A few months ago, inside a product review meeting at a large technology company, a deceptively simple question stalled the room. A senior engineer had just presented an update to a recommendation system—nothing unusual, a set of refinements designed to improve engagement. At the end of the presentation, someone asked: why does this particular piece of content appear first, and not another one?
What followed was not a disagreement, or even a technical explanation, but a hesitation that felt almost philosophical. The engineer began outlining the factors—user history, interaction patterns, model predictions, content characteristics—but the explanation kept expanding, branching into dependencies, interactions, probabilities. After a minute or two, it became clear that the answer was not incomplete. It was structurally impossible to compress into a single cause.
The system did not produce decisions in the way we expect decisions to be made.
This moment, which is quietly familiar inside the industry, rarely makes it into public discussions about technology. We tend to imagine digital power in terms inherited from earlier eras: someone sets the rules, someone enforces them, someone benefits. Even the most critical accounts—whether from regulators, economists, or journalists—often assume that behind every outcome there is an identifiable locus of control.
But the architecture that now governs visibility online does not behave that way.
What you see is not chosen in a single place. It is assembled.
Over the past decade, digital platforms have evolved into systems where outcomes emerge from the interaction of multiple layers: ranking algorithms trained on behavioral data, interface designs that privilege certain actions, advertising mechanisms that introduce economic incentives, and user activity that continuously feeds back into the system. Each component is intelligible in isolation. Together, they form something closer to an ecosystem than a mechanism.
This is not entirely without precedent. Economists have long described markets as distributed systems. Friedrich Hayek’s central insight was that no single actor possesses all the information necessary to coordinate economic activity, and that prices serve as signals that aggregate dispersed knowledge. But even in Hayek’s framework, the mechanism of coordination remained legible.
In digital systems, the coordinating mechanism is no longer price alone. It is visibility—and visibility is governed by a structure that is neither fully transparent nor fully centralized.
Herbert Simon, writing decades before the internet, warned that human decision-making operates under conditions of bounded rationality. We simplify, we approximate, we rely on partial models of reality. What is striking today is not just that individuals are bounded in this way, but that the systems we build reflect the same limitation. Engineers understand components. Product teams understand metrics. Marketers understand performance signals. But the system as a whole exceeds any single perspective.
Even inside the companies that build these systems, there is no complete, real-time understanding of how outcomes are produced.
This is where the current conversation about artificial intelligence begins to feel oddly misplaced. Much of the attention—both in industry and in public discourse—has shifted toward tools: how to prompt models, how to generate outputs, how to integrate AI into workflows. These are practical questions, but they operate at the surface level of interaction.
What remains largely unexamined is the architecture within which these tools operate.
When a new model is introduced into a recommendation system, or when predictive optimization is added to an advertising platform, it does not replace the existing structure. It becomes another layer within it. Each layer is optimized locally, often by different teams, according to different metrics. The result is a system that becomes more capable, but not necessarily more coherent.
A senior machine learning researcher once described large-scale platforms to me as “systems we steer rather than control.” The distinction is subtle, but important. Steering implies influence without full command, adjustment without complete predictability.
This has consequences that extend beyond engineering.
In marketing, for example, the shift is already visible. Traditional models assumed that influence operated through messaging: identify an audience, craft a proposition, deliver it through a channel. But in digital environments, the channel is no longer neutral. It actively shapes which messages are seen, in what sequence, and with what frequency.
Research from the Ehrenberg-Bass Institute has long emphasized the importance of “mental availability”—the likelihood that a brand comes to mind in a buying situation. In digital systems, that availability is increasingly mediated by algorithmic retrieval. It is not just about being remembered. It is about being surfaced.
Marketers, in practice, have adapted to this reality more quickly than most theoretical frameworks. They run continuous experiments, adjusting creative formats, timing, targeting strategies. They observe which signals the system appears to reward and align themselves accordingly. A campaign that performs well is not simply one that persuades, but one that fits the logic of distribution.
And yet, even here, understanding is partial.
Executives speak of “working with the algorithm,” but the phrase is more metaphor than method. What they are really doing is responding to outputs, inferring patterns, and iterating. Success is often recognizable only in retrospect.
This creates a peculiar dynamic. The system produces highly structured outcomes—certain products gain visibility, certain narratives spread, certain behaviors are reinforced—but the process by which these outcomes emerge remains only partially understood, even by those most deeply involved in shaping them.
Shoshana Zuboff has argued that digital platforms exercise a form of power that operates by shaping behavior rather than issuing commands. That insight remains important, but it assumes a level of intentional design that is becoming harder to sustain as systems grow more complex. Increasingly, influence arises not from a single strategy, but from the interaction of many small optimizations.
The political theorist Hannah Arendt once distinguished between power as something held by individuals and power as something that arises from structures. In the digital environment, power seems to have shifted decisively toward the latter. It is embedded in the configuration of the system—in how signals are weighted, how feedback loops are constructed, how incentives are aligned.
This is why it is so difficult to challenge.
Regulatory debates often focus on discrete issues: content moderation, competition, data privacy. These are important, but they tend to treat platforms as if they were traditional institutions, capable of making clear, centralized decisions. In reality, many of the most consequential outcomes are not the result of explicit choices, but of systemic interactions.
You cannot point to a single moment and say: this is where it happened.
What makes this especially complicated is that users themselves are part of the system. Every interaction—every click, scroll, pause—feeds back into the models that shape future visibility. Behavior becomes input, input becomes output, and the cycle continues.
In that sense, the system is not simply acting on users. It is co-evolving with them.
Which brings us back to the original question.
Who decides what you see online?
The most accurate answer is also the least satisfying. There is no single decision-maker. What you see is the result of a distributed process in which algorithms, interfaces, economic incentives, and human behavior interact continuously, producing outcomes that are structured but not centrally directed.
This does not mean that power has disappeared. If anything, it has become more pervasive. But it has also become more difficult to identify, because it no longer presents itself as control.
It presents itself as the natural shape of the system.
And perhaps that is the most important shift. We are no longer navigating a space where decisions are made in visible places. We are navigating a system that produces decisions as a byproduct of its own operation.
Which is why the question—who decides—feels increasingly hard to answer.
Not because there is no answer.
But because the answer no longer fits the way we expect decisions to work.
