
The field was formally named in a 1955 proposal for the Dartmouth Summer Research Project on Artificial Intelligence, held the following year. Its organizers proposed investigating whether aspects of learning, language, reasoning, abstraction and creativity could be described precisely enough for machines to simulate them.
The ambition was enormous. Progress proved much slower.
Early AI researchers produced programs capable of proving theorems, solving constrained problems and manipulating symbols. These systems inspired predictions that machines capable of broad human-level intelligence might be only a generation away.
They were not.
From Symbolic AI to the AI Winters
For decades, symbolic AI dominated the field. Intelligence was represented through rules, logical structures and carefully encoded knowledge. Expert systems later demonstrated that computers could outperform or assist specialists inside narrowly defined domains.
The results were real. The extrapolations were not.
Systems that worked impressively inside controlled environments often failed when confronted with ambiguity, incomplete information or ordinary common sense. Building and maintaining large collections of rules was expensive. Computers remained limited, data were scarce, and many systems could not adapt beyond the knowledge explicitly given to them.
Investment and enthusiasm eventually declined. Periods of disappointment and reduced funding became known as the AI winters. Symbolic AI had been expected to produce general intelligence quickly; instead, it revealed how difficult intelligence actually was.
Yet AI did not disappear. Research continued beneath the surface.
That quiet period is crucial to Amara’s Law. Public attention tends to measure technological progress through visible products and dramatic demonstrations. Long-term transformation, however, depends on less glamorous foundations: better algorithms, larger datasets, faster processors, cheaper storage, network infrastructure and generations of trained researchers.
While public expectations cooled, those foundations accumulated.
The Deep-Learning Breakthrough
By the 2000s and early 2010s, several previously separate developments began reinforcing one another: digitized data expanded, graphics processors made parallel computation more accessible, and neural-network training methods improved.
In 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton demonstrated a large deep convolutional neural network on the ImageNet image-classification challenge. Their system achieved a major improvement over previous approaches and helped make deep learning central to modern AI research.
The breakthrough did not mean that machines suddenly understood the visual world. It meant that a long-underestimated research direction had crossed an important practical threshold.
The pattern repeated in 2016, when AlphaGo defeated a professional Go champion using deep neural networks, reinforcement learning and tree search. Go had long been considered exceptionally difficult for computers because of its enormous search space. The achievement arrived earlier than many specialists had expected.
In 2017, the paper Attention Is All You Need introduced the Transformer architecture. By using attention mechanisms instead of relying entirely on recurrent or convolutional structures, Transformers enabled far more effective parallel training for language tasks. That architectural change became a foundation for the large language models that followed.
None of these developments alone created the present AI revolution. Together, they formed a compounding chain.
ChatGPT and the Return of Maximum Hype
On November 30, 2022, OpenAI released ChatGPT as a public research preview. Unlike earlier AI systems hidden inside recommendation engines, search tools and enterprise software, ChatGPT placed a powerful language interface directly in front of ordinary users.
People could converse with a machine, request explanations, generate code, summarize documents and produce apparently original text.
AI was no longer an abstract promise. It had become an experience.
OpenAI’s original announcement also described important limitations: ChatGPT could produce convincing but incorrect answers, react unpredictably to small changes in wording and generate different answers to the same question. Those weaknesses were visible from the beginning, even as public discourse often ignored them.
The release triggered a new wave of extreme predictions. Artificial general intelligence was said to be imminent. Search engines, schools, programmers, artists, journalists, lawyers and entire categories of office workers were declared obsolete—sometimes within months.
Once again, the demonstration was real. The timeline was speculative.
The Short-Term Overestimation of AI
The short-term mistake is not believing that AI matters. It is confusing a convincing demonstration with a complete, reliable and economically deployable system.
A model that performs a task in a controlled test is not automatically ready to assume responsibility inside a hospital, court, school, bank or industrial process. Deployment requires validated data, security, integration, monitoring, accountability, legal clarity and human supervision.
Generative AI systems remain uneven. They can solve difficult mathematical or programming problems and then fail on apparently simple tasks. They can produce brilliant explanations beside fabricated references. They may appear coherent without possessing a stable model of truth.
The 2026 Stanford AI Index describes this as a jagged frontier. Frontier systems achieved striking results in areas such as competition mathematics, scientific reasoning and software engineering, yet remained unreliable on other tasks. AI agents reached roughly 66 percent success on a benchmark involving real computer operations—an enormous improvement, but still a failure rate of approximately one in three attempts.
That is not sufficient reliability for many high-consequence applications.
The short-term overestimation of AI therefore produces several recurring errors:
Companies mistake access to a language model for an AI strategy.
Managers automate poorly designed processes instead of redesigning them.
Investors treat benchmark improvements as guaranteed profits.
Commentators mistake fluent language for human-equivalent reasoning.
Institutions deploy systems before establishing who is responsible when they fail.
The gap between an impressive prototype and a dependable institution is where much of the current hype will die.
The Long-Term Underestimation of AI
The opposite mistake may be even greater.
When AI fails to deliver every promised transformation immediately, skeptics conclude that the entire field is another bubble. But technologies do not need to become conscious, perfect or universally autonomous to reshape civilization.
The internet did not transform society because every website worked perfectly. Smartphones did not need human intelligence to reorganize communication, commerce and attention. AI does not need to become a digital person before it begins restructuring institutions.
By 2026, Stanford reported that organizational AI adoption had reached 88 percent and that generative AI had reached approximately 53 percent population-level adoption within three years—faster than either the personal computer or the internet.
The most consequential effects may emerge not from a single superintelligent machine but from millions of ordinary AI systems embedded in routine decisions.
AI will increasingly influence which applicants receive interviews, which students are flagged as struggling, which patients receive additional screening, which financial transactions appear suspicious, which information becomes visible and which actions are considered normal.
These systems will not merely automate existing institutions. Over time, institutions will be redesigned around them.
That is the long-term side of Amara’s Law.
AI and the Future of Work
Predictions of immediate mass unemployment illustrate short-term overestimation. Claims that AI will leave employment largely unchanged illustrate long-term underestimation.
The decisive unit is often not the occupation but the task.
A lawyer may continue to exist while spending less time searching documents. A programmer may remain employed while writing less routine code and reviewing more machine-generated work. A teacher may continue teaching while AI systems generate exercises, analyze performance and personalize explanations.
The International Labour Organization estimated in 2025 that one in four workers worldwide had an occupation with some exposure to generative AI. It concluded that because most occupations contain tasks requiring human input, job transformation was more likely than complete redundancy. Exposure was also distributed unequally across professions, countries and genders.
This does not make the disruption harmless.
Automation can remove entry-level tasks through which inexperienced workers once learned a profession. Productivity gains can be concentrated among owners rather than employees. Workers may be monitored, evaluated or disciplined by systems they cannot inspect. Entire professions may retain their names while losing bargaining power, autonomy or social status.
The long-term question is not simply how many jobs AI destroys. It is how AI changes the content, value and control of human work.
AI as Institutional Infrastructure
The deepest technological transformations eventually become boring.
Electricity disappeared into walls. Digital networks disappeared into nearly every business process. Search engines became an invisible layer between humanity and knowledge.
AI may follow the same trajectory.
Today, people consciously open an AI application and write a prompt. Tomorrow, AI systems may be embedded throughout operating systems, vehicles, laboratories, classrooms, factories, hospitals and government services. People may interact with hundreds of models without knowing their names—or even realizing that a model made part of the decision.
At that point, the most important AI systems may no longer resemble chatbots.
They may be prediction layers, automated negotiators, scientific assistants, surveillance mechanisms, allocation engines and semi-autonomous agents operating inside larger technical and bureaucratic systems.
The transformation will become less theatrical and more structural.
Amara’s Law Is Not an Argument for Passivity
Amara’s Law is sometimes misinterpreted as advice to ignore technological hype and wait patiently for the future.
That would be dangerous.
If society underestimates long-term effects, postponing governance until those effects are obvious may leave institutions permanently behind. Infrastructure becomes difficult to change once organizations, markets and populations depend on it.
Responsible decisions therefore require two forms of discipline at the same time:
Do not believe every short-term promise.
Do not dismiss the long-term transformation.
NIST’s Generative AI Profile reflects this lifecycle approach by encouraging organizations to incorporate trustworthiness and risk management into the design, development, deployment and evaluation of AI systems rather than treating safety as an afterthought.
Good governance should not attempt to freeze AI at its current stage. It should create institutions capable of adapting as capabilities, uses and risks evolve.
How to Apply Amara’s Law to AI Strategy
Leaders should begin by separating technical capability from social diffusion.
A model may become technically capable of performing a task years before organizations can integrate it reliably. Conversely, a system does not need frontier-level intelligence to spread rapidly when it is cheap, convenient and embedded in widely used platforms.
Forecasts should therefore use several time horizons.
Over the next two years, decision-makers should examine practical reliability, integration costs, data quality and measurable returns.
Over five years, they should consider workflow redesign, changing skill requirements, new competitors and the disappearance of routine tasks.
Over ten or twenty years, they should ask how AI could alter institutions, education, scientific discovery, political power, inequality and the distribution of human agency.
The final question should not be “Will this model replace people?”
It should be:
What kind of system will emerge when people, organizations and AI adapt to one another for an entire generation?
Conclusion: The Hype Is Wrong About the Clock
Artificial intelligence may not eliminate most professions next year. Fully autonomous agents may not reliably manage businesses, conduct science or operate critical infrastructure as quickly as their promoters suggest. Artificial general intelligence may remain conceptually ambiguous and technically elusive.
That is the short-term correction.
The long-term correction is more unsettling.
AI does not need to become omniscient to reorganize knowledge. It does not need to become conscious to redistribute power. It does not need to replace every worker to weaken professions, eliminate entry points and transform the value of expertise. It does not need to govern humanity openly if institutions quietly become dependent on its recommendations.
Amara’s Law does not tell us that artificial intelligence is overhyped.
It tells us that we are probably looking at the wrong timescale.
We may be exaggerating what AI will accomplish by 2028 while failing to imagine how completely it could reshape civilization by 2046.
The immediate future will probably be less revolutionary than the headlines promise.
The distant future may be far more revolutionary than we are prepared to admit.
Frequently Asked Questions
What is Amara’s Law?
Amara’s Law is the observation that people tend to overestimate the impact of a new technology in the short term while underestimating its impact over the long term.
Who created Amara’s Law?
The principle is named after Roy Charles Amara, an American engineer, futurist and longtime leader of the Institute for the Future.
How does Amara’s Law apply to artificial intelligence?
AI capabilities are frequently expected to transform industries immediately. Real deployment is slowed by reliability, integration, regulation and organizational limitations. Over longer periods, however, AI may quietly restructure work, education, science and institutional decision-making.
Is artificial intelligence overhyped?
Some short-term claims are overhyped, especially predictions of immediate universal automation or perfectly autonomous AI agents. However, this does not mean that AI’s long-term impact is exaggerated. Its cumulative social and institutional effects may be underestimated.
What is the best historical example of Amara’s Law in AI?
The history of neural networks is a strong example. After periods of high expectations and disappointment, neural-network research continued developing alongside better hardware, larger datasets and improved algorithms. It later became the foundation of deep learning and generative AI.
References and Further Reading
Amara, Roy. “What We Have Learned About Forecasting and Planning.” Futures, 1988.
Institute for the Future. “Amara’s Law.”
McCarthy, John, Marvin Minsky, Nathaniel Rochester and Claude Shannon. “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence,” 1955.
Krizhevsky, Alex, Ilya Sutskever and Geoffrey Hinton. “ImageNet Classification with Deep Convolutional Neural Networks,” 2012.
Silver, David, et al. “Mastering the Game of Go with Deep Neural Networks and Tree Search.” Nature, 2016.
Vaswani, Ashish, et al. “Attention Is All You Need,” 2017.
International Labour Organization. “Generative AI and Jobs: A 2025 Update,” 2025.
Stanford Institute for Human-Centered Artificial Intelligence. AI Index Report 2026.
National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, 2024, updated 2026.
Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.
Copyright 2026 AI-Talks.org
