The AI Time Horizon Problem: Why We May Overreact to the First Shocks of AI and Underreact to Its Deeper Transformation

Mark Howard

— 11 min read
Share

New technologies are hard to see clearly when they first arrive.

Part of the problem is that they do not arrive all at once. They arrive first as demonstrations, products, headlines, fears, promises, market valuations, workplace experiments, and political arguments. Only later do they become routines, institutions, infrastructure, business models, habits, dependencies, and power structures.

This creates what we might call the AI time horizon problem.

When a powerful technology appears, people naturally judge it by what they can see. And what they can see first are the early shocks: job displacement, scams, cheating, misinformation, data centers, surveillance, strange new products, and unsettling demonstrations of machine capability.

Those concerns are not imaginary. Many are real. But they are also incomplete.

The deeper effects of a general-purpose technology usually unfold more slowly. They depend on adoption, complementary investment, institutional redesign, public policy, legal rules, social norms, labor-market adjustment, and new business models. The first wave tells us that something important has arrived. It does not tell us where society will end up.

That distinction matters for The Race.

The central question of The Race is whether AI will concentrate power or disperse capability. But that question cannot be answered only by looking at today’s most visible disruptions. The more important question may be whether today’s temporary advantages are beginning to harden into tomorrow’s infrastructure.

That is where the time horizon problem becomes a democratic problem.

Amara’s Law and the Wrong Clock

The best-known version of this idea is often called Amara’s Law, usually attributed to futurist Roy Amara: we tend to overestimate the effect of a technology in the short run and underestimate its effect in the long run.

The phrase is useful. But it should not be treated as a law of nature. It is better understood as a warning about attention.

In the short run, new technologies are easy to overstate because they are novel, vivid, and surrounded by hype. Their most dramatic examples travel fastest. A chatbot passes a bar exam. A student uses AI to write an essay. A company announces an AI-related layoff. A fake image goes viral. A new data center strains a local power grid.

These events are concrete. They can be shown, shared, feared, and debated.

Long-term transformation is harder to see. It is slower, more distributed, and often hidden inside institutions. It appears in procurement rules, software defaults, worker training, business processes, school policies, insurance practices, developer ecosystems, and regulatory assumptions. It is not one dramatic event. It is a thousand quiet decisions.

That is why democratic institutions can be pulled in two bad directions at once.

One danger is panic. If we treat the first visible shocks as the whole story, we may respond with fear, bans, symbolic regulation, or a politics of technological rejection.

The other danger is complacency. If the first wave looks messy, uneven, overhyped, or disappointing, we may assume the deeper transformation is not really happening.

Both reactions miss something important.

The first shocks of AI may be overinterpreted. But the long-term restructuring may still be underestimated.

Technology Panics Are Historically Common

Public anxiety around new technologies is not new.

Novels, newspapers, telephones, radio, television, video games, the internet, smartphones, and social media have all generated waves of concern about attention, morality, youth, truth, social connection, violence, labor, and democratic life.

The point is not that those concerns were always wrong. Sometimes they identified real harms. Sometimes they were early warnings. Sometimes they were exaggerated. Often they were all three at once.

But the pattern is familiar. A new technology appears. People focus on its most visible disruptions. Research and policy rush toward the loudest anxieties. Public debate narrows around the most emotionally charged examples. Then, years later, society discovers that the technology’s deepest effects were more complicated than the first panic suggested.

Amy Orben describes this as the “Sisyphean cycle of technology panics”. Her point is especially useful for thinking about AI. The problem is not concern itself. A democratic society should be concerned about powerful technologies. The problem is when concern becomes too narrow, too repetitive, or too focused on immediate moral shock to guide useful institutional action.

AI is already producing its own version of this cycle. Public attention moves quickly from one visible fear to another: students cheating, artists being displaced, workers being replaced, deepfakes spreading, chatbots hallucinating, models manipulating users, or data centers consuming land, power, and water.

These are serious issues. They deserve attention. But they are also the front edge of a larger transformation.

The democratic task is not to wave away the early fears. It is to ask what they reveal — and what they may be causing us to miss.

Why Short-Term Risks Feel So Large

Risk-perception research helps explain why the first shocks of technology can feel so large.

Paul Slovic’s work on the perception of risk shows that people do not evaluate risk only by calculating probabilities. They also respond to qualities such as dread, lack of control, catastrophic potential, unfamiliarity, and whether a risk seems imposed rather than chosen.

AI has many of these qualities.

It is opaque. It can feel uncontrollable. It is associated with job loss, deception, surveillance, and the loss of human agency. It is being deployed by powerful firms and institutions that many people do not fully trust. Its harms can feel diffuse and difficult to trace. Its capabilities can feel almost magical, which makes both utopian and catastrophic stories easier to believe.

So public anxiety about AI should not be treated simply as ignorance. It may be registering something real: a fear that decisions affecting people’s lives are moving into systems they cannot understand, challenge, or control.

That fear deserves respect.

But it also needs direction. A society can spend a great deal of energy reacting to the most visible dangers while still failing to govern the systems that produce them.

For The Race, the distinction is crucial. We need to ask two questions at the same time.

What visible harms are appearing now?

And what deeper structures of power are forming underneath them?

A healthy democratic response needs both.

Historical Example: Electrification

Electrification is one of the clearest historical examples of the time horizon problem.

Electricity did not transform the economy simply because the electric motor was invented. The technology had to diffuse through factories, homes, cities, transportation systems, appliances, lighting, and industrial processes. Businesses had to redesign production around electric power. Cities had to build generating and distribution systems. Households had to adopt appliances. Workers had to learn new routines. Complementary technologies had to mature.

Paul David’s famous comparison between the electric dynamo and the computer is useful here. His point was that transformative technologies often appear before their productivity effects become visible in economic statistics. In the early stages, the technology can seem both impressive and underwhelming. The breakthrough exists, but society has not yet reorganized around it.

That is a helpful caution for the present moment.

If AI is a general-purpose technology, then its deepest effects will not come only from individual people using chatbots. They will come from the redesign of organizations, markets, public agencies, education systems, legal processes, infrastructure, and everyday interfaces.

That kind of transformation takes time. It also takes choices.

Electrification was not just a story about invention. It was a story about systems: who built them, who governed them, who paid for them, who gained access, and who captured the benefits.

AI will be a systems story too.

Historical Example: The Computer and the Internet

The computer age followed a similar pattern.

For years, computers spread through business and government without immediately producing the sweeping productivity gains many expected. This gave rise to the “productivity paradox,” often summarized by Robert Solow’s observation that computers were everywhere except in the productivity statistics. The Richmond Fed’s discussion of the productivity puzzle uses that observation to frame the difficulty of tracing new technology into broad economic measures.

The paradox was not simply that computers failed to matter. It was that their impact required complementary change: new software, new management systems, new supply chains, new skills, new organizational forms, and new ways of measuring value.

The internet followed a related path. Early internet debate focused on novelty, speculation, pornography, chat rooms, online shopping, piracy, and the dot-com bubble. Those were real parts of the story. But they were not the whole story.

The deeper transformation came later.

Search became the doorway to information. Platforms became the infrastructure of communication and commerce. Social media altered political life. Cloud computing changed firm structure. Smartphones merged identity, location, payment, communication, and attention into a single device. Advertising markets reorganized around data.

Many of these long-term effects were not obvious from the first wave of websites.

That is the lesson worth carrying forward. The chatbot may not be the final form of AI. It may be the first familiar doorway.

The deeper question is what sits behind the doorway: models, compute, data, agents, cloud infrastructure, operating systems, procurement contracts, workplace systems, educational platforms, and civic institutions.

General-Purpose Technologies and Complementary Change

The reason these patterns recur is that some technologies are not merely new tools. They are general-purpose technologies.

Economists Timothy Bresnahan and Manuel Trajtenberg described general-purpose technologies as technologies that are pervasive, improve over time, and generate complementary innovation across many sectors. Elhanan Helpman and Trajtenberg later explored how growth driven by general-purpose technologies can depend on waves of improvement, diffusion, and adaptation rather than a single moment of invention.

That matters because AI is increasingly discussed as a general-purpose technology. It is not confined to one industry or one task. It can be used in law, medicine, education, finance, public administration, customer service, software development, logistics, media, science, and everyday personal work.

But general-purpose technologies do not transform society by existing. They transform society when people and institutions reorganize around them.

That is where the democratic stakes appear. The question is not only who invents the technology. It is who has the capacity to use it well, who sets the terms of access, who controls the infrastructure, who shapes the rules, and who captures the gains.

The Productivity Paradox and AI

Brynjolfsson, Rock, and Syverson argue that AI may present a modern version of the productivity paradox: a clash between impressive technical capability and slow, uneven appearance in aggregate economic statistics. In Artificial Intelligence and the Modern Productivity Paradox, they identify several possible explanations, but emphasize implementation lags as especially important.

That is exactly the kind of lag The Race should pay attention to.

Implementation lags are not empty time. They are the period when organizations figure out how to use the technology. They are also the period when power can shift.

During a lag, firms choose vendors. Agencies sign contracts. Workers are trained, augmented, monitored, or displaced. Schools select platforms. Developers build around particular APIs. Users accumulate data and habits inside particular systems. Regulators define categories. Courts interpret rights. Standards bodies make choices. Public institutions either build capacity or become dependent on private infrastructure.

In other words, the lag is not a pause before the real story begins.

The lag is where much of the real story gets written.

Brynjolfsson, Rock, and Syverson develop this point further in their work on the Productivity J-Curve, arguing that general-purpose technologies such as AI require complementary intangible investments — business process redesign, new products, new business models, and human capital — before their benefits become visible.

This is why AI’s uneven early productivity effects should not be taken as proof that AI is simply overhyped. They may instead mean that the complementary changes have not yet fully arrived. And when they do arrive, the distribution of benefits will depend heavily on choices made earlier.

This is the bridge from the time horizon problem to the window of contestability.

The Democratic Risk: Looking at the Wrong Part of the Curve

The AI debate is often pulled toward two extremes.

One side says AI is about to change everything immediately. The other says the technology is mostly hype because the promised transformation has not yet appeared at scale.

The time horizon problem suggests that both claims may be wrong.

AI may be overhyped in the short term and still underappreciated in the long term. The first business uses may be clumsy. Many tools may disappoint. Some companies may exaggerate their adoption. Some productivity gains may be hard to measure. The first consumer products may feel like toys. Many startups may fail.

And yet, beneath that noisy surface, important things may be hardening.

A few firms may gain durable control over compute. A few platforms may become default interfaces to information. A few cloud providers may become the operating layer for AI adoption. A few model ecosystems may become embedded in enterprise workflows. A few agents may accumulate personal and organizational memory. A few procurement patterns may make public agencies dependent on proprietary systems. A few standards may quietly determine who can compete.

That is the democratic risk.

Not that every frightening AI headline is correct.

Not that everything has already been decided.

But that society may spend the early years arguing about the most visible shocks while the deeper architecture of power is being built elsewhere.

The Window of Contestability

The time horizon problem points toward a second concept: the window of contestability.

The window of contestability is the period after a powerful technology emerges but before its dominant platforms, standards, defaults, business models, and institutional dependencies become locked in.

During this window, democratic institutions have more leverage. Public policy can shape procurement. Standards can preserve interoperability. Public agencies can build internal capacity. Workers can be supported through transition. Communities can negotiate infrastructure impacts. Schools can choose tools that protect student agency. Markets can remain open to smaller firms. Civic institutions can define public-interest uses before private platforms define them for everyone.

After the window narrows, action is still possible. But it becomes harder.

Once systems are embedded, changing them is costly. Once users depend on a platform, exit becomes difficult. Once public services are built around proprietary infrastructure, oversight becomes weaker. Once developers build for a dominant ecosystem, alternatives struggle. Once defaults become habits, people stop seeing them as choices.

This is why the time horizon problem is not just a communications problem. It is a governance problem.

Democratic institutions likely have more time than panic suggests, but less time than complacency assumes.

What This Means for The Race

The Race is not a project about whether AI is good or bad. It is a project about direction.

Does AI concentrate power, or does it disperse capability?

The AI time horizon problem helps explain why that question is difficult to answer in real time. The early evidence will always be mixed. Some developments will expand access. Others will deepen dependence. Some tools will help small organizations. Others will strengthen dominant platforms. Some public uses will build civic capacity. Others will outsource public judgment to private systems.

The task is not to predict the future with false precision. The task is to watch the hardening process while it is still happening.

That means asking different questions:

Are people gaining meaningful new capabilities, or are they becoming dependent on systems they cannot understand or leave?

Are public institutions building capacity, or are they becoming customers of private intelligence infrastructure?

Are workers being augmented, or are firms using AI mainly to reduce labor power?

Are standards open, or are dominant platforms defining the rules?

Are communities shaping AI deployment, or are they merely reacting to decisions made elsewhere?

Are today’s AI tools becoming tomorrow’s infrastructure?

Those questions move the discussion beyond hype and panic. They focus attention on the choices that will determine whether AI strengthens democracy or weakens it.

Conclusion: Do Not Panic, But Do Not Wait

The lesson of Amara’s Law is not that early fears are foolish. Nor is it that long-term transformation is inevitable. The lesson is that time matters.

New technologies are often misunderstood because society looks at the wrong clock.

The first clock measures shock: the headline, the demonstration, the layoff, the scandal, the product launch.

The second clock measures transformation: diffusion, dependency, institutional redesign, legal settlement, market structure, and power.

AI is now moving on both clocks.

The short-term shocks deserve attention. But the deeper democratic question is whether the systems forming around AI remain contestable.

That is why this paper is only the first step in a larger series. The next paper will describe the window of contestability more clearly: what it is, how it works, and why democratic institutions have more leverage before defaults, contracts, standards, habits, and dependencies harden. A third paper will then develop a set of contestability indicators to help us watch for the signs that today’s temporary advantages are becoming tomorrow’s durable structures of power.

The danger is not only that we will fear AI too much.

The danger is that we will fear the wrong things too loudly, and notice the deeper hardening too late.

References and Further Reading

Read more