Revenue = Users × Time × Ad Rate. The variable platforms control is time — and time requires emotional states.
An advertising-dependent platform's revenue can be expressed as a product of three variables: the number of users on the platform, the time each user spends there, and the revenue the platform extracts per unit of that time (a function of inventory price, as described in AE-002). All three matter. But they are not equally tractable.
User count growth is bounded by global demographics and market saturation. Once a platform has reached a substantial fraction of its addressable market — as Meta, YouTube, and Google have — marginal user acquisition becomes expensive and slow. The platform cannot double its user base again without a market-changing event.
Ad rates are partly exogenous. They reflect advertiser demand, competition for ad budgets among platforms, macroeconomic conditions, and the targeting precision of the platform's data infrastructure. Platforms can improve targeting quality, but ad rate improvements are constrained by advertiser willingness to pay and competitive alternatives.
Time-on-platform is the variable the platform controls most directly through design. The feed algorithm, the notification architecture, the content recommendation system, the infinite scroll implementation, the autoplay defaults — these are all mechanisms for controlling how long users stay. Time is the lever. Everything else follows from the revenue function's math: more time per user means more inventory, and more inventory means more total revenue regardless of user count.
This is why the most senior product and engineering resources at attention-economy platforms are concentrated on engagement: not because engagement is intrinsically valuable, but because engagement is the primary driver of the revenue variable that the platform controls. It is not that platforms are indifferent to users. It is that the revenue function rewards them for treating user time as the primary resource to be maximized.
In the internal vocabulary of major platforms, "time-on-platform," "daily active users," "average session duration," and "daily active users to monthly active users ratio" (DAU/MAU) are the primary operational metrics. These are not the only metrics tracked — conversion rates, advertiser ROI, content upload volumes, and many others are also monitored — but they are the metrics most directly linked to revenue and therefore most directly linked to executive accountability and engineering priority.
Former platform employees and executives have described in congressional testimony and published accounts the degree to which these metrics governed product decisions. When Facebook's early News Feed algorithm produced shorter average sessions because users had read their feed and left, the response was not to treat this as evidence of user satisfaction with the product. It was to redesign the feed to surface more content before users reached the end of their available feed. The implicit assumption — that more time on platform is always better for the platform — is structural, not accidental.
The metrics cascade through the organization. Product managers are evaluated against engagement targets. Algorithm teams are measured by session duration impact. Content policy teams must navigate between removing harmful content (which sometimes reduces engagement) and maintaining the algorithmic infrastructure that makes engagement optimization possible. In this environment, the implicit answer to "how much harm is acceptable?" is always framed against "how much engagement would be sacrificed to prevent it?"
Platform feed algorithms are, in simplified terms, systems for predicting which content, shown to which user at which moment, will maximize the platform's target engagement metric. The specific metric varies: Meta's News Feed has at various times optimized for clicks, comments, reactions, "meaningful social interactions" (weighted interactions between friends and family), and video view time. YouTube's recommendation algorithm optimizes for watch time and session length. TikTok's For You page optimizes for video completion rate and re-watch rate.
In each case, the algorithm is a prediction machine: it predicts which content will produce the most of the target metric and ranks content accordingly. Over time, the algorithm learns what kinds of content produce high metric returns. The learning is not value-neutral. It is metric-constrained: the algorithm learns to favor content that works by the measure it is trained to maximize.
Research examining what kinds of content these algorithms favor consistently finds an emotional content premium. Content that elicits strong emotional responses — whether positive (inspiration, awe, humor) or negative (anger, fear, disgust, outrage) — generates more of every engagement signal than content that leaves users in neutral emotional states. The algorithm, optimizing against engagement signals, learns to surface emotionally activating content without any explicit instruction to do so. The outcome is structural: the recommendation infrastructure, left to optimize its stated metric, will converge on emotionally charged content because that content performs better by the metric that the revenue function requires the algorithm to maximize.
The evidence for an emotional content premium in platform distribution is robust and multi-sourced. Research published in peer-reviewed journals, platform-published data, internal documents released through litigation, and documented A/B test outcomes converge on the same finding: emotional arousal increases engagement signals across virtually every platform and every content category studied.
Studies of Twitter/X sharing behavior find that tweets incorporating moral-emotional language — words conveying virtue or vice, care or harm, loyalty or betrayal — generate systematically higher retweet rates than their matched non-emotional equivalents. Work by NYU researchers documented approximately 20% more retweets per unit of moral-emotional language. The effect is consistent across political affiliations, content categories, and time periods.
Research on YouTube recommendation behavior documents that the recommendation system systematically surfaces more extreme, more emotionally intense versions of content adjacent to what a user initially watches. A viewer watching moderate political commentary is more likely to be recommended progressively more extreme content in subsequent recommendations — not because platform policy favors extremism, but because more extreme content generates higher watch time and completion rates among audiences already interested in the topic.
Facebook's own internal research, partially disclosed through the Wall Street Journal's 2021 reporting, showed that the algorithm's "meaningful social interaction" optimization — intended to favor friends-and-family content over passive news feed consumption — increased the prevalence of divisive and emotionally charged political content because divisive content generates more comments and reactions than neutral content, and comments and reactions were the metric "meaningful" was operationalized to measure.
Outrage is a specific case of the emotional content premium with particular structural importance. Content that elicits outrage — moral indignation at a perceived injustice, violation of norms, or threatening behavior — generates several engagement behaviors simultaneously and at high rates:
The profitability of outrage is not a platform intent. No platform has, to the available record's knowledge, implemented a policy to favor outrage content. The profit is structural: the revenue function rewards engagement, engagement algorithms learn the content that maximizes it, and outrage content maximizes most engagement metrics most of the time. The algorithm did not decide to favor outrage. It learned that outrage works.
Platforms now explicitly downrank engagement-bait and outrage content. Meta's "meaningful social interactions" update, YouTube's responsibility algorithm changes, and TikTok's content quality signals all represent genuine attempts to decouple engagement metrics from harmful content amplification.
The objection accurately describes policy intentions and some real algorithmic changes. Platforms have made genuine adjustments, and some categories of engagement-bait content have been measurably reduced. The structural problem is not that platforms are lying when they describe these changes. It is that any algorithm optimizing against engagement metrics in a human behavioral environment will encounter the same empirical reality: emotionally activating content generates higher engagement signals. The platforms can modify which engagement signals they optimize for and how they weight them, but they cannot eliminate the underlying behavioral fact that human attention is captured most effectively by emotional arousal. So long as the revenue function depends on attention time, the pressure to surface engagement-maximizing content will remain — attenuated by policy interventions, but not eliminated.
The revenue function creates a structural dilemma for platforms with genuine wellbeing commitments: any significant reduction in engagement optimization directly reduces revenue. A platform that replaced its engagement-optimized feed with a chronological feed, eliminated autoplay, added friction to extended sessions, and disabled emotionally targeted notifications would produce, by most estimates, a substantial reduction in time-on-platform — and therefore a proportional reduction in advertising revenue.
This reduction cannot be absorbed through efficiency gains in the other revenue variables. User count is saturated. Ad rates cannot rise fast enough to compensate for time lost. The revenue function does not permit a significant engagement reduction without a proportional revenue reduction, absent a new revenue model.
This is the dependency the named condition describes: the platform's revenue is structurally dependent on engagement optimization, which is structurally dependent on emotional activation, which produces documented harms as an operational output. The dependency is not a choice made by bad actors. It is the predictable consequence of building a business on the sale of human attention in a competitive market.
Internal: This paper is part of The Attention Economy (AE series), Saga VIII. It draws on and contributes to the argument documented across 55 papers in 12 series.
External references for this paper are in development. The Institute’s reference program is adding formal academic citations across the corpus. Priority papers (P0/P1) have complete references sections.