Imagine two consumers looking at the same wine list. At the store, a French syrah sells for twenty dollars and an Australian shiraz for ten. The shopper believes the French wine is perhaps fifty percent better but twice the price, so they buy the shiraz. Weeks later, the same two wines appear in a restaurant, both marked up by forty dollars. Now the shopper believes exactly what they believed before—the French wine is still fifty percent better—but it is only twenty percent more expensive. They order the French wine.
Nothing about the shopper’s beliefs changed between the two visits. What changed was which difference stood out. The economists Pedro Bordalo, Nicola Gennaioli, and Andrei Shleifer use this example to open their model of salience in decision making, and it captures something that ordinary intuition keeps trying to deny: attention is a filter, and whatever gets through that filter is weighted as though it were more real than everything it displaced. This article asks why shifting what feels important can accomplish more than shifting what a person holds to be true.
Salience is not a metaphor
The most rigorous work on this subject comes from economics, where salience has been given a precise and testable meaning. Bordalo, Gennaioli, and Shleifer define something as salient when it stands out relative to a reference point, and they show that people overweight salient attributes when they choose. Their 2012 model, published in the Quarterly Journal of Economics, applies this to decisions under risk: people’s probability weights get distorted toward lottery states that grab attention, which helps explain familiar anomalies like the Allais paradox and preference reversals. Their 2013 companion paper in the Journal of Political Economy extends the same logic to ordinary purchases, where a good’s salient attributes are the ones furthest from the average of the choice set. The consumer, in their account, is a “local thinker” who overweights what is unusual about the options in front of them.
Two implications follow. The first is that context matters more than the individual usually perceives. Adding an irrelevant option to a menu can change what people buy, not because their preferences shifted, but because the new option changed which attribute now looks distinctive. The second is that salience is a lever independent of information or belief. A marketer does not have to convince you the French wine is better. They only have to make quality salient rather than price, and your choice moves.
This is the mechanism behind the title of this essay. If behavior is governed by what is salient, then whoever controls salience—whoever decides which facts, threats, opportunities, and social signals feel most pressing—holds a lever on behavior that does not require touching a single belief.
Salience as a brain process
The economics is a model of choice. Neuroscience supplies a plausible substrate, and it must be handled carefully because the temptation to overstate it is strong. In a widely cited 2007 paper in the Journal of Neuroscience, William Seeley and colleagues identified what they called a salience network: a set of brain regions, anchored by the dorsal anterior cingulate cortex and the orbital frontoinsular cortices, that co-activate when a person detects behaviorally relevant stimuli. The network has been associated with the process of deciding what deserves attention and what can be ignored, and the authors reported that individual differences in this circuitry correlated with measures of anxiety.
The careful reading is important. This is a discovery about which brain regions participate in the detection of relevance, drawn from structural and functional imaging at modest scale. It is not a description of a switch that an outsider can flip to decide what a person cares about. The value of the finding is that it makes the salience concept concrete: there is a real, measurable process by which some signals rise and others recede, and that process is part of how a brain allocates its limited capacity. When we later ask whether a technology could influence that process, the question has a physical referent rather than a purely philosophical one.
Why beliefs are the wrong target of attack
Most people assume that influence means changing what someone thinks. This assumption produces a great deal of misplaced anxiety and an equal amount of misplaced confidence. Belief change is hard, slow, and visible. It requires argument, evidence, or sustained social pressure, and it leaves a trail. The person who has been argued out of a position usually knows it. They can cite the moment, describe the reasons, and, if they change their mind again, explain why.
Salience works differently, and that difference is the whole point. A person whose beliefs are intact but whose sense of urgency has been rearranged will act differently while sincerely reporting no change of mind. The heavy smoker who knows the risks does not need to be persuaded that smoking harms them. They need the harm to feel abstract at the moment of lighting up and the pleasure to feel immediate. Whoever can tilt that balance in the moment of decision does not have to win an argument. They win the moment.
This is why moral language spreads the way it does. In a 2017 study in PNAS, William Brady, Julian Wills, John Jost, Joshua Tucker, and Jay Van Bavel analyzed more than 563,000 tweets about gun control, same-sex marriage, and climate change. They found that the presence of “moral-emotional” words—words like “hate” or “greed” that combine moral judgment with feeling—increased a message’s diffusion by roughly twenty percent per additional word. Crucially, the effect was bounded by group membership: moral-emotional language spread more within liberal and conservative networks than between them. The finding is not that people adopted new positions. It is that certain ways of framing what matters made those messages travel through networks of people whose beliefs were already fixed. Salience worked on the already-convinced, and it worked by making one more thing feel urgent enough to pass along.
Ranking is industrial-scale salience control
Before any neuroscience enters the picture, the modern attention economy already operates on salience at a scale no prior institution could match. A recommender system’s central function is to decide, out of effectively infinite content, what a person sees first. That is a salience decision by definition, made billions of times a day.
The best available evidence that this changes what people encounter comes from a 2022 PNAS study by Ferenc Huszár and colleagues. Exploiting a natural experiment on Twitter, they compared a randomized control group whose feed remained reverse-chronological against users exposed to the algorithmic timeline, covering roughly two million accounts across seven countries. The algorithmic feed amplified mainstream right-leaning political content more than mainstream left-leaning content in six of the seven countries, while showing no evidence that far-left or far-right content was amplified more than moderate content. The result is often flattened into a partisan talking point, which does the science a disservice. What it actually demonstrates is narrower and more important: the ranking algorithm changed which political content was salient to millions of people, and the direction of that change was specific to this platform and this period.
That framing matters because it separates the demonstrated claim from the speculative one. We know that ranking systems shape the information environment at population scale. We do not know that they rewrite anyone’s political beliefs. The evidence better supports a model in which the feed changes what feels relevant, talked about, and worth reacting to, which is exactly the salience channel.
At the level of the individual interface, the same principle operates through design. The Federal Trade Commission’s 2022 staff report, Bringing Dark Patterns to Light, documents interfaces that steer users by making one path frictionless and another exhausting. The report describes how some dark patterns roughly doubled sign-up rates compared with neutral designs, and how their effects compound when combined. Notice what a dark pattern is doing: it is not arguing for a choice. It is arranging the environment so that one option feels like the obvious next step. It manipulates salience by manipulating effort and prominence, and it does so without ever engaging the user’s beliefs.
The emotional channel
The purest demonstration that platforms can move a person’s felt state—closer to salience than to belief—remains the 2014 Facebook emotional contagion experiment by Adam Kramer, Jamie Guillory, and Jeffrey Hancock. Working with 689,003 users, the researchers reduced the proportion of positive or negative emotional content in the News Feed and found that users’ own subsequent posts shifted in valence accordingly. It is a small effect measured at enormous scale, and it is a direct test of whether manipulating what a person sees changes how they feel.
The study must be cited with its baggage. It drew intense ethical criticism and now carries a formal Expression of Concern from PNAS regarding informed consent, published in 2014. That is not a reason to discard the finding, but it is a reason to treat it as one contested data point rather than a settled law. The lesson the episode teaches cuts in two directions: large platforms can measurably shift the emotional tone of their users’ expression, and the ethical and scientific infrastructure for doing such experiments responsibly was, at the time, plainly inadequate.
Emotion is the connective tissue between salience and behavior because emotion is how urgency gets encoded. Something that feels bad or exciting or outrageous is, by definition, something the mind treats as important. An environment engineered to maximize engagement will disproportionately surface content that triggers strong feeling, not because anyone decided to make users angry, but because intense feeling is what the metric rewards. The metric does not need to know what a person believes. It only needs to know what makes them react.
Why salience control is hard to notice and hard to contest
If salience control is this powerful, why does it generate so little resistance? Four properties do the work.
It leaves no trace in the target’s self-report. A person can accurately describe their beliefs and be completely blind to the fact that their priorities were rearranged by a sequence of exposures. There is nothing to notice, because nothing they can introspect changed.
It exploits normal, healthy cognition. Attention is limited; that is not a bug to be fixed but a basic condition of being a mind. Any system that must allocate scarce attention has to privilege some inputs over others. Salience control hijacks a process we cannot opt out of.
It scales without cost. Once a ranking system is built, applying it to a billion users costs almost nothing additional. A human propagandist can reach thousands. A salience-controlling algorithm reaches everyone in the market, continuously, in their own language.
It is deniable. A platform can truthfully say it never told anyone what to think. It ranked content by predicted engagement. The claim is not false, and it is not the whole story either. This gap between the literal truth and the operative effect is where most modern influence lives.
To these, add the compounding of personalization. The 2017 experiments by Sandra Matz and colleagues showed that matching messages to personality profiles improves persuasiveness at scale, and the 2024 follow-up work she co-authored showed that large language models can generate such personalized messages automatically, with recipients not discounting them merely because a machine wrote them. Combine automated personalization with an engagement-ranked feed and you have a system that can select both which content is salient and how it is framed, tailored to an individual, at near-zero marginal cost.
The counterargument, taken seriously
The strongest objection to all of this is that it overreads the evidence and underrates human agency. On the evidence, the objection has real force. The effects being discussed are typically modest in magnitude. Salience and context effects, in the economics literature, are about how choices shift at the margin, not about overriding considered judgment. Microtargeting research, including the 2023 PNAS study by Ben Tappin and colleagues, shows that personalization can improve persuasion substantially yet produces no additional benefit from targeting more than one trait and yields advantages that depend heavily on context. The famous emotional contagion result is a small average shift in expressed valence, and it is ethically contested.
On agency, the objection is also partly right. People adapt. They notice that a feed makes them angry and step away. They develop norms for reading advertising skeptically. Salience control is not hypnosis, and describing it as such is its own kind of manipulation—one that flatters the audience’s helplessness while excusing its choices.
But the objection proves less than it seems. It shows that salience control is not total. It does not show that it is weak. Marginal effects, applied to billions of decisions over years, constitute the difference between one world and another. And the fact that adaptation is possible does not mean it happens by default, because the systems being adapted to are also adapting. The relevant comparison is not between a manipulated population and a perfectly rational one. It is between two information environments and the behavior each produces.
What salience control is not
It is worth being explicit about the claims this essay is not making, because the subject attracts more mythology than analysis.
It is not claiming that a hidden cabal decides what the public will care about. The mechanism is decentralized. Platforms compete for engagement; advertisers bid for attention; creators chase the algorithm; and the resulting salience landscape is an emergent product of millions of incentives rather than a master plan. That decentralized origin makes the phenomenon harder to fix, not easier, because there is no single author to hold responsible.
It is not claiming that beliefs are irrelevant. Beliefs set the boundaries within which salience operates. You cannot make a person act on an idea they reject outright. What salience determines is which of their existing beliefs and values gets to drive behavior at a given moment.
It is not claiming that salience control is new. Political movements, religions, and advertisers have always tried to make their concerns feel urgent. What is new is the instrumentation: continuous measurement, individualized delivery, and rapid feedback. The phenomenon is ancient; the precision is not.
And it is not claiming that controlling salience is inherently wrong. A good teacher makes the important thing feel important. A public health campaign that makes a neglected risk salient saves lives. The problem is not the mechanism but the alignment of the objective with the interests of the person whose attention is being shaped.
Where the lever actually sits
If salience is the lever, the strategic question is who gets to hold it and toward what end. Several concrete tests follow.
The first is whether people can see and adjust their own salience settings. The EU’s Digital Services Act gestures at this: Article 27 requires platforms to disclose the main parameters of their recommender systems in plain language, and Article 38 requires very large platforms to offer at least one recommender option that is not based on profiling. The second provision is the more interesting one, because it creates a genuine alternative to the optimized feed rather than just a disclosure about it. A feed you can inspect is not the same as a feed you can leave.
The second is whether ranking objectives are governed rather than merely published. Disclosure tells us that engagement drives ranking; it does not change what engagement-driven ranking does. The deeper question is whether platforms can be required to optimize for something other than the metric that rewards outrage—whether salience can be pointed at a target that serves the user’s stated goals rather than the platform’s engagement numbers.
The third is whether the effects compound across systems. A person does not encounter one algorithm. They encounter a phone, a feed, a streaming queue, a news app, each independently tuned to capture attention. Individually defensible choices can add up to an environment no one designed and no one governs. Regulation built around single platforms may miss the cumulative effect.
The fourth is literacy that is honest about mechanism. Telling people “algorithms influence you” is nearly useless, because it offers no test and no response. Telling them specifically how salience works—that context and framing shift choices without changing beliefs, and that the tell is a decision that feels automatic and a priority that arrived from nowhere—gives them something to do with the knowledge.
What follows
The uncomfortable implication of salience research is that the most effective influence does not need to touch what a person believes, and therefore does not need to lie. It needs only to decide what they notice. A system can be scrupulously truthful, present only accurate information, and still shape behavior profoundly by choosing which true things to put in front of which person at which moment. This is why fact-checking alone cannot address the problem: correcting a false belief is irrelevant if the operative influence was never about belief in the first place.
The boundary that matters is the same one that runs through the rest of this series. Attention is the scarce resource from which every other mental capacity is built. Whoever allocates it exercises a form of power over human beings that is quieter than coercion, more durable than argument, and more difficult to contest than censorship, because nothing is forbidden and nothing is false. The question is not whether such power will exist. It already does. The question is whether the people whose attention is being allocated will have any say in what it is allocated toward, and whether the institutions that shape that allocation answer to anything beyond the next quarter’s engagement.
Sources and further reading
- Bordalo, P., Gennaioli, N., & Shleifer, A. (2012). “Salience Theory of Choice Under Risk.” The Quarterly Journal of Economics, 127(3), 1243–1285. https://academic.oup.com/qje/article-abstract/127/3/1243/1922370
- Bordalo, P., Gennaioli, N., & Shleifer, A. (2013). “Salience and Consumer Choice.” Journal of Political Economy, 121(5), 803–843. https://www.journals.uchicago.edu/doi/10.1086/673885
- Seeley, W. W., Menon, V., Schatzberg, A. F., Keller, J., Glover, G. H., Kenna, H., Reiss, A. L., & Greicius, M. D. (2007). “Dissociable intrinsic connectivity networks for salience processing and executive control.” Journal of Neuroscience, 27(9), 2349–2356. https://pmc.ncbi.nlm.nih.gov/articles/PMC2680293/
- Huszár, F., Ktena, S. I., O’Brien, C., Belli, L., Schlaikjer, A., & Hardt, M. (2022). “Algorithmic amplification of politics on Twitter.” Proceedings of the National Academy of Sciences, 119(1), e2025334119. https://www.pnas.org/doi/10.1073/pnas.2025334119
- Brady, W. J., Wills, J. A., Jost, J. T., Tucker, J. A., & Van Bavel, J. J. (2017). “Emotion shapes the diffusion of moralized content in social networks.” Proceedings of the National Academy of Sciences, 114(28), 7313–7318. https://www.pnas.org/doi/10.1073/pnas.1618923114
- Kramer, A. D. I., Guillory, J. E., & Hancock, J. T. (2014). “Experimental evidence of massive-scale emotional contagion through social networks.” Proceedings of the National Academy of Sciences, 111(24), 8788–8790. (See the accompanying Expression of Concern: https://www.pnas.org/doi/10.1073/pnas.1412469111) https://www.pnas.org/doi/10.1073/pnas.1320040111
- Federal Trade Commission. (2022). Bringing Dark Patterns to Light (staff report). https://www.ftc.gov/reports/bringing-dark-patterns-light
- Tappin, B. M., Wittenberg, C., Hewitt, L. B., Berinsky, A. J., & Rand, D. G. (2023). “Quantifying the potential persuasive returns to political microtargeting.” Proceedings of the National Academy of Sciences, 120(25), e2216261120. https://www.pnas.org/doi/10.1073/pnas.2216261120
Loading comments…