Where Grievance Finds a Story

The injury usually arrives before the causal diagram. A person feels the bill, the layoff, the rent increase, or the disappearing margin in their life long before they can map the machinery that produced it.

The bill is higher. The job is less secure. The rent outruns the raise. The product gets smaller. The service gets worse. The institution says the relevant explanation lives somewhere inside a thicket of market conditions, regulatory constraints, financing costs, technical limitations, macroeconomic forces, and specialist vocabulary.

That gap creates a market for stories. Complexity leaves an empty chair at the table, and somebody will happily sit in it.

Some stories help people trace causes. Others hand the pain a convenient target. The more difficult the real mechanism is to inspect, the easier it becomes to redirect anger toward whichever person, group, symbol, or outrage is easiest to recognize.

Economic conditions can change political behavior, but the relationship is not mechanically uniform. Autor, Dorn, Hanson, and Majlesi examine historical US trade exposure and political outcomes. Their findings connect greater exposure to import competition with changes in representation and political alignment, with important differences across places. The periods in that research precede the current AI transition. The study does not identify a preferred political response or predict a future election. [S142]S142 · Importing Political Polarization? The Electoral Consequences of Rising Trade ExposureDavid Autor; David Dorn; Gordon Hanson; Kaveh Majlesi · 2020peer-reviewed political economy · economic-grievance evidenceClose

The useful point is direct: material conditions can change the environment in which people interpret institutions and political choices. The story attached to that pain then comes from somewhere else: history, identity, organizations, friends, media, political rhetoric, recommendation systems, and whatever explanation reaches the person at the right emotional moment.

The story that keeps the anger horizontal

This is the part I find most corrosive.

When the causes of financial pressure are distributed across monetary policy, tax rules, procurement systems, market concentration, lobbying, labor arrangements, pricing systems, and decades of institutional decisions, there is no single face to be angry at. The machinery is real and the pain is real, but the causal map is expensive to learn.

Culture-war stories are cheap. They arrive with names, faces, slogans, and no requirement that anyone understand a balance sheet.

They arrive prepackaged with a villain, a tribe, a vocabulary, a set of grievances, and somebody to blame by dinner. A person who is anxious about housing, wages, healthcare, debt, status, or the future does not need to understand a balance sheet before a much simpler story tells them who supposedly took something from them.

That is what I mean when I say opacity and propaganda can work together even without a master plan. Institutions can benefit from complexity. Political actors can benefit from grievance. Media systems can benefit from attention. Platforms can benefit from engagement. Each actor can pursue its own objective and the combined system can still redirect social anger sideways, toward other people living under the same pressures.

The result is grotesquely efficient. Structural pain becomes interpersonal hostility. Economic anxiety becomes identity conflict. Institutional distrust becomes a search for traitors. The machinery stays abstract. The neighbor gets a face.

The story that travels

Research on moral-emotional language helps explain one part of the attention process. Brady and colleagues’ 2017 study found that such language was associated with greater diffusion of messages about polarizing issues, particularly within ideological networks. Their later work examined how social feedback and network norms can shape the expression of outrage. These findings concern expression and diffusion. They do not establish that every shared message changes a belief. [S148][S149]S148 · Emotion shapes the diffusion of moralized content in social networksWilliam J. Brady et al. · 2017PNAS empirical study · attention/diffusion mechanismS149 · How social learning amplifies moral outrage expression in online social networksWilliam J. Brady et al. · 2021Science Advances preregistered studies · platform-learning mechanismClose

A 2025 preregistered replication and meta-analysis combined 27 studies and about 4.8 million observations. Its pooled estimate associated an additional moral-emotional word with roughly 13% greater expected sharing, with substantial variation across contexts. The measure is an incidence-rate ratio for sharing, not a percentage increase in polarization, anxiety, or persuasion. The paper explicitly treats the meta-analytic association as distinct from a causal effect. [S150]S150 · Estimating the effect size of moral contagion in online networks: a preregistered replication and meta-analysisWilliam J. Brady et al. · 2025peer-reviewed replication/meta-analysis / PNAS Nexus · replication evidenceClose

Those are different stages. A message can be shown, noticed, clicked, shared, believed, remembered, or acted on. Sharing is not believing. Clicking is not endorsing. Exposure is not persuasion. An engagement metric usually sees only part of that chain.

The commercial problem begins when the observed stage becomes the target. A system rewarded for sharing can get extremely good at producing more sharing without learning whether the message is useful, representative, calming, truthful in proportion, or remotely good for the person consuming it. The metric can count the handoff. It cannot tell you what the handoff did.

This mechanism is broader than any ideology or platform. People bring their own commitments and conflicts. The ranking system decides which parts of that human material receive repeated distribution. It does not create every grievance. It can decide which grievances become ambient.

Attention is not a representative sample

A feed is a selection of events, not a census of ordinary life. It is a highlight reel with no obligation to show you the denominator. A person may encounter many unusual incidents because each is individually compelling. The selection can be accurate at the level of the incidents and misleading as an impression of prevalence.

Consider an illustrative stream containing a sequence of rare failures. Each report is real. The stream leaves out the much larger number of uneventful interactions. A reader may come away believing that failure is typical. The problem is not necessarily fabrication. It is the missing denominator. Reality has been cropped.

The same issue appears in an AI answer that retrieves vivid examples without explaining how they were selected. A fluent synthesis can turn a collection of memorable cases into an apparent general rule. The answer may need fewer examples and a better account of representativeness rather than more confident prose.

This connects directly to the earlier discussion of corporate reputation. Visible scandals are not a random sample of all conduct. Visible acts of generosity are not a complete inventory of an institution’s behavior. A system that learns from what becomes public inherits the selection process as well as the content.

The problem becomes harder when feedback changes production. If a certain expression reliably receives attention, people may produce more of it. The later archive then contains both the original conditions and the adaptation to the attention system. A model trained on that archive can learn what gets expressed, not merely what people privately believe or how often the described events occur.

That is a plausible synthesis of selection and feedback, not an estimate of one universal platform effect. The empirical task is to identify the mechanism in a particular setting and distinguish it from other causes.

A result that complicates the easy story

A large Facebook field experiment during the 2020 US election period reduced participants’ exposure to like-minded sources by about one-third. It changed the information environment and reduced exposure to uncivil content and sources repeatedly sharing misinformation. It did not detect effects on eight preregistered political-attitude measures, including measures of affective polarization and ideological extremity. The intervention involved 23,377 participants and ran from late September through late December 2020. [S143]S143 · Like-minded sources on Facebook are prevalent but not polarizingBrendan Nyhan et al. · 2023Nature field experiment · counterevidence / platform mechanismClose

That result breaks the easy chain from exposure to attitude. The feed changed substantially. The preregistered political-attitude measures did not move detectably in the experiment. [S143]S143 · Like-minded sources on Facebook are prevalent but not polarizingBrendan Nyhan et al. · 2023Nature field experiment · counterevidence / platform mechanismClose That matters. The mechanism is more complicated than “algorithm shows outrage, therefore person becomes polarized.”

The larger system has several interacting causes. A person encounters information through friends, family, work, media, organizations, prior commitments, and platforms. A feed enters that ecology. So does a politician. So does a boss. So does a chatbot.

It also illustrates why evidence should correct the mechanism locally. A null result for an attitude measure does not erase evidence that the feed changed. Evidence that a feed changed does not authorize a claim that an attitude changed. Keeping both findings intact gives a more accurate account than either celebration or dismissal.

Connection has a quality as well as a quantity

More communication is not automatically better connection. A person can exchange many messages and still lack the kind of relationship they need. Another can find real community through a digital service. The medium alone does not determine the experience.

A 2025 longitudinal meta-analysis reports small bidirectional associations between loneliness and problematic media use. The direction runs both ways in the studies considered: loneliness can precede problematic use, and problematic use can precede later loneliness. The reported beta-based effects are around 0.09 to 0.10. That is not evidence that ordinary media use has the same effect for everyone. [S144]S144 · Loneliness and Problematic Media Use: Meta-Analysis of Longitudinal StudiesJia Yuin Fam; Niko Mannikko · 2025peer-reviewed longitudinal meta-analysis · psychosocial evidenceClose

Research on doomscrolling likewise includes associations with anxiety and pessimism, but observational designs leave important questions about direction and confounding. A person who is already distressed may seek more threatening information. Repeated exposure may also affect their experience. A correlation alone does not allocate the contribution between those possibilities. [S145]S145 · Doomscrolling evokes existential anxiety and fosters pessimism about human nature? Evidence from Iran and the United StatesReza Shabahang et al. · 2024peer-reviewed behavioral study · psychosocial evidenceClose

The systems question is what gets rewarded. If a service measures continued use more readily than the quality of the user’s life outside the service, it can have an incomplete account of success. A person spending less time on a platform because they found what they needed may look less valuable under an engagement metric than a person who remains unsettled and keeps searching.

No executive needs to write “make users miserable” into a quarterly plan. A system can simply fail to distinguish satisfied departure from frustrated departure, or useful engagement from compulsive return. If continued use is the measurable objective, misery can become commercially indistinguishable from enthusiasm. The metric knows that you stayed. It does not know why.

The conversation changes the relationship

A feed presents material. A conversational system can respond to the person’s own account, remember context within its available memory, and help construct an explanation. That gives it a different role in the person’s thinking. The benefit can be substantial: a patient explanation, help organizing a problem, or a chance to explore an idea without embarrassment.

It also changes the route by which influence can occur. The system does not merely choose which item to display. It can choose which distinctions to introduce, which possibilities to emphasize, and which questions to ask next. Those choices can shape the path of a conversation even when no explicit recommendation is made.

The OpenAI and MIT Media Lab research published in March 2025 found that affective use was uncommon overall and concentrated among a smaller subset of users. Its observational analyses and randomized study answer different questions. Associations involving heavier voluntary use should not be described as effects caused by random assignment. The findings do not support a universal claim that chatbot use makes people lonely or dependent. [S146][S147]S146 · Investigating affective use and emotional well-being on ChatGPTOpenAI · 2025first-party large-scale usage study · emerging evidenceS147 · How AI and Human Behaviors Shape Psychosocial Effects of Chatbot UseMIT Media Lab · 2025longitudinal controlled study · emerging causal/observational evidenceClose

The narrower concern is worth keeping. A system that becomes useful in emotionally important moments may acquire a kind of trust that exceeds its reliability in a later task. The user may have good reasons for trusting some functions and still need a way to recognize when the system is outside those functions. The transition can be difficult to notice because the interface and voice remain the same.

Imagine someone who has received excellent writing help and then asks the same assistant to interpret a difficult relationship or a complex financial event. The earlier success is evidence about the earlier task. It does not automatically validate the later interpretation. Trust likes to travel lighter than evidence does. A single conversational relationship can span domains with very different evidence requirements.

I care about that distinction because I use these tools constantly. Their usefulness is the reason the issue matters. Nobody builds a civilization-scale dependency on a system that is obviously useless.

The missing option can matter more than the wrong fact

An answer can be accurate in every sentence and still present an incomplete choice. A menu can be perfectly truthful about every item printed on it while quietly omitting the door. Suppose a customer asks how to lower a bill. The assistant accurately describes three plans from one provider but never considers cancellation, a different provider, or a service that meets the need in another way. Whether that answer is adequate depends on the task and the assistant’s role.

A seller-side assistant may be doing exactly what it was built to do. A user-side adviser would be leaving out relevant alternatives. The factual accuracy of the descriptions does not settle the conflict. The issue is the boundary around the options.

The same problem can occur in explanations of social conditions. A system can accurately summarize one causal account while omitting another that materially changes the interpretation. A responsible analysis does not require equal space for every possible claim. It requires attention to evidence that changes a load-bearing explanation.

That is not a demand to flatten every dispute into “both sides.” Weak claims do not deserve equal weight because someone says them loudly. Evidence that changes a load-bearing explanation does deserve attention. The Facebook experiment constrains one causal story without making the attention economy benign.

From repeated exposure to repeated interpretation

The future issue is not only how often people encounter a message. It is how often they rely on one system to organize what messages mean. Repeated interpretation can create a stable frame: which causes are salient, which alternatives seem practical, which uncertainties matter, and which questions go unasked.

This is an extrapolation about a possible form of dependency. It does not establish that current users have lost independent judgment. People can question tools, consult other sources, and use assistance to become more informed. The question is what conditions make that independence easier or harder to exercise.

A system that exposes sources, separates evidence from interpretation, and helps a person compare alternatives can increase agency. An opaque system that continuously personalizes the explanation can do the reverse. The dangerous version does not need to lie very often. It only needs to become the place a person habitually goes to decide what things mean. Interpretation can become infrastructure before anyone notices they have started depending on the plumbing.

The transition from platform to reasoning infrastructure therefore deserves its own analysis. The important asset is no longer only a stream of attention. It can be the place where a person turns to decide what a situation means. Whoever operates that infrastructure has responsibilities that cannot be measured solely by response quality, engagement, or the number of tasks completed.