After Prediction: How Cheap AI Creates Categories That Could Not Exist Before

Most public accounting of artificial intelligence begins with a subtraction. A task once performed by a person is performed by a machine, a job is threatened or lost, and the significance of the technology is measured by the size of the loss. The measure is not wrong. It is early. It captures the first visible consequence of a falling input price and mistakes it for the whole event.

This paper argues for a different governing claim:

The power of AI is not primarily the replacement of human labor. It is the creation of new categories as the cost of prediction approaches zero.

Labor replacement is a first-order substitution effect. A machine performs a prediction that a human previously performed, and the existing task becomes cheaper. That matters economically, and for the people displaced it matters personally. But substitution preserves the old category. The car is still a car, the warehouse is still a warehouse, the clinic is still a clinic — each simply staffed by fewer people. The deeper transformation begins when prediction becomes cheap, continuous, and abundant enough that products, institutions, physical assets, markets, and behavior are redesigned around it. At that point entirely new categories become possible, and the old nouns begin to lose their shape.

The movement this paper traces has seven steps. First, a scarce human prediction is identified. Second, a machine produces that prediction more cheaply. Third, the price of the existing task falls. Fourth, the value of complementary inputs rises. Fifth, the system is redesigned around abundant prediction. Sixth, a new economic category appears. Seventh, prediction becomes invisible inside the new category. The labor-replacement story ends near step three. The category-creation story begins after it.

The argument proceeds from the historical analogy of cheap arithmetic (Section 2) and the economics of a falling prediction cost (Section 3) to a test for genuine category creation (Section 4), then to two prospective case studies — autonomous mobility and predictively positioned supply chains (Sections 5 and 6) — and their consequences for labor and accountability (Sections 7 and 8), before confronting counterarguments (Section 9) and concluding.

A note on method. This paper mixes four kinds of claim and tries to keep them separate: documented present conditions, current experimental or early commercial evidence, consequences derived from an economic framework, and speculative second- and third-order scenarios. Future-tense claims are deliberately conditional. The framework predicts; it does not guarantee.

Before the computer was a machine, it was a job title. Banks, observatories, laboratories, and aerospace programs employed people — at NASA’s Jet Propulsion Laboratory, mostly women — whose occupation was calculation: computing trajectories, reducing data, tabulating results (NASA, n.d.). As electronic machines took over the arithmetic, the occupation dissolved into new technical roles, and by about 1970 the human “computer” had largely been retitled into engineering and programming work (Holt, 2016).

The scale of what followed is difficult to overstate. Nordhaus (2007) estimates that computer performance improved relative to manual calculation by a factor of between 1.7 trillion and 76 trillion, with a roughly two-trillion-fold improvement in constant-dollar cost, most of it after 1945. Arithmetic — once a scarce professional service — became effectively too cheap to meter.

The first wave of cheap arithmetic was pure substitution. Machines calculated payrolls, ballistics tables, and engineering stresses that people had calculated before, only faster. That ledger — human calculators eliminated, calculation cheaper — would be accurate and radically incomplete, because cheap arithmetic did not stay inside the calculator.

Digital photography is the clearest demonstration. A digital photograph exists because light can be represented as numbers and because enormous quantities of inexpensive computation can then be applied to those numbers: capture, demosaicing, color correction, compression, storage, copying, transmission, search, and editing. When Steven Sasson assembled the first self-contained digital camera at Eastman Kodak in December 1975, it recorded a 0.01-megapixel black-and-white image to cassette tape in 23 seconds (Estrin, 2015). Every stage of that pipeline, and of the modern pipeline that descends from it, is arithmetic performed at a price no human calculator could approach.

Two disciplines are required here. First, arithmetic alone did not cause digital photography. Image sensors, optics, memory, batteries, displays, networks, and manufacturing were all necessary; cheap computation was a foundational enabling input within a larger technological system, not a sole cause. Second, the transformation cannot be measured by the labor it displaced. Kodak’s global employment peaked at about 145,300 in 1988; the company entered Chapter 11 bankruptcy in January 2012 (Rochester Business Journal, 2017). Darkroom and film-processing work declined with it. But counting those jobs explains almost nothing about what happened to photography. At the peak of the film era, in 1999, roughly 80 billion photographs were taken worldwide (Kodak estimate, cited in Evans, 2015). Analysts now project on the order of two trillion photographs per year, about 94 percent of them taken on smartphones (Broz, 2025) — devices that shipped 1.24 billion units in 2024 alone (International Data Corporation, 2025).

The significance of digital photography is what an image became. Instantly reviewable, perfectly reproducible, editable, searchable, transmissible, and embedded in software, the photograph stopped being a scarce chemical artifact and became a computational object — the raw material of new industries, new behaviors, and new institutions that film chemistry could not have supported. The category changed. And, crucially, the arithmetic disappeared from view. No one experiences a smartphone camera as arithmetic, although arithmetic is nearly all it is.

This gives the paper its governing parallel:

Cheap arithmetic changed what information could be. Cheap prediction changes what systems can become.

The most important AI products will eventually stop looking like prediction, just as digital photography does not look like arithmetic to its users. When that happens, the temptation will be to conclude that the prediction revolution ended. The correct conclusion will be that it succeeded.

The economic identity of modern artificial intelligence is best captured in a single sentence: recent advances in machine learning amount to a quality-adjusted drop in the price of prediction (Agrawal et al., 2018, 2019a). Prediction — filling in missing information from the information one has — is an input into decision-making under uncertainty. It has always been economically ubiquitous — a doctor predicts disease, a dispatcher predicts arrival, a driver predicts the motion of a pedestrian, an underwriter predicts loss — and until recently it was locked inside the scarcest general-purpose instrument available: human attention and judgment.

When the price of a useful input falls by orders of magnitude, two things happen. The first is substitution: machine prediction replaces human prediction in existing tasks. The second, historically larger, is demand expansion: the input is applied in places where it was previously too expensive to use at all. This is the oldest pattern in the economics of technology. Jevons (1865) observed it with coal-burning efficiency; Bessen (2015) documented it with automated teller machines, which cut the number of tellers per bank branch while the number of branches — and, for decades, total teller employment — rose. Cheap arithmetic followed the same path: the job title disappeared while the operation multiplied.

Radiology offers a live illustration in the prediction era. In 2016, Geoffrey Hinton suggested that training of radiologists should stop, because deep learning would soon outperform them (Mukherjee, 2017). A decade later, the U.S. Food and Drug Administration had authorized more than 1,400 AI-enabled medical devices, roughly three-quarters of them in radiology (U.S. Food and Drug Administration, 2026) — and the American College of Radiology was describing a persistent radiologist shortage, with a 2025 workforce study counting 37,482 radiologists serving Medicare beneficiaries in 2023 and projecting substantial workforce growth through 2055 under every modeled scenario (Christensen et al., 2025; Rula, 2026). Cheaper image interpretation did not shrink the demand for interpretation; it expanded the volume of imaging and moved the bottleneck to clinical judgment, integration, and accountability. The prediction layer expanded the service layer.

Two conceptual distinctions govern everything that follows.

First, prediction is not action. A prediction machine predicts; an agent acts. Cheap prediction changes a real system only when it is joined to authority, physical execution, capital, infrastructure, energy, verification, and feedback from the world. This is why the value of complements rises as the price of prediction falls: better prediction raises the return on the judgment that sets objectives and on the action that carries predictions into reality (Agrawal et al., 2019b, 2022).

Second, “approaches zero” is an economic threshold, not a metaphysical claim. Prediction never becomes literally free; models are expensive to train, run on energy-hungry hardware, and degrade outside their training distributions. The threshold that matters is the point at which prediction stops being the binding constraint in system design — when designers can assume prediction the way an architect assumes electricity. Once that threshold is crossed, scarcity does not disappear. It relocates.

Every case examined below therefore asks the same two questions. Once prediction becomes abundant, what becomes scarce? And what genuinely new category becomes possible?

Substitution and category creation are easy to conflate, because every new category passes through a substitution phase on its way into existence. A test is needed.

The weak version of any AI story preserves the old noun: a car without a driver, a warehouse with fewer dispatchers, a clinic with fewer diagnosticians. These may be real and consequential outcomes, but they are not the destination of this paper. The strong version shows the noun itself becoming inadequate: transportation becomes guaranteed movement, the vehicle becomes programmable space, inventory becomes a location-flexible service level.

A genuine new category should satisfy most of the following six conditions:

1. It would have been uneconomic or operationally impossible while prediction was expensive.
2. Prediction is continuous and embedded rather than sold as a separate forecast.
3. The customer buys a new outcome rather than a cheaper version of the old task.
4. The product’s geometry, ownership, timing, or institutional boundary changes.
5. New complements, scarcities, and failure modes become economically central.
6. The category remains valuable after the novelty of “AI” disappears from its description.

The final condition deserves emphasis, because it is the one digital photography teaches. The technology succeeded so completely that users stopped thinking about the computation underneath it. A category still marketed primarily as “AI-powered” is a category still in its substitution phase. If a case remains “the same product with fewer workers,” it has not yet demonstrated the thesis. The two case studies that follow apply this test explicitly.

5.1 The old bundle

The automobile is designed around a human prediction machine. A driver supplies a bundle of functions: perception, continuous prediction, physical control, legal responsibility, exception handling, and — in commercial service — customer care. The prediction component is not “how to drive” in the abstract but a continuous family of narrow forecasts: where each nearby object is likely to move, which path is traversable, what another road user is about to do, whether a maneuver will remain within an acceptable risk boundary.

Much of the vehicle’s geometry and much of the surrounding urban form exist because this prediction machine is human. Seats face forward around a control position. Visibility, instrumentation, and crashworthiness are organized around a person who must see and steer. Ownership is attractive partly because reliable transportation requires a vehicle and a willing driver to be in the same place at the same time — and the cheapest way to guarantee a driver is to be one. The parked car is the physical residue of that bundling: a machine that must wait wherever its one authorized prediction engine happens to be.

5.2 The substitution stage

Driver assistance is substitution: the human still drives, aided at the margin. A driverless version of the conventional car is still largely substitution — the same steel box, the same trips, one input replaced.

That substitution stage is now commercially real, which is what makes the case worth writing. As of early 2026, Waymo reported roughly 500,000 paid robotaxi rides per week across ten U.S. metropolitan areas, expanding to an eleventh in July 2026 (TechCrunch, 2026; Electrek, 2026). Peer-reviewed analysis of its first 56.7 million rider-only miles found crash rates substantially below matched human benchmarks — 79 percent fewer any-injury-reported crashes, 81 percent fewer airbag-deployment crashes — with the caveat that the study’s authors are Waymo-affiliated (Kusano et al., 2025). Commercial driverless operation exists; it is early, geographically bounded, and concentrated in favorable conditions. On the framework’s terms, however, a robotaxi that looks like a car is still the old category. The interesting economics begin when designers stop treating the human control position as the organizing constraint.

5.3 The category break: the vehicle changes identity

The break arrives when the vehicle is redesigned around passengers rather than around a permanent human control position. This is no longer pure design fiction: in March 2022 the U.S. National Highway Traffic Safety Administration finalized amendments to its occupant-protection standards specifically to accommodate vehicles built without traditional manual driving controls (National Highway Traffic Safety Administration, 2022). A regulator does not rewrite crashworthiness standards for a vehicle class no one intends to build.

Once the control position is optional, the cabin can be organized around the purpose of the trip. A vehicle could become a mobile office, a clinic that travels to the patient, a bedroom that makes an overnight trip productive, a store or workshop that positions itself where demand is predicted. The framework’s economic movement is precise: cheap motion prediction reduces the value of the control position and raises the value of the space that replaces it.

Transportation itself can shift from vehicle ownership toward guaranteed access to movement. When a vehicle can reposition itself to the next user, the economically important product is no longer necessarily the car; it may be the assurance that movement is available when needed — a service level, not an object. Ownership does not vanish in this scenario, but it stops being the only way to secure reliable mobility, just as owning a generator stopped being the way most people secured electricity.

5.4 Urban land: measured, not mythologized

The urban stakes are real but must be stated with geographic precision. A popular claim holds that 14 percent of American cities is parking; careful research shows the truth is local and varied. A 2025 study of fifteen U.S. cities found off-street surface parking occupying between 3.4 percent (Oakland) and 10.7 percent (Anaheim) of parcel land, with central business district shares ranging from 2.3 percent in Boston to 31.7 percent in Tulsa (Qiam & Lehe, 2025). County-scale inventories found parking equivalent to about 14 percent of incorporated land in Los Angeles County — some 18.6 million spaces, 3.3 per registered vehicle (Chester et al., 2015) — and about 10 percent of regional land in metropolitan Phoenix (Hoehne et al., 2019). A study of seven Silicon Valley cities estimated roughly 14 percent of land devoted to parking (Gabbe et al., 2021). Across 102 U.S. city centers, the median share of land dedicated solely to parking is about 26 percent (Parking Reform Network, 2026a). Each figure counts something different — surface lots, all off-street parking, parcel land, central districts — and the ranges matter more than any single number.

What cheap repositioning changes is not the existence of parking but its binding logic. A vehicle that can move itself does not need to be stored beside its current user for most of the day. Parking then changes location, duration, and function: from ubiquitous adjacent storage toward consolidated staging, charging, and maintenance — some of it on cheaper peripheral land. Parking does not universally disappear; vehicles must still wait, charge, be cleaned, and recover from disruptions, and many existing structures will resist conversion for reasons of ceiling height, ramps, structure, and local demand. What can decline is the requirement that every destination hold storage for every visitor’s vehicle.

The curb changes character with it. Long-duration storage gives way to a high-turnover interface among passengers, buildings, deliveries, and services — a dynamically allocated transfer surface. Buildings may need fewer parking decks but more precisely managed pickup zones, loading bays, accessible boarding areas, and charging connections. Ingress becomes a routing problem rather than a storage problem.

5.5 The countervailing cost: empty movement

Every claim above must be netted against the cost of empty vehicle movement. A vehicle that avoids expensive central parking may consume road capacity and energy traveling empty to a peripheral lot, circulating, or repositioning toward predicted demand. Simulation studies treat this squarely: an early agent-based study found each shared autonomous vehicle could replace around eleven conventional vehicles but generated up to 10 percent additional travel distance (Fagnant & Kockelman, 2014); the International Transport Forum’s Lisbon simulations found total vehicle-kilometers ranging from 6 percent above baseline (shared rides with strong transit) to 89 percent above baseline (single-passenger vehicles without transit) (International Transport Forum, 2015). Later work treats repositioning, charging, waiting time, and parking as a joint optimization problem and shows that prediction helps: coupling repositioning with charging strategies cut empty travel meaningfully in Austin simulations (Dean et al., 2022), and demand-prediction dispatching reduced trip-level deadheading by 53 to 82 percent in ride-hailing datasets (Kontou et al., 2020). Cheap prediction can reduce empty travel by telling vehicles when waiting beats circulating. It cannot make empty travel free.

5.6 Branching futures: density or dispersion

The urban outcome is genuinely branched, and the branch point is not the prediction technology. One path combines shared autonomous fleets, curb and congestion pricing, strong transit, and deliberate parking redevelopment; it could support denser, more pedestrian urban centers. The other path combines privately owned autonomous vehicles, cheap empty repositioning, and a lower perceived cost of travel time; it could increase vehicle-miles and accelerate sprawl. The mechanism behind the second path is documented in miniature: in a naturalistic experiment that gave households a free chauffeur for a week — simulating the experience of not having to drive — vehicle-miles traveled rose 83 percent overall, with one-fifth of the increase coming from zero-occupancy trips, albeit in a small pilot of thirteen subjects (Harb et al., 2018). A person who can sleep or work through a commute may tolerate a much longer one.

Policy is therefore a co-author of the category. The quiet removal of parking minimums — roughly a fifth of more than 3,000 examined zoning codes have abolished or reduced parking mandates citywide, and over a hundred cities have removed all minimums (Parking Reform Network, 2026b) — is exactly the kind of institutional change that determines whether land released from storage becomes housing and public space or merely more vehicle circulation. The ownership model and public policy may matter as much as the prediction technology.

5.7 The test applied

Run the six conditions. A network of purpose-built, self-repositioning vehicles selling guaranteed movement and usable interior space (1) was operationally impossible while motion prediction required an onboard human; (2) embeds prediction continuously in every trajectory, dispatch, and staging decision rather than selling it as a forecast; (3) sells a new outcome — arrival, plus usable time and space in transit — rather than a cheaper taxi; (4) changes the vehicle’s geometry, the ownership model, and the institutional boundary between vehicle, curb, and building; (5) makes curb access, charging, fleet maintenance, remote exception handling, and liability the economically central scarcities; and (6) would remain valuable if no one ever again said “AI” about it — its users would simply say the city has rooms that move.

The new category, stated compactly: programmable physical space connected to guaranteed movement.

The new scarcities: reliable physical actuation, energy and charging access, curb and staging space, maintenance capacity, emergency handling, legal authority to operate, public trust, and resilience when maps, sensors, communications, or models fail. Safe prediction does not make movement free. The driver’s prediction becomes cheap; the value of the surrounding operating system rises.

Signals that would show the category forming: parking requirements removed from zoning codes; conversion of lots and garages, tracked by neighborhood; curb dwell time and pickup-conflict metrics; empty vehicle-miles per passenger-mile; charging and staging land per fleet vehicle; changes in commute distance and residential location; and whether released land becomes housing and public space or additional circulation. Signals that would weaken it: intervention rates that refuse to fall, operating domains that refuse to widen, and per-mile costs that keep the service a luxury.

6.1 The old system: inventory as purchased certainty

A warehouse is not a monument to wasted motion, and it is not primarily a consequence of expensive human driving. Inventory is a rational buffer against uncertainty — in demand, production, supplier reliability, transportation, and physical lead times. U.S. manufacturers hold on the order of 25 days of material inventory against delivery uncertainty and 28 days of finished goods against production and demand uncertainty (Thomas, 2023; Thomas & Kandaswamy, 2015). Storage is one way of purchasing certainty: goods are held so that a future order can be fulfilled even when prediction or execution fails.

Human driving adds real constraints — U.S. property-carrying drivers are generally limited to 11 hours of driving within a 14-hour window after 10 consecutive hours off duty (Federal Motor Carrier Safety Administration, 2025) — but driver time is only one component. Loading, unloading, ports, transfer facilities, receiver schedules, road capacity, maintenance, energy, and working capital shape the network just as firmly. The old supply chain is best understood as a system that uses inventory, time, information, and transportation as substitutes for one another under uncertainty.

6.2 The predictions

Supply-chain prediction is a family of linked forecasts: what will be demanded, where and when demand will occur, when a shipment will arrive, whether a supplier, machine, route, or port will fail, how long loading and transfer will take, and which inventory position best protects the desired service level at least cost. Cheap prediction allows these forecasts to be refreshed continuously and at fine resolution. But the machine still does not move the goods. Prediction must be joined to purchasing authority, production capacity, vehicles, actuators, roads, ports, energy, labor, and permission to reroute or substitute.

6.3 First-order effects — and the absolutes to reject

The first-order effects are better use of the existing network: routing closer to actual demand, arrival estimates that coordinate docks and labor, maintenance before likely failure, rebalancing before a local stockout. Autonomous line-haul operation relaxes one constraint on continuous movement; Aurora’s driverless trucks now complete thousand-mile Fort Worth–Phoenix runs in about fifteen hours — beyond what a single human driver’s regulated day allows (Korosec, 2026).

Several absolute claims should be rejected at this stage. Perpetual “rolling warehouses” are not automatically economical: a stationary container on inexpensive land usually consumes less energy and road capacity than a container circulating without a destination, and trucking’s marginal cost — about $2.26 per mile in 2024 (American Transportation Research Institute, 2025) — is a price paid per mile whether or not the motion creates value. Autonomous trucks will not achieve 100 percent utilization: a system at full utilization has no slack for peaks, maintenance, charging, or queue recovery, and waiting times explode as utilization approaches capacity. Removing driver-hour limits does not instantly double system capacity, because driver time is not the only binding constraint. And perfect just-in-time logistics does not eliminate delays: prediction can reveal a storm or port closure earlier, but it cannot conjure an alternate ship, transformer, or loading crew when every alternative is constrained. Prediction can reduce uncertainty about a delay without eliminating the delay.

Information can arrive at software speed. Capacity cannot.

6.4 The category break: predictive positioning

The category break is not a warehouse circling the highway. It is the conversion of inventory position itself into a continuously optimized decision variable — predictive positioning.

The elements are individually documented. Amazon’s anticipatory-shipping patent describes shipping goods toward a destination geographic area before the final delivery address is known, with the address completed while the package is in transit (Spiegel et al., 2013). Cross-docking — moving goods through a facility without letting them dwell as storage — has been a competitive weapon since Walmart ran most of its goods through warehouses that repacked and dispatched “often without ever sitting in inventory” (Stalk et al., 1992). Driverless line-haul freight is in early commercial service: Aurora began commercial driverless trucking between Dallas and Houston in May 2025 and reported roughly 250,000 cumulative driverless miles and ten route pairs by early 2026 (Aurora Innovation, 2025, 2026), while Kodiak delivers customer-owned driverless trucks in the Permian Basin (Transport Topics, 2025).

What abundant prediction does is fuse these elements into a single adaptive system. Inventory can leave a regional facility before its final destination is resolved, because the range of probable destinations is itself a prediction worth acting on. Goods already in transit can be rerouted as demand resolves. Cross-docks and micro-fulfillment nodes reduce dwell time where prediction is confident; large regional storage persists where it is not. Storage, transit, and fulfillment stop being separate line items and become one continuously coordinated position. Inventory location becomes conditional and adaptive — a service level expressed in space and time rather than a stack of goods behind a door.

The governing question of the architecture is not “how do we eliminate warehouses” but: when is the option value of moving inventory greater than the cost of keeping it in motion?

6.5 Warehouses change purpose rather than disappear

The evidence to date is emphatic that better forecasting does not abolish buffering — it transforms it. During precisely the decades in which demand forecasting improved most, U.S. warehousing employment grew from roughly 645,000 in 2010 to a peak near 1.9 million in 2022, standing at about 1.84 million in mid-2026 (U.S. Bureau of Labor Statistics, 2019, 2026), and some 1.8 billion square feet of industrial space was delivered after 2020 alone (Cushman & Wakefield, 2024). Meanwhile the aggregate inventories-to-sales ratio drifted down over the long run — from about 1.56 in 1992 to the high 1.2s in 2026, with a sharp rebuild after the 2021 shortages (U.S. Census Bureau, 2026). Firms hold relatively less routine inventory per dollar of sales while building far more capacity for fast-moving, finely positioned fulfillment. The building did not disappear; its function migrated from long-term storage toward high-velocity transfer, sorting, returns, and — increasingly — charging and energy infrastructure for electric and autonomous fleets.

6.6 The just-in-time paradox and selective buffering

Abundant prediction now creates a genuine strategic tension. Better prediction reduces the inventory needed against ordinary, well-modeled variation. But highly optimized systems can become fragile against rare disruptions, and shared prediction can make fragility systemic. The bullwhip effect showed decades ago how rational local responses to demand signals amplify distortion up a supply chain (Lee et al., 1997). Model abundance adds a new correlation channel: when many firms use similar models, similar data, and similar objectives, they can order, reroute, and de-stock in unison — the same herding and concentration dynamic financial regulators now flag for AI in markets (Financial Stability Board, 2024).

The 2021 semiconductor shortage is the canonical recent stress test. Global automakers cut roughly 10.5 million vehicles from 2021 production plans (AutoForecast Solutions, as reported in Automotive News, 2022), at an estimated revenue cost of $210 billion (AlixPartners, as reported in LeBeau, 2021). The instructive exception was Toyota — the inventor of just-in-time — which after the 2011 Tōhoku earthquake required suppliers to stockpile two to six months of semiconductors under its business-continuity plan, and rode out the early shortage while leaner competitors idled lines (Shirouzu, 2021). Toyota’s lesson is not that just-in-time failed; it is that just-in-time was never supposed to mean zero buffers everywhere. It meant buffers placed deliberately.

That is the winning architecture the framework predicts: less routine inventory against predictable variation, more deliberate reserves against tail risk — selective buffering, not perfect just-in-time. NIST’s supply-chain analyses point the same direction, warning against treating lean efficiency as sufficient when networks are fragile (Kaschak, 2022).

The marginal prediction can spread instantly. The complementary system cannot.

6.7 The test applied

Run the six conditions. Predictively positioned inventory (1) was impossible while demand and arrival predictions were expensive, slow, and coarse — no one ships toward an unresolved destination on a quarterly forecast; (2) embeds prediction continuously in every positioning decision rather than selling forecasts to planners; (3) sells a new outcome — availability within a time window — rather than cheaper storage; (4) changes the timing and institutional boundary of fulfillment, merging storage, transit, and retail at the margin; (5) makes physical capacity, energy, working capital, interoperability, cybersecurity, and resilience the central scarcities; and (6) would remain valuable with the word “AI” deleted — customers would simply say the thing they need is somehow always nearby.

The new category, stated compactly: inventory whose position across production, storage, and transit is continuously adapted to predicted demand and risk.

Signals that would show the category forming: inventory days falling for well-predicted product classes while strategic reserves persist or grow for high-consequence goods; rising share of shipments rerouted while in transit; warehouse dwell time falling relative to cross-dock throughput; empty and loaded vehicle-miles tracked per fulfilled unit; human interventions per thousand completed shipments falling; and performance during rare disruptions, not merely average conditions. Signals that would weaken it: rolling stock used as storage at scale (motion without option value), or systemic failures traced to correlated model behavior outpacing the resilience investments meant to contain them.

The framework treats labor displacement as a consequence, not the thesis — but the consequence deserves careful statement, because it is where the public debate lives.

The correct unit of analysis is the task, not the occupation. A professional driver is a bundle: perception and prediction, but also inspection, loading, communication, passenger care, security, local knowledge, and legal responsibility. Automation does not encounter “a job”; it encounters this bundle, and it separates the parts it can perform from the parts it cannot. The first effects are therefore substitutions of particular tasks within occupations — specific routes, specific operating conditions — rather than the simultaneous disappearance of job titles. Federal projections reflect this: U.S. employment of delivery drivers is projected to grow 8 percent from 2024 to 2034, and heavy-truck driving 4 percent, in part because the volume of delivery itself keeps rising (U.S. Bureau of Labor Statistics, 2025a, 2025b).

Some driving work migrates from continuous onboard control to remote supervision and exception handling. This is now a regulatory category, not a speculation: California’s autonomous-vehicle rules define training and licensing responsibilities for remote operators, extended in 2026 to distinct “remote driver” and “remote assistant” roles (California Department of Motor Vehicles, 2026), and the U.S. Department of Transportation identifies remote supervision, maintenance, data management, and cybersecurity among plausible new roles (U.S. Department of Transportation, 2021). How many vehicles one person can safely support is an empirical variable, not a slogan; no staffing ratio should be assumed. The economically important quantity is human exception-attention required per autonomous vehicle-hour produced. If exceptions are genuinely rare, labor requirements fall sharply; if supervisors must maintain continuous vigilance, the system has relocated the driver rather than absorbed the driving.

Two cautions temper optimism. First, capabilities that become more valuable to a service do not automatically command higher wages; a skill’s value to the product and a worker’s bargaining power are different things, and federal workforce analyses explicitly flag job quality and worker matching as open questions (U.S. Department of Transportation, 2021; U.S. Government Accountability Office, 2019). Second, transitions can be geographically and temporally concentrated even when aggregates look benign. Prediction software replicates at software speed; vehicles, infrastructure, regulation, insurance, and training do not. Fleet turnover alone stretches deployment over a decade or more (U.S. Department of Transportation, 2021). The World Economic Forum’s projection of 170 million jobs created and 92 million displaced by 2030 — a net gain of 78 million — is not an AI-only forecast and should not be quoted as one, but its shape matches the historical pattern: churn is the story, not simple subtraction (World Economic Forum, 2025).

Displacement measures the effect of AI on the old economy. It does not measure the value created by the new one. Both statements can be true, and policy should take both seriously.

The legal transformation is often summarized as “negligence dies, product liability replaces it.” The evidence supports something more interesting: responsibility expands along the chain of control.

In a conventional crash, the central question is usually the conduct of one human driver at one moment. In an automated system, the candidate defendants multiply: a remote operator who retained a duty to intervene; a fleet operator responsible for maintenance and deployment decisions; a manufacturer whose system may be defective; a software or model provider whose update introduced a common failure; a mapping or infrastructure provider; a regulator or standards body that defined acceptable operating conditions. These regimes coexist rather than replace one another, and the balance among them depends on jurisdiction and on who controlled the driving task at the relevant moment. The working hypothesis:

Accountability increasingly follows control.

Two structural changes are already observable. The first is that risk becomes correlated and version-level. Human driving errors are largely independent; a defect shared by a software version is fleet-wide by construction. When a Cruise vehicle dragged a pedestrian in October 2023, the consequence was not one lawsuit but a recall of all 950 affected vehicles, suspension of the company’s California permits, and ultimately a reported multimillion-dollar settlement paid by the operator (Cruise LLC, 2023; California Department of Motor Vehicles, 2023). Waymo has filed successive recalls — 444 vehicles, then 672, then 1,212 — each a software defect remedied fleet-wide by update (Waymo LLC, 2024, 2025). The recall, a mechanism built for defective parts, is becoming a mechanism for defective behavior.

The second is that the same centralization makes risk more observable. Automated vehicles record detailed telemetry, and U.S. regulators require designated manufacturers and operators to report qualifying crashes involving automated systems, allowing pattern detection across fleets (National Highway Traffic Safety Administration, 2025). Jurisdictions are moving at different speeds and in different styles: the European Union’s 2024 Product Liability Directive expressly treats software as a product and its developers as manufacturers within a strict-liability regime, with member-state transposition due by December 2026 (Directive (EU) 2024/2853); the United States, by contrast, still has no comprehensive federal autonomous-vehicle statute — a proposed federal evaluation program was withdrawn in June 2026 — leaving a patchwork of federal safety standards and state regimes, including California’s 2026 rules that authorize heavy-duty autonomous trucks and, for the first time, direct traffic citations to autonomous fleets (California Department of Motor Vehicles, 2026; National Highway Traffic Safety Administration, 2026). Meanwhile American juries have begun assigning manufacturers a share of fault in driver-assistance crashes: in 2025 a federal jury found Tesla one-third responsible in a fatal Autopilot crash, a verdict sustained against post-trial challenge in early 2026 and now on appeal (NPR, 2025; CNBC, 2026). Enforcement, in short, moves upstream — from citing a driver after a violation toward certifying, auditing, restricting, and penalizing the organizations and systems that determine vehicle behavior.

The trolley problem, vivid as it is, should not carry this discussion. The consequential choices are ordinary and prior: operating-domain limits, speed policy, following distance, uncertainty thresholds, fallback behavior, testing regimes, maintenance intervals, disclosure of known limitations, telemetry retention, and evidence preservation. Courts and regulators do not need a machine to name its sacrifice victim in order to ask whether a system was reasonably designed, validated, monitored, and restricted in light of foreseeable risk. The legal center of gravity moves from judging a split-second human reflex toward judging the prior architecture of control.

Prediction abundance makes responsibility architectural.

The thesis should be held to its strongest objections. Seven deserve standing acknowledgment.

Prediction may stay expensive where it matters most. Cost collapses within well-modeled operating domains, but rare, safety-critical edge cases resist statistical treatment precisely because they are rare. The operating-domain boundary is where prediction remains scarce — visible in practice when a truck maker asked Aurora to return an observer to the cab of its driverless trucks weeks after commercial launch (Bloomberg, 2025). The framework survives because categories can form inside bounded domains; but their growth rate is set by how fast those boundaries widen, and that speed is not guaranteed.

Cheaper travel may mean more travel. If autonomous mobility lowers the perceived cost of time in motion, induced demand could raise congestion and extend sprawl, as the chauffeur experiment and fleet simulations both suggest (Harb et al., 2018; International Transport Forum, 2015). The category can still form — programmable space is valuable in a sprawling region too — but its social value depends on pricing and policy choices outside the technology.

Empty repositioning may offset parking gains. The urban ledger is land released from storage minus land and road capacity consumed by curbs, charging, staging, and empty movement. Nothing in the framework guarantees the ledger is positive in every city; it predicts only that the option becomes available.

Physical capacity can remain the bottleneck. Ports, transformers, chips, road space, production lines, and working capital do not respond to software-speed information. Where capacity binds, better prediction reallocates shortage rather than eliminating it.

Optimization can purchase efficiency with fragility. Systems tuned tightly to prediction perform beautifully inside the distribution and can fail severely outside it — the just-in-time paradox generalized.

Correlated models can fail together. Shared models, data, and objectives create a common-mode failure channel that did not exist when predictions were made by millions of idiosyncratic humans (Financial Stability Board, 2024; Lee et al., 1997). Version-level defects in vehicle fleets are the physical instance of the same structure (Waymo LLC, 2025). New categories arrive carrying new systemic risks, and the resilience investments they require are part of their true cost.

Regulation, trust, and liability can slow category formation — sometimes fatally for a firm, as Cruise’s suspension showed (California Department of Motor Vehicles, 2023). This is not an anomaly of the transition; under the complement logic it is the transition. Authority and trust are among the scarce complements whose price rises when prediction gets cheap.

Finally, labor displacement and its distribution remain morally and politically important even though they are not, on this argument, the source of AI’s long-run value. A framework that explains why displacement is the smaller economic story does not excuse anyone from managing the smaller story well.

The argument survives these objections in a specific sense: each one shapes where, when, and whether particular categories form, and who benefits — none restores the premise that AI’s significance can be measured by counting replaced workers.

The two case studies converge on a recognizable pattern. A category is forming when the noun starts to fail: when “car” undersells a room that moves, and “warehouse” undersells a network that keeps goods perpetually well-positioned. It is forming when the unit of purchase changes — from vehicle to guaranteed movement, from storage to availability. It is forming when the geometry changes: cabins without control positions, curbs engineered for turnover, cross-docks displacing storage aisles. It is forming when liability migrates toward the entities that exercise control, because responsibility is a reliable tracer of where decision-making actually resides. And it is forming when new scarcities become the strategic terrain — curb space, charging, exception attention, resilience reserves — while the prediction itself drops out of the conversation.

The last signal is the decisive one, and the photographic analogy explains why. Nobody markets a smartphone camera as “arithmetic-powered.” The computation won so completely that it vanished into the category it created. The most reliable sign that prediction has transformed a system will be that its users stop mentioning prediction at all — attention will be paid, human and institutional, only to the unresolved remainder: the exceptions, the objectives, the accountabilities. A system that instead multiplies dashboards, alerts, approvals, and supervisory queues has made prediction abundant without making it invisible; it has automated the work while leaving humans to carry it. Where human attention concentrates is therefore a diagnostic: attention flowing to genuinely consequential exceptions signals a category maturing; attention consumed by routine confirmation signals a substitution stalled midway.

These claims are falsifiable, and Sections 5 and 6 specified the instruments: parking conversion and curb metrics, empty miles per passenger-mile, intervention rates per vehicle-hour, inventory days by risk class, in-transit rerouting shares, disruption performance. If a decade from now the vehicles still look like cars, the warehouses still work like warehouses, and the ratios have not moved, the thesis of this paper — not merely its timing — will deserve revision.

The first wave of every general-purpose input is imitation. Cheap arithmetic first did what human calculators did; cheap prediction first does what human predictors do. Judged during the imitation phase, both look like labor-replacement technologies, and the accounting fills with subtraction.

But arithmetic did not matter because it emptied the computing rooms. It mattered because, once abundant, it changed what an image, a ledger, a song, and a message could be. The categories it created — digital photography among them — are invisible monuments to an input that became too cheap to notice. Prediction is now crossing the same threshold, and the early evidence runs through this paper: vehicles that regulators no longer assume have drivers, inventory that ships toward addresses not yet written, responsibility migrating from reflexes to architectures.

Prediction is not action, and abundance is not magic. Every category examined here still depends on scarce complements — energy, curb space, capacity, trust, authority, resilience — and on institutions willing to redesign themselves. The categories will form unevenly, shaped by policy as much as by models, and they will carry new failure modes commensurate with their new powers. That is what it has always meant for an input to become infrastructure.

The measure to watch, then, is not how many of the old tasks the machines take over. It is how many things start existing that could not have existed before. The power of AI will be visible not when machines imitate the categories humans already built, but when abundant prediction creates categories that could not exist before.