For most of the twentieth century, mass media concentrated the power to decide what large numbers of people would see. Newspapers, radio, and television set a common field of attention. The choice was coarse. An editor or a network decided what entered the shared space. People could still turn the page, change the channel, or leave the room.
What replaced that system is different in kind. Contemporary platforms do not primarily broadcast a single message. They continuously allocate individual attention according to predictions of what will keep each person engaged longer. The ranking function is the product. The data extracted from watching, pausing, sharing, and reacting improves the next prediction. Advertising is how the resulting attention is sold. In 2025 Meta’s advertising revenue reached $196.2 billion. U.S. social media advertising overall reached $117.7 billion. These are not side effects. They are the business model.
This is the aha. The systems are not neutral pipes. They are attention markets. What gets seen is what the ranking function predicts will keep eyes on the screen. Responsibility requires seeing. When the field of attention is shaped by engagement signals, large parts of ordinary reality become optional. Local conditions, long-term costs, other people’s actual experience, the state of the ground and water and air — these drop out unless they happen to generate rapid reaction. The gap is between the power to shape the field and the willingness to account for what the field does to the capacity of people to see and answer for what is in front of them.
The technical literature is consistent on the mechanism. Engagement-based ranking amplifies content that produces strong, rapid reactions. A pre-registered audit of Twitter’s ranking system found that, relative to a reverse-chronological baseline, the engagement-ranked feed amplified emotionally charged, out-group hostile content. Political posts were more likely to express anger and hostility toward the opposing group. Users reported feeling worse about that group afterward. They did not, when asked, prefer the political content the algorithm had selected. Separate experiments show that simply lowering a post’s rank can cut engagement by roughly 40 percent even when the content is identical. The architecture steers attention before conscious preference is consulted.
These effects are not confined to politics. Multiple systematic reviews and meta-analyses find positive associations between high-engagement or problematic social media use and internalizing symptoms such as depression and anxiety in adolescents. A 2024 JAMA Pediatrics meta-analysis covering 143 studies and more than a million adolescents reported significant associations in both clinical and community samples. Effect sizes are often modest. Heterogeneity is high. Causation is not uniform. The pattern is nonetheless recurrent enough that treating it as noise is no longer credible. Problematic use and passive consumption show the more consistent links.
The economic logic is straightforward. Attention is costly for the user and valuable to the platform. A platform can monetize it directly through advertising or allocate it to creators to induce more content, which generates more attention. Revenue-maximizing systems will sometimes surface content that is suboptimal for particular viewers in order to strengthen creator incentives. Higher advertising profitability increases the value of each unit of attention and shifts incentives further toward extraction. The externalities — thinner sustained focus, elevated affective polarization, documented associations with adolescent mental-health burdens — are borne elsewhere. They do not appear as costs on the platform’s income statement.
Recent decisions illustrate the gap between stated corrections and unchanged incentives. In January 2025 Meta ended its U.S. third-party fact-checking program and replaced it with a Community Notes system modeled on X. The company described the change as a restoration of free expression. Content policies on immigration, gender, and other topics were simplified. What remained was the ranking system optimized for engagement. Meta’s Oversight Board later noted delays in note publication, the limited number of published notes, and dependence on the broader information environment. In the first six months of the U.S. rollout roughly 900 notes were published. Professional fact-checking labels had previously been applied to tens of millions of posts in the European Union over comparable periods. The moderation layer changed. The ranking objective did not.
X under Elon Musk followed a related path. The public case centered on free speech and reduced institutional bias. Community Notes were retained. Verification became paid. Advertising relationships shifted. The ranking function continued to privilege rapid engagement. Claims of restored speech and the continued operation of an attention market are not contradictions. They describe different layers of the same architecture.
Shoshana Zuboff’s account of surveillance capitalism supplies the broader frame. Platforms extract behavioral data not only to target ads but to predict and increase the probability of profitable next actions. When the prediction engine is also the distribution engine, observation and governance of attention converge. The result need not be classical censorship. It is a continuous, personalized reallocation of the field in which people form judgments.
This is the sense in which the present system functions as control of attention rather than control of data. It requires no single central directive. It requires only that the dominant systems for allocating attention remain optimized for engagement and advertising yield while the costs remain external. The language of free expression, community moderation, or user choice can coexist with that structure. The structure itself determines which signals travel, which fade, and which capacities for attention and responsibility are strengthened or eroded.
The people who set the objective functions are not passive. They choose what to measure and what residual effects to accept. Those choices can be examined directly.
What evidence shows that current ranking objectives leave users more capable of sustained attention to the conditions that affect their lives and the living systems they depend on?
When a moderation system is replaced and the change is called freedom, what ranking function remains, and who is accountable for the distribution patterns it continues to produce?
How are the documented associations between high-engagement use and elevated rates of depression and anxiety in adolescents incorporated into design decisions, and where do those costs appear in internal accounting?
If the capacity of people to notice and respond to local and planetary conditions were treated as a constraint rather than an externality, what specific changes to optimization targets would follow?
These are design questions. They do not require assuming malice. They require treating the measured effects of attention allocation as relevant to the people who control the allocation. The technical capacity to build systems at this scale has already been demonstrated. Whether that capacity will be used to internalize the costs it currently externalizes remains an open decision.
The work is the gap.
— Sean Claesgens
Claesgens Integrated, LLC