Drag Is Not Rescue: friction changes the timing, not the destination

The strongest objection argues about time, not mechanism

The six essays before this one set out a structural argument. Unit cost dominance has been crossed for a large share of professional cognitive tasks. Interface collapse carries the advantage into workflows. The Multiplayer Prisoner’s Dilemma means no actor can hold the propagation back. The Sorites Collapse Principle and Categorical Recursion close the regulatory route. The Successor System closes the route of structural alternatives. One route remains open in most serious objections, and this essay closes it.

The most sophisticated reply to the thesis does not contest the mechanisms. It contests the timeline. It accepts that the pressure exists and points to the many forms of friction that slow, channel or localise deployment. Integration is hard. Liability is unsettled. Professions defend their ground. Privacy law restricts data. Power and chips are physically scarce. Cultures resist. Local labour markets adapt slowly. Premium niches absorb some people. New kinds of job appear. Politics may deliver redistribution. Each of these is real, and each slows the propagation somewhere.

The conclusion the objection wants is that friction buys time, time allows adaptation, and adaptation preserves enough of the wage-demand circuit to avoid the discontinuity. This essay disagrees. Friction modulates timing. It does not restore the circuit. The question is not whether friction exists. It is whether friction restores mass productive necessity, and it does not.

The friction inventory, taken at its strongest

Start by stating every form of friction as strongly as it can be stated.

Integration is expensive. Deploying AI inside an enterprise workflow is not a matter of buying API access. It takes data pipelines, security review, change management, training, validation of outputs, integration with existing systems, error handling, audit infrastructure and a degree of reorganisation that often outlasts the purchase. Moving from pilot to production at scale takes months and sometimes years.

Liability is unsettled. Who is responsible when an AI-drafted contract contains a material error, when an AI recommendation harms a patient, when an AI-run workflow breaches a regulation? Until those questions are answered, cautious firms keep human review whose main purpose is to allocate liability rather than to improve quality. It is expensive but defensible to insurers, regulators and courts.

Professions gatekeep. Bar associations, medical boards and accounting bodies control credentials and define professional conduct. They have institutional reasons to preserve human roles, and they can require human judgement at specific points, forbid some kinds of delegation, or impose disclosure duties. Inside a regulated profession those rules have force.

Privacy law constrains data. The GDPR, HIPAA and similar regimes restrict what data can cross institutional and national boundaries. Many deployments need data localisation, anonymisation or restricted-processing architectures that add cost and limit capability.

Compute and energy are finite. Training is limited by chips, data-centre capacity, power and grid connections, and large-scale inference faces milder versions of the same limits. Compute is growing fast, energy more slowly. Some deployments wait for capacity that does not yet exist.

Culture resists. Workers, managers, customers and citizens have preferences about dealing with people. Some markets pay for human service. Some institutions resist for reasons of identity, pride or habit.

Local labour markets differ. Specialisation, networks, language, regulation and culture mean that different places adjust on different timescales. A technology hub and a regional manufacturing town do not move together.

Premium human niches exist. High-end legal practice, bespoke medical care, luxury hospitality, craft production and concierge services pay a premium for people. They are real labour markets with real wages.

New jobs form. Every earlier wave of automation created categories that had not existed, and this one is creating some: safety researchers, model operations specialists, agent designers, deployment auditors, verifiers.

Redistribution is available. Welfare states, tax and transfer systems, basic-income pilots, sovereign funds and public compute are all real routes by which the worst outcomes of an unmanaged transition could be avoided.

None of this is illusory. The claim is that none of it, alone or in combination, restores general-purpose cognitive labour as the mass scarcity that supported middle-class absorption.

Each friction decays, routes around, or stays local

Take them in turn.

Integration cost is an adoption tax, paid once and then amortised. The first deployment is expensive, the tenth routine, the hundredth automated. It is high now because firms are doing it for the first time. It falls as they learn, as vendors standardise the tooling and as consultancies package deployment as a service.

Liability settles. The open questions will be closed by case law, regulation, insurance products, indemnities and professional practice. The settlement will not favour deployment everywhere, but where AI judgement proves more reliable than human judgement, courts and insurers will treat its use as the standard of care. At that point the incentive reverses, and not deploying becomes the legal risk.

Professional gatekeeping runs on the same logic as the Multiplayer Prisoner’s Dilemma. Professional bodies are made of practitioners under competitive pressure. A body that restricts AI slows adoption inside its jurisdiction while its members lose work to adjacent providers, foreign competitors and unregulated alternatives. The drift towards accommodation is structural.

Privacy law changes the architecture of deployment rather than preventing it. Federated learning, on-device inference, anonymisation pipelines and data clean rooms let regulated industries deploy with adjusted infrastructure. The constraint is real and routable.

Compute and energy limits slow particular deployments at particular moments without bending the trajectory. The cost of inference has fallen by orders of magnitude in a few years, and energy infrastructure is being built at a pace not seen in decades because the demand justifies it.

Cultural resistance is real but uneven. Some markets prefer human service and pay for it. Most markets, over time, choose price. The human premium survives at the edge of every market and at the centre of a few. The centre of most markets has always been cost-sensitive.

Local labour markets move at different speeds in the same direction. The relevant claim is not uniform collapse. It is that the wage premium for cognitive work weakens wherever interface-mediated cognitive work becomes substitutable. The OECD’s Employment Outlook 2026 adds a point the friction argument needs and does not have. When technology shocks hit local labour markets, adjustment often happens through transitions into joblessness and fewer openings for newcomers, rather than through displaced workers moving into other sectors, and the scars last. Slow local adjustment is not the same as absorption. It is often its absence.

Premium niches absorb a small fraction of the displaced. High-end legal practice in the United States employs perhaps a hundred thousand lawyers, against more than a million in the profession as a whole. Premium categories scale to thousands or tens of thousands each, not to the tens of millions that mass absorption requires. They are residual scarcity, not mass scarcity.

New job formation produces real roles, some well paid. They do not replace the cognitive work being drained from the entry-level pipeline, and many of the new categories are themselves exposed to the next generation of models. The numbers do not close.

Verification is bounded by the same arithmetic. The share of people who can meaningfully oversee complex multi-step agentic systems at a professional standard is a minority of the workforce. That is large enough to matter and far too small to restore mass absorption at the scale the postwar circuit needed. The displaced population is not the population that can be retrained into verification at a wage the deployment will pay.

Redistribution is the structural alternative the previous essay examined. It can preserve consumption. It does not preserve the circuit. What emerges from large-scale redistribution may be humane, fair and stable, or authoritarian, brittle and captured. Either way it is a successor, not a continuation.

The most protected sectors are publishing their own deployments

One version of the argument deserves direct treatment. It holds that some sectors are friction-locked, so they will keep a fragment of the wage-demand circuit alive while others collapse. The usual candidates are pharmaceutical documentation, medicine, law, defence, classified work and other heavily regulated cognitive domains. The claim is empirical. These sectors will not see deployment at scale, it says, because their friction is structural rather than temporary.

The deployment record contradicts it in exactly the sectors the argument names first.

The clearest case is Novo Nordisk’s NovoScribe system, described in Anthropic’s published case study and in AWS’s account of the same deployment. Regulatory writing in pharmaceuticals has every property the friction argument relies on: heavy regulation, severe liability, conservative culture, sensitive data, demanding audit and regulator review at each stage. The deployment happened. According to the case study, clinical study reports that took up to fifteen weeks across forty to fifty professionals can now be produced in minutes by a team of three, and the resources needed for device verification protocols fell by ninety-five percent. These are the vendor’s and customer’s figures.

The same case study records months of development spent ingesting unstructured legacy data. That is friction behaving exactly as this essay says it does. It set the integration timeline. It did not stop the deployment. Friction is the adoption tax, not the barrier.

Others in protected domains describe the same trajectory. Brandon White, chief executive of Axiom Bio, was quoted in OpenAI’s GPT-5.5 launch saying that “if OpenAI keeps cooking like this, the foundations of drug discovery will change by the end of the year.” Mainstay, quoted in OpenAI’s GPT-5.4 launch, reported a 95 percent first-attempt success rate navigating regulated property-tax and homeowners’ association portals, finishing sessions about three times faster on about 70 percent fewer tokens. Harvey reported 91 percent on its BigLaw Bench test of transactional legal analysis on the same model generation. All of this is vendor and customer reporting, and the evidence register treats it that way.

What matters is the pattern. Named deploying companies across several model providers describe the same things: delegation, long-horizon execution, work that continues without a person in the loop, smaller teams, and deployment in regulated, high-stakes sectors. The friction argument needs friction to be sufficient to preserve mass productive necessity. The sectors cited as most protected are the ones now publishing case studies of their own deployment.

Friction does not add up to rescue

A critic can concede every point above and argue that the frictions together still produce enough drag to save substantial labour markets. Integration plus liability plus gatekeeping plus privacy plus compute plus culture plus locality plus niches plus new jobs plus redistribution might, in combination, slow things enough for the circuit to survive in a modified form.

That fails for three reasons.

The frictions do not combine into a stable floor. Integration cost amortises. Liability settles. Gatekeeping drifts. Privacy constraints are routed around. Compute grows. Culture adapts. Locality moves in the same direction at different speeds. Some frictions will persist locally. Liability may harden in particular high-stakes domains, energy may bind in particular jurisdictions, and political resistance may intensify in particular sectors. The total drag may stay large in places. It does not become a mechanism for restoring mass productive necessity, because the frictions that persist do not reverse the trajectory.

The frictions do not touch the mechanism. The Multiplayer Prisoner’s Dilemma operates however slowly deployment proceeds. As long as deploying is competitively dominant, every actor that delays loses ground to actors that do not. Friction spreads deployment across time and across actors. It does not produce an equilibrium in which deployment stops.

And the frictions do not restore the circuit even if they slow the propagation indefinitely. Suppose friction were permanent and deployment ran at half the pace the thesis assumes. The circuit would stop reproducing over forty years instead of twenty. The direction is the same. The structural condition is the same. The question of what replaces the circuit still arrives, later. Slow destruction is still destruction.

The propagation is a damped curve, not a step

The honest model is a damped exponential rather than a step function. The structural pressure runs in one direction. The damping is real and varies by domain, geography and political regime. What happens in any place is the pressure integrated against the local damping. The change spreads, at different speeds.

The current measurements fit that shape. The St. Louis Fed’s Real-Time Population Survey records generative AI assisting 1.7 percent of total US work hours in mid-2026, up from 1.4 percent at the end of 2024. That is early on the curve, and self-reported, so it likely undercounts. Meanwhile the leading indicator is moving faster than the aggregate. Stanford’s Digital Economy Lab, in an August 2026 revision, puts employment of 22 to 25 year olds in AI-exposed occupations 19 percent below where it would be had it kept pace with less-exposed peers, and finds the gap widening. Low aggregate penetration alongside a widening gap at the entry point is what a damped propagation looks like from inside.

So the next ten to thirty years are best understood neither as a single event called the discontinuity nor as a smooth evolution that preserves the old structure. They are a transition with a fixed direction and uneven speed, visible locally as wage stagnation and narrowing entry ladders rather than mass dismissals, and raising the successor question wherever it arrives. Friction is the local damping. It changes the speed. The essays before this one describe the direction.

Friction buys time, and the time has a use

Friction is not the absence of structural pressure. It is the existing system’s response to it. Workers, firms, professions, regulators and states generate friction because they feel the pressure. Friction is evidence of the pressure, not evidence against it.

The policy consequence is that managing friction should not be confused with preserving the circuit. A regulator who writes liability rules that slow AI deployment in healthcare is doing useful work. That regulator is not preserving the wage-demand circuit. They are buying time in which other choices can be made about what comes next. The time is valuable, finite, and should be spent.

Treating friction as rescue uses the time to postpone exactly the decisions the time makes possible. That is the worst available use of it. The right use is to design the successor deliberately rather than deny that one is needed. This is the request the whole sequence builds towards. Stop debating whether the circuit can be saved. Use the time friction provides to decide what replaces it, and on whose terms. The essays in What comes after take up that question.

What the sequence establishes

The sequence makes one claim in four steps.

The first premise is that unit cost dominance is being crossed. AI plus verification already produces a substantial and growing share of professional cognitive work at lower cost and equal or better quality than people working alone, across the kind of work that once absorbed the middle class. That is the subject of the first essay.

The second is that the advantage propagates. Because models now operate the software that work runs through, task-level dominance becomes workflow recomposition, which suppresses hiring and breaks training ladders without whole occupations disappearing. That is the second essay.

The third is that the propagation cannot be restrained as a competitive equilibrium. Restraint is dominated for workers, firms, sectors and states, observation of defection comes too late, and the enforcers are the actors most exposed to the cost. That is the third essay.

The fourth is that no available preservation route restores the circuit. Category-based regulation fails because the line between assistance and replacement exists nowhere inside a workflow and the deployed object migrates across legal categories. Structural alternatives preserve consumption rather than productive necessity. Friction slows the change without reversing it. That is the work of essays four to seven together.

If those premises hold, the conclusion follows. The wage-demand circuit is no longer self-reproducing under the new technological condition. It can continue institutionally for some time after its reproduction mechanism fails, and that continuation is managed transition rather than survival. A successor system arrives. Which one, who designs it and on whose terms remain open, and they are the questions worth arguing about.

What would refute this

This essay is refuted by a specified form of friction that is permanent rather than decaying, that cannot be routed around, and that restores mass productive necessity rather than delaying its erosion. The test is in the protected sectors themselves. If the domains cited as friction-locked showed stable or rising entry-level hiring, stable junior-to-senior ratios and a stable wage share over several years, while capability and deployment kept advancing elsewhere, friction would be doing more than modulating timing.

The aggregate version would be refuted by the leading indicators turning. If the Stanford early-career gap closed, entry-level hiring in AI-exposed work recovered, and the share of consumption financed by wages rather than transfers or debt held steady as deployment deepened, then the drag would be rescue after all. A slow timeline on its own does not count.