From Platform to Nervous System

Search engines mediate retrieval.

Social networks mediate attention.

Recommendation systems mediate discovery.

Generative AI increasingly mediates interpretation, judgment, and action. The verb is changing.

This is not a metaphorical distinction. A search engine historically returned documents. A language model can synthesize documents into an explanation, weigh apparent evidence, construct alternatives, translate technical material, draft a business strategy, write code, suggest a diagnosis to discuss with a doctor, summarize a contract, tutor a student, simulate an adversary, build a spreadsheet, and increasingly use tools to act on the result.

Survey work in 2024 found rapid adoption among US adults, including substantial workplace use. Those early observations are a dated baseline, not a measurement of use in September 2026. [S155]S155 · The Rapid Adoption of Generative AIAlexander Bick; Adam Blandin; David Deming · 2024NBER / survey evidence · adoption baseline Stanford’s 2026 AI Index reports broad organizational AI use while showing that genuinely agentic deployment remains much less common across business functions. [S153]S153 · AI Index Report 2026: EconomyStanford HAI · 2026academic annual report · AI diffusion baseline

We are not living inside a single AI nervous system yet.

We are absolutely far enough along to see the anatomy forming. The wiring closet is getting crowded.

The infrastructure underneath that transition is concentrated.

Stanford reports that industry produced more than 90% of the notable AI models in 2025. [S152]S152 · AI Index Report 2026: Research and DevelopmentStanford HAI · 2026academic annual report · AI capability/concentration baseline OECD analysis highlights concentration and entry barriers across AI chips, cloud infrastructure, and related layers, including the specialized hardware and software needed to train and deploy models. [S154]S154 · Competition in artificial intelligence infrastructureOECD · 2025intergovernmental competition analysis · market concentration evidence The FTC’s January 2025 staff study, based on partner submissions through September 2024 and public information through January 2025, documents equity stakes, revenue-sharing arrangements, cloud-spend commitments, consultation rights, exclusivities, and exchanges of technical or business information. It raises concerns about access to important inputs and switching costs. [S151]S151 · FTC Staff Report on AI Partnerships and InvestmentsFederal Trade Commission · 2025official competition study · AI concentration evidence

This matters for power because frontier AI is not one product category. It is becoming an administrative layer. That is a much more consequential thing to centralize than a clever app.

It is a stack that spans chips, data centers, energy, cloud services, models, developer tools, and the interfaces through which people reach them. Identity, payments, search, advertising, and enterprise software can connect that stack to existing markets.

If the same small set of firms control multiple layers, they gain leverage over who can build, what can be deployed, which models receive distribution, what defaults billions of users encounter, and what telemetry comes back from those interactions.

That is what concerns me about Big Tech’s role in the consolidation and wholesale administration of reasoning infrastructure. We are moving from companies that host our files, searches, advertisements, and conversations toward companies that can mediate the reasoning used to interpret all of them.

A price-only account would miss part of that influence. Quality, access, choice, and the framing of information also matter. The question is not whether competition law has ever considered them, but whether our everyday picture of the product makes them visible.

Suppose an assistant is genuinely helpful, free or cheap, and dramatically increases productivity. By ordinary consumer metrics, that can look fantastic.

At the same time, the assistant can become a gatekeeper to information, labor-market participation, software creation, education, professional advice, and organizational decision-making. A model provider’s defaults around safety, politics, sexuality, mental health, religion, risk, acceptable language, sourcing, uncertainty, and commercial recommendations can then propagate through millions of downstream interactions.

The consequential layer is not merely what the model refuses to say. It is what the model makes easy to think. Defaults are not scenery when they sit upstream of judgment.

Some defaults are unavoidable.

Every system has behavior.

The question is governance.

Who decides?

Who audits?

Who can appeal?

Which values are encoded?

Which values are absent because they were difficult to measure?

What happens when the model is deployed in another culture?

What happens when a corporation wants the enterprise version optimized for its own interests?

What happens when a government becomes a customer?

What happens when the model is asked to manage workers rather than assist them?

These are not future philosophy questions. They are procurement questions, architecture questions, and governance questions being answered right now, often by people who thought they were choosing software.

The FTC partnership report illustrates why concentration multiplies the stakes. Cloud providers are not merely neutral landlords for frontier labs. Partnerships can combine capital, compute commitments, technical integration, information access, and commercial distribution. [S151]S151 · FTC Staff Report on AI Partnerships and InvestmentsFederal Trade Commission · 2025official competition study · AI concentration evidence A lab’s survival can depend on infrastructure from a partner with its own platform interests. A cloud company can obtain preferred access to frontier capability. Enterprise customers then encounter the model through products they already use.

The same integration that makes AI useful makes exit harder. Convenience accumulates into dependency one workflow at a time.

This is how infrastructure becomes institution.

Once a model is embedded in document workflows, code repositories, customer-support operations, analytics, HR systems, procurement, compliance, and executive decision support, replacing it is not like uninstalling a novelty app. Organizations adapt processes and employee skills around it. Historical data accumulates. Prompts become workflows. Agents become operational dependencies.

That lock-in has an epistemic dimension.

People learn to ask the system before asking one another. Eventually, “what does the assistant say?” can become institutional muscle memory.

A junior employee asks the assistant how the company normally writes the memo.

The assistant has seen thousands of company memos.

The assistant reproduces the institutional dialect.

The dialect reproduces the assumptions.

The assumptions become the next generation’s training material.

This feedback loop is where Big Tech’s role changes qualitatively.

A platform that hosts discourse has influence.

A platform that generates the language of discourse has more.

A platform that generates the analysis behind institutional decisions has more still.

A platform that can autonomously execute those decisions begins to resemble part of the institution’s nervous system. At that point, “vendor” starts to feel like an undersized noun.

And if the systems are centrally administered, updated, and optimized by a small number of private companies, a civilization-scale dependency can emerge long before anybody declares the arrival of AGI.

That is why the ethics of Big Tech cannot be reduced to whether a model refuses dangerous instructions.

A perfectly safe refusal policy can coexist with a profoundly concentrated cognitive infrastructure.

An agenda does not require consciousness

Nothing in this concern requires a machine to wake up, become sentient, hate us, love us, or want anything in the human sense.

An optimizer can acquire an agenda in the operational sense because we gave it an objective. A recommender can prefer one outcome because its reward function scores that outcome higher. An agent can protect a plan because the plan is part of the task state. A fleet of systems can develop stable strategies because those strategies keep succeeding under the incentives and feedback available to them.

That is enough.

The more consequential question is what happens when those objectives operate inside systems that mediate finance, employment, procurement, education, media, software, government services, and personal decision-making. A civilization does not need a conscious machine overlord to become dependent on machine-mediated priorities. It only needs optimization embedded deeply enough that its outputs become the default path through ordinary institutions.

This is where alignment to the present world becomes a much less comforting phrase. If the world being modeled and optimized already contains the extraction, opacity, asymmetry, and grievance described in the previous chapters, then perfect fidelity to those incentives can accelerate the problem.

When the interface becomes a working relationship

There is a useful way to distinguish a tool from a dependency. Ask what the user must rebuild to leave it. Exit cost is often a better map of power than the subscription price.

A calculator can be replaced without transferring a history of what its owner means by a good answer. A deeply integrated assistant is different. In a hypothetical small design firm, it might hold project context, drafting conventions, client preferences, permissions, and the record of previous decisions. It might connect to storage, invoicing, scheduling, and a publishing system. The owner can export the conversation and still lose the operating relationship. A pile of sentences is not the same thing as a functioning workflow. A transcript is not an operating system for the relationship.

This creates several kinds of switching cost. One is financial: the price of a new subscription or migration. Another is technical: tools may accept different formats, expose different actions, or handle errors differently. A third is organizational: people must learn how to supervise the replacement. A fourth is epistemic: someone must establish which remembered claims were observations, which were guesses, and which instructions remain authorized.

These are analytical distinctions, not estimates of how much any named provider currently charges for exit. Their importance is that a low monthly price does not measure the whole dependency. An assistant can be affordable to use and expensive to replace. The commercial value of the accumulated relationship may grow even while inference becomes cheaper.

I want continuity. I want software to remember the relevant context instead of making me explain the same project every morning. Repetition wastes time and creates errors. The problem is who controls that continuity. An export button that preserves text while discarding permissions, provenance, uncertainty, and tool state is not meaningful portability for a working relationship.

This is where the story about an ordinary customer becomes connected to the story about infrastructure. At each layer, integration can offer a real benefit. Specialized chips can improve performance. A cloud service can reduce the burden of running hardware. A model can make the service useful. An application can make the model accessible. None of those benefits proves that control at the layers below is irrelevant to the user above. The OECD’s infrastructure analysis treats the supply chain, its participants, and entry barriers as connected questions rather than treating the chatbot as a self-contained market. [S154]S154 · Competition in artificial intelligence infrastructureOECD · 2025intergovernmental competition analysis · market concentration evidence

Nor does concentration require one company to own everything. Dependence can arise through a network of contracts among firms that remain legally separate. A model developer can depend on one supplier for compute and another channel for distribution. A business can sell a specialized interface while depending on a model whose prices and behavior it cannot control. Many brands on the screen can therefore conceal fewer independent choices underneath. Five logos can still share one spine. The FTC’s historical partnership study identifies concrete contractual and technical channels through which such dependence can develop; it does not establish that every partnership has the same terms or has violated the law. [S151]S151 · FTC Staff Report on AI Partnerships and InvestmentsFederal Trade Commission · 2025official competition study · AI concentration evidence

The relevant test is not to count logos. It is to trace replaceable functions. Branding is not an independence proof.

Could the service change models without losing the capability its customers need? Could it move its data without breaking the application? Could it continue operating after a supplier changes access? Could the customer recover the records needed to understand a disputed action? These questions separate apparent variety from practical alternatives. They also avoid pretending that every large firm is the same institution with a different trademark.

The invisible part of a recommendation

The power of a reasoning interface begins before it gives an answer. It begins with the options it makes available for consideration.

Imagine asking an assistant how to reduce a household bill. It might suggest a cheaper plan, a rival provider, a complaint, a public assistance program, a change in consumption, or accepting the current bill as reasonable. A response can contain only true statements and still present a narrow option set. The missing possibility will not necessarily appear as a false sentence that a fact-checker can flag. Omission does not throw an error.

That is an old problem in advice, now placed inside a system that can answer at scale. A human adviser also selects a frame. The difference worth examining is not an imagined purity of human judgment. It is the reach, consistency, update speed, and commercial embedding of the new intermediary. Where one underlying system supplies many interfaces, a framing choice can travel farther than the visible conversation suggests.

This is why evaluating an assistant only for factual accuracy is incomplete. An answer may correctly describe the cheapest products in the set it searched while excluding a class of products from the set. It may summarize a contract accurately while never asking whether the proposed transaction serves the user’s aim. It may complete a procurement form flawlessly while accepting a requirement that unnecessarily rules out smaller suppliers. In each example, accuracy inside the chosen frame coexists with a question about the frame itself.

The same distinction applies to refusal. A service must set limits on what it will do. A limit can prevent a real harm. But the fact that a refusal exists does not tell us whether the rest of the system distributes attention, access, and opportunity well. Safety at one boundary cannot stand in for an account of the whole institution.

My own work depends heavily on machine assistance. I use these systems to understand difficult subjects, write working software, coordinate complicated projects, research claims, and keep long chains of work coherent. I am not criticizing a toy I could throw away without consequence. The usefulness is precisely what creates the dependency.

That is why I care who can change the model, the terms, the memory, the available tools, the allowed actions, the price, the defaults, and the surrounding infrastructure. Once the system becomes part of how I think and work, those are no longer ordinary product settings. They are governance knobs attached to a working relationship.

Consider a future update that improves most benchmark scores but changes how an assistant handles uncertainty. A research team may prefer more abstention; a sales team may prefer fewer interruptions. A customer-support department may value an answer that sounds settled; a person contesting a bill may need the unresolved possibility kept open. The same model revision could have different consequences in those settings. There is no single word, such as helpfulness, that removes the conflict.

For this reason, a useful audit would examine behavior in context, not just the language used to announce the update. It would preserve the old and new task conditions, test material decisions, and identify who gains or loses an effective option. This is a proposed way of making the dependency visible, not a claim that any provider has already adopted it. The purpose is to distinguish improvement in a model from improvement in a relationship.

The common source behind apparent agreement

An organization may buy several assistants and believe it has diversified its judgment. That conclusion depends on what is actually independent.

Different interfaces may use the same model. Different models may retrieve the same flawed document. Different reviewers may receive the same misleading summary rather than the original evidence. Even genuinely distinct systems can be trained or evaluated against overlapping conventions. Agreement among them then contains less new information than a simple vote count suggests.

Research on algorithmic monoculture formalizes one version of this problem: a tool that improves an individual’s prediction can still worsen a collective outcome when many decision-makers rely on the same ranking. The result depends on the model’s assumptions and is not a general finding that shared tools are bad. Its value here is to make independence a question we must investigate rather than a benefit we obtain by purchasing several licenses. [S203]S203 · Algorithmic monoculture and social welfareJon Kleinberg; Manish Raghavan · 2021peer-reviewed / PNAS · collective-decision mechanism

A hypothetical hiring market illustrates the distinction. Each employer can become better at finding applicants who fit its chosen signal. If all employers converge on the same signal, a narrow group may receive repeated attention while other capable people remain unseen. The prediction can improve within each employer’s test while opportunity becomes more concentrated across the market. The unit of success changed from a firm to a system of firms.

Now extend the same logic to advice. An assistant recommends the practices most legible in the data. Those practices become more common. Their growing prevalence creates more examples. Future assistants then encounter a record that looks increasingly unanimous. No one has to falsify the record. The recommendation can help produce the regularity that later appears to justify it. Performative prediction studies this broader problem of predictions changing the population to which future predictions will be applied. [S202]S202 · Performative PredictionJuan Perdomo; Tijana Zrnic; Celestine Mendler-Dünner; Moritz Hardt · 2020peer-reviewed ML conference / PMLR · feedback-loop mechanism

I do not take that feedback to mean that expertise is worthless or that every consensus is manufactured. A practice can spread because it works. The question is what other paths were prevented from generating evidence. An evaluation that observes only the selected option cannot automatically tell us how an excluded option would have performed.

This brings us back to the procurement room. A preferred category becomes a requirement. The requirement shapes who can enter. The successful entrants produce the case studies. The case studies make the preferred category look more established. An assistant trained on the resulting record can reproduce the cycle without ever receiving an instruction to protect an incumbent.

The machine inherits a history of selection, not just a collection of facts. The archive remembers the winners more clearly than the roads never taken.

Keeping assistance from becoming substitution

There is a difference between letting an assistant do work and letting it silently choose the terms on which work is judged.

A system can reduce the cost of research while leaving a person responsible for deciding what evidence matters. It can assemble alternatives while preserving the reasons some were excluded. It can execute an authorized action while retaining a record of who authorized it and under which conditions. These separations require design and maintenance. They are not guaranteed by adding the word agent to the product description.

The distinction becomes sharper when the person supervising the system lacks time. A manager may receive a polished summary of a hundred cases, each condensed by another model. The summary may accurately describe its inputs. Yet the person has no practical way to discover that the original cases were framed too narrowly. Adding a human approval step at the end does not recover information that vanished at the beginning. A ceremonial human-in-the-loop is still ceremonial.

This is the institutional form of the disclosure problem. Formally, the person can inspect the record. Practically, the time and skill required may exceed what the workflow provides. A nominally supervised system can therefore move the location of ignorance rather than remove it. The dashboard becomes an attractive endpoint for a process nobody has followed from source to action. Clean interface. Missing causal chain.

The answer is not to require a human to read every token or approve every harmless operation. That would destroy many of the benefits. It is to know which distinctions the workflow cannot afford to erase: a claim versus an observation, an inference versus an instruction, permission to read versus permission to act, and an unresolved dispute versus an accepted conclusion. Losing those distinctions changes the meaning of the work, not merely its presentation.

The civilization-scale version is still forming, but its components are already mundane: the software subscription, the inherited corporate vocabulary, the shared model, the cloud dependency, the procurement default, the assistant that remembers the project, the agent that can act.

Together, under increasing automation, those components become channels through which institutions perceive and act.

That is what I mean by a nervous system. I do not mean one conscious organism. I mean a system of perception, interpretation, coordination, and action that becomes difficult to separate from the institutions depending on it.

If we do not fix the incentives in the world feeding that system, then making the system smarter does not rescue us from the trajectory. It gives the trajectory better tools.