24 Aug 2026 20 min read Institutional Decay

America is at War With Itself

"Our algorithms exploit the human brain's attraction to divisiveness." -Internal Facebook Presentation, 2018.

American flag pictured between strands of barbed wire fence.

By Matt Stone

In September 2025, Pew asked 3,445 Americans whether politically motivated violence was increasing in this country. Eighty-five percent said yes. Republicans said yes at 86 percent. Democrats said yes at 85 percent.

That is as close to national consensus as anything gets anymore.

Then Pew asked why. Democrats' most common answer was rhetoric from Trump and Republicans. Republicans' most common answer was rhetoric from Democrats and liberals. Fifty-three percent called left-wing extremism a major problem. Fifty-two percent said the same about right-wing extremism.

Both sides agree the house is on fire. Both sides think the other one is holding the match. Everybody wants safety, security, food on the table, the ability to take care of their families. Why do we think the other half of the country does not want the same? Personally, I live by the principle that if someone is not hurting anyone else, then they should be left the fuck alone. That concept seems so foreign to Americans today. For some reason, it has become incredibly easy to turn Americans against their neighbors. In the following sections, I break down why.

Part I: The Wiring Is Not Broken

Start with the thing nobody wants to hear: the instincts tearing this country apart are not defects. They are solved problems. They are ancient answers to questions our ancestors could not afford to get wrong.

The evolutionary psychologist John Tooby spent his career on what he called coalitional instincts. His argument, laid out in a 2017 essay, is that humans are built to form alliances, and that alliance membership has to be signaled or it is worthless. As he put it: "To earn membership in a group you must send signals that clearly indicate that you differentially support it, compared to rival groups."

The uncomfortable implication is that Tooby argued the best loyalty signals are the ones that cannot be explained any other way. "Communications of practical and functional truths are generally useless as differential signals, because any honest person might say them regardless of coalitional loyalty." Saying the sky is blue proves nothing about your allegiance. Saying something absurd, something no neutral observer would ever say, proves everything.

Which means that in a sufficiently sorted society, the incentive is not to be right. The incentive is to be unmistakably ours. Terminally unique. Belief becomes a uniform.

Underneath that sits a more basic piece of hardware. In 2001, Robert Kurzban, Tooby, and Leda Cosmides ran an experiment in PNAS testing whether the mind's tendency to sort people by race is fixed or contingent. It turned out to be contingent. In the authors' own summary, by "creating a social context in which race was uncorrelated with coalitional alliances, we were able to drastically decrease the extent to which subjects noticed and remembered other people's race." A 2020 meta-analysis found the effect holds up across contexts, though it is small: a correlation of about 0.137 across nine effect sizes from five studies, dropping to roughly 0.106 once you correct for likely publication bias.

Small, but real, and the implication is enormous. The mind is not tracking race. It is tracking sides, and it will grab whatever marker is available to figure out which side you are on. Give it a better marker and it drops the old one.

For most of American history, the available markers were geography, church, union hall, and race. Over the last fifty years, they collapsed into one. Lilliana Mason's work on social sorting documents it: religious, racial, and ideological identities that used to cut across party lines now stack neatly behind them. Using ANES data, she found anger toward opposing candidates rose from about 40 percent in 1980 to about 60 percent in 2012, and that people holding cross-pressured identities showed dramatically dampened emotional responses, sometimes near zero.

That is the whole tragedy in one finding. The person who is a union Republican, or a churchgoing Democrat, cannot get as angry, because getting angry costs them a piece of themselves. We have been steadily eliminating those people. Not by force. By sorting.

The consequences are measurable and they are not confined to politics. Iyengar and colleagues, reviewing the field in 2019, found the partisan feeling-thermometer gap grew from 22.64 degrees in 1978 to 40.87 degrees in 2016, driven almost entirely by growing coldness toward the other side rather than warmth toward your own. In 1960, 4 to 5 percent of Americans said they would be upset if their child married across party lines. By 2010, that was roughly a third of Democrats and half of Republicans.

One honest caveat, there is a popular claim that conservatives are wired for higher threat sensitivity than liberals. It does not hold. A 2022 Nature Human Behaviour study using five behavioral measures across roughly 4,000 U.S. respondents found "no consistent evidence" for a relationship between negativity bias and ideology. A 17-country physiological study in the American Political Science Review, measuring skin conductance while people watched negative news, failed to find strong support either.

The wiring is not partisan. It is human. Both sides have it. That is exactly what makes it exploitable.

Part II: The Amplifier

Now put that wiring in front of a system whose only job is to predict what you will react to.

In 2021, Steve Rathje, Jay Van Bavel, and Sander van der Linden analyzed 2,730,215 posts from news accounts and members of Congress across Facebook and Twitter. Posts about the political out-group were shared or retweeted roughly twice as often as posts about one's own side. Each additional out-group word increased the odds of sharing by 67 percent.

The comparison numbers are the ones that matter. Out-group language was 4.8 times stronger a predictor of engagement than negative emotional language, and 6.7 times stronger than moral-emotional language. It was not enough to be angry. It was not enough to be righteous. You had to be angry at them.

Facebook knew. In 2018, an internal presentation reported by the Wall Street Journal in May 2020 contained a slide reading: "Our algorithms exploit the human brain's attraction to divisiveness." The report warned that left alone, the platform would feed users "more and more divisive content in an effort to gain user attention." A team called Common Ground was assigned to fix it. The team was dissolved. Executives declined the changes, in part out of concern that the fixes would disproportionately affect conservative accounts and produce accusations of bias.

It is important to be precise about how this happened, because the sloppy version of the story is worse than useless. In 2017, Facebook weighted the new reaction emojis at five times the value of a like. Not just the angry one. All of them: love, haha, wow, sad, angry, weighted identically. A staffer asked in writing whether this would "lead to News Feed having a higher ratio of controversial than agreeable content." By 2019, company data scientists had confirmed that posts drawing angry reactions were disproportionately likely to contain misinformation, toxicity, and low-quality news. The angry weight was eventually set to zero, and when it was, users saw less misinformation and less graphic content with no measurable drop in engagement.

Nobody sat in a room and decided to weaponize rage. Somebody made an arbitrary product decision, the optimizer ran with it for three years, and the wiring did the rest. That is the actual mechanism, and it is more disturbing than conspiracy, because conspiracy at least implies someone is steering.

The 2018 "meaningful social interactions" change did the same thing at scale. Jonah Peretti of BuzzFeed told Facebook directly that the change was "forcing us to produce worse content." An internal Facebook research memo, reported by the Wall Street Journal in September 2021, recorded what happened in Poland: "One party's social media management team estimates that they have shifted the proportion of their posts from 50/50 positive/negative to 80% negative, explicitly as a function of the change to the algorithm." Reporting by the Washington Post on the same document trove found Poland's two largest parties, Law and Justice and Civic Platform, calling the situation unsustainable and describing an online "social civil war."

Read that again. A political party rebuilt what it said to the public around a ranking function, and according to the same reporting, some European parties shifted not only their messaging but their actual platforms. The human was not removed from the process. The human was moved to the end of it.

And the machine is still doing it. A 2025 study in PNAS Nexus by Smitha Milli and colleagues ran a controlled experiment with 806 Twitter users, showing each of them their top ten tweets from the engagement algorithm alongside their top ten from a plain chronological feed. The algorithm amplified anger by 0.47 standard deviations and out-group animosity by 0.24. Forty-six percent of algorithmically selected political tweets contained out-group hostility, versus 38 percent of chronological ones.

Then the researchers asked users which tweets they preferred. Users were less likely to prefer the political tweets the algorithm had chosen for them.

That is the finding I keep coming back to. The system is not giving people what they want. It is giving people what they cannot look away from, and those are not the same thing, and the gap between them is where a country goes to wither and die.

Part III: The Study That Cannot Be Run

Here is where this gets complicated, and where I want to be careful, because the standard move at this point is to wave at a few null results and call it balance.

The best-publicized evidence against everything I just wrote comes from a set of randomized experiments. In 2023, a consortium of academic researchers working with Meta published a package of studies in Science and Nature. They swapped algorithmic feeds for chronological ones. They stripped out reshared content. They cut exposure to like-minded sources. In 2024 the same project published a deactivation experiment: roughly 19,900 Facebook users and 15,600 Instagram users paid to shut their accounts off for six weeks running into the 2020 election.

What people saw changed enormously. What people believed did not. No meaningful movement on affective polarization, issue polarization, or candidate evaluations.

Meta has been citing that ever since.

There are three problems with it, and they are not small.

The control group was not a control group.

While that experiment was running, Meta was quietly running one of its own. Around the 2020 election the company deployed what it internally called "break glass" measures: emergency changes to the news feed built to suppress hate speech, election delegitimization keywords, and publishers with low quality scores. There were roughly 63 of them. Around a dozen have ever been publicly documented. They overlapped the study window of September 24 to December 23, 2020, and the published study did not account for them.

A team led by Przemyslaw Grabowicz at UMass Amherst, with Filippo Menczer among the coauthors, went back over the data and estimated that a substantial share of the reduction in untrustworthy news exposure credited to the normal algorithm was actually attributable to those temporary emergency settings.

Let us sit with what that means. The experiment did not compare a chronological feed against Facebook's algorithm. It compared a chronological feed against Facebook's algorithm on its best behavior, during the one stretch in the company's history when executives had their hands on the dials and were actively turning them the other way. Then the result was reported to the public as evidence that the dials do not matter much.

Attitudinal movement is the wrong instrument.

This is the part I want to state plainly, because it is the crux. Asking whether six weeks off Facebook moves a feeling thermometer is not a test of whether fifteen years on Facebook built the thermometer. The data is not useful for painting a realistic picture of the impact social media has had on the masses.

A 2025 paper by Joseph Bak-Coleman, Stephan Lewandowsky, Philipp Lorenz-Spreen, Arvind Narayanan, Amy Orben, and Lisa Oswald sets out why the design cannot answer the question people think it is answering. Four reasons, and every one of them lands here.

Effects are not linear. Society-level outcomes are not the sum of individual-level ones, so you cannot scale from a subject to a country.

Effects are path-dependent. Removing a cause does not run it backward. Their analogy is a smoking cessation study: take the cigarettes away for six weeks and irritability goes up, which on this logic would prove smoking is good for your mood.

Effects unfold on the system's timescale, not the study's.

And the assumption holding the whole design together fails outright. A randomized trial requires that treating one person does not alter another person's experience. Social media is nothing but that. Deactivate a subject and their friends keep posting content the algorithm shaped, cable news keeps chasing what performed online that morning, and the party they vote for has already rewritten its messaging around a ranking function. The control group is soaking in the treatment.

There is no unexposed America to compare against. That is not a limitation of the study. That is the finding.

And the null result is not even consistent.

In 2020, Hunt Allcott, Luca Braghieri, Sarah Eichmeyer, and Matthew Gentzkow paid people to deactivate Facebook for four weeks before the 2018 midterms. Deactivation significantly reduced polarization of views on policy issues. Their overall political polarization index fell by 0.16 standard deviations, which the authors themselves noted was roughly 42 percent as large as the entire increase in American polarization between 1996 and 2016.

Four weeks. Forty-two percent of two decades.

It also reduced factual news knowledge and increased subjective wellbeing, which is its own uncomfortable sentence.

And the authors of the 2024 deactivation study said the quiet part out loud in their own paper: the polarization effects they could not detect likely accumulate over a longer period than they were able to observe.

So no, the evidence does not exonerate the platforms. What it shows is that we have been trying to measure a twenty-year climate with a six-week thermometer, in a room where the operator was adjusting the thermostat and did not mention it.

One finding does survive, and it deserves to be sat with rather than waved off. Boxell, Gentzkow, and Shapiro found in 2017 that affective polarization grew fastest among Americans 65 and older, roughly three times the rate of the 18 to 39 group, and the older group was the one least likely to be getting campaign news online. That is real. It is also a measure of an individual's internet use, not of their exposure to an information environment. The retiree watching cable news is watching segments selected partly by what performed online, forwarded links by their kids, and hearing from a congressman whose messaging was rebuilt around engagement metrics. You do not have to touch the machine to live downstream of it.

One more thing, from the opposite direction, because it cuts against the doom framing and belongs in the record. In June 2026, UC Davis researchers led by Garen Wintemute reported on a nationally representative survey of more than 8,000 adults tracking attitudes toward political violence from mid-2024 to mid-2025. Personal willingness to damage property, threaten someone, or injure someone did not increase. Willingness to kill declined slightly. Their conclusion was that extreme positions and personal readiness to commit violence "remain limited to a small minority," and that the country is not in an inevitable slide toward widespread political violence.

Eighty-five percent of Americans believe violence is rising. Almost none of them are willing to commit any. Hold both of those at once, because the gap between them is the whole subject of this article. The public is having reality shaped for them. People believe the country is becoming more dangerous, yet the national murder rate dropped roughly 18% in 2025 to 4.1 per 100,000 residents, tying 1955 and 1956 for the lowest recorded rate since national estimates began.

Social media did not create the division. It built the most precise map of it that has ever existed, made that map available on a rental basis, and lowered the price of exploiting it to roughly zero. And the most parasitic corporations on the planet began using it to do exactly that.

Part IV: Who Is Renting the Map

In October 2019, the Senate Intelligence Committee released Volume 2 of its bipartisan Russia investigation. The finding was blunt: "The Committee found the IRA targeted African-Americans more than any other group or demographic." The majority of operations "focused on exacerbating existing tensions on socially divisive issues, including race, immigration, and Second Amendment rights."

And the campaign did not stop when the votes were counted. IRA activity spiked after Election Day 2016: up 238 percent on Instagram, 84 percent on YouTube, 59 percent on Facebook, 52 percent on Twitter.

That detail tells you everything about the objective. If the goal were electing a candidate, the operation ends in November. It did not end in November. The candidate was the pretext. The division was the product.

It is still running. On September 4, 2024, the Justice Department seized 32 internet domains used in a Russian operation known as Doppelganger, run by the Social Design Agency, Structura National Technology, and ANO Dialog, directed out of the Russian Presidential Administration under First Deputy Chief of Staff Sergei Kiriyenko. Court filings named internal projects including the "Good Old USA Project" and a "U.S. Social Media Influencers Network Project." The operation used fabricated influencers, fake profiles impersonating American citizens, and, per the Deputy Attorney General, AI-generated narratives and AI-generated advertisements aimed at specific American demographics and regions.

Heading into the 2026 midterms, the tooling is scaling and the defenses are not. A Brennan Center report published July 2, 2026 found that China, Iran, and Russia are all preparing to influence these midterms with AI-generated content, while federal agencies are no longer funding election security improvements, providing cybersecurity support, or sharing threat information. States, per the report, are preparing for this largely on their own for the first time since 2016.

But foreign operations are a rounding error next to the domestic economics, and this is the part Americans consistently refuse to look at.

Meta reported full-year 2025 revenue of $200.97 billion. Of that, $196.18 billion was advertising. Roughly 98 cents of every dollar. The company reported 3.58 billion daily active people across its apps in December 2025.

That is the business. Attention is inventory, engagement is the yield curve, and out-group hostility is the highest-performing input in the entire system. Not because anyone wants it to be. Because the wiring makes it so, and the optimizer found the wiring.

Meanwhile the correction mechanisms are being dismantled. Meta ended third-party fact-checking in January 2025 in favor of crowdsourced Community Notes. In March 2026, Meta's own Oversight Board issued an advisory opinion warning that Community Notes are not an adequate substitute. Their numbers: in the first six months of the U.S. rollout, Meta published roughly 900 Community Notes. Over the same period, professional fact-checkers in the EU labeled approximately 35 million Facebook posts. There is no strike, no reach penalty, and no demonetization for a post that receives a note.

And the ground truth is disappearing underneath all of it. Northwestern's Medill School reported in October 2025 that 213 U.S. counties have no local news source at all, and 1,524 more have only one, usually a weekly. About 50 million Americans live with limited or no access to local news. Since 2005, nearly 3,500 newspapers have closed or merged, close to 40 percent of the country's local papers, and more than 270,000 newspaper jobs are gone.

Take away the shared local fact and the shared local reporter, hand everyone a national feed optimized for out-group hostility, and you have not just divided a country. You have removed the last thing both neighbors were reading.

Part V: Why You Cannot See It Happening to You

This is the layer that matters most, and the one almost nobody writes about, because it implicates the writer and the reader equally. Manipulation that works does not feel like manipulation. It feels like being right.

In 2019, More in Common surveyed Americans on what they believed the other side actually thinks. Democrats estimated 53 percent of Republicans hold extreme views. The real figure was 34 percent. Republicans estimated 56 percent of Democrats hold extreme views. The real figure was 29 percent. Across the board, Americans imagine roughly twice as many extremists on the other side as actually exist.

Two findings in that report that deserve more attention.

First: among Democrats, more education made the gap worse. Democrats with postgraduate degrees were about three times as inaccurate about Republicans as Democrats who had not finished high school. Among Republicans, education showed no significant relationship either way.

Second: the more news you consume, the less you know. Americans who followed the news "most of the time" had a perception gap nearly three times larger than those who followed it "only now and then," 28 points versus 10. People who post about politics on social media were 61 percent further off than people who do not.

Being informed, in the way we currently define informed, makes you even further removed from reality. If you do not acknowledge your own biases and learn to separate fact from opinion, your view becomes incredibly skewed.

The 2018 study by Douglas Ahler and Gaurav Sood put numbers on the caricature. Americans overestimated the size of party-stereotypical groups by an average of 342 percent. Respondents guessed that 38.2 percent of Republicans earn over $250,000 a year. The actual figure is 2.2 percent. They guessed 31.7 percent of Democrats are LGBT. The actual figure is 6.3 percent. Union members among Democrats: guessed 39.3 percent, actually 10.5.

You are not arguing with the other party. You are arguing with a cartoon of the other party that you assembled out of the most engaging examples an optimizer could find, because the most engaging examples are by definition the least representative ones.

And exposure does not fix it. Chris Bail's 2018 PNAS field experiment surveyed 1,652 partisans and paid a subset to follow a bot retweeting the other side for a month. Republicans exposed to liberal content became substantially more conservative, by up to roughly 0.6 standard deviations at high compliance. Democrats moved slightly left, not significantly.

Meeting the enemy did not humanize the enemy. It confirmed the enemy. Because what they encountered was not their neighbor's actual politics. It was the loudest, most badge-like version of it, which is exactly what Tooby predicted a coalition would produce and exactly what an engagement optimizer would surface.

Then layer on the third-person effect. A 2008 meta-analysis covering 106 studies and 372 effect sizes found that people consistently believe media influences others more than it influences them, at an average effect size of d = 0.646. And the effect gets stronger the more harmful the message appears: for content people judge undesirable, propaganda being the textbook case, it rises to d = 0.862.

In other words, the more obviously manipulative something looks, the more confident you become that it is working on everyone but you.

Put it together and you get a population that is certain it is being lied to, certain the lies are working, and certain the victims are somebody other than them.

That is the trap. Not that we are angry. That we experience the anger as evidence.

Part VI: What It Would Actually Take

In October 2024, a team led by Jan Voelkel ran what may be the most useful study on this question yet. They solicited 252 proposals for reducing polarization, selected 25, and tested them on more than 32,000 Americans in a single megastudy.

The good news: 23 of 25 treatments significantly reduced partisan animosity. The best performers were embarrassingly simple. Show people a sympathetic, relatable person who disagrees with them. Remind people that most Americans are an exhausted majority rather than committed combatants. Most Americans are struggling to buy gas, groceries, or a home.

The bad news, in three parts.

Only six of the 25 moved antidemocratic attitudes at all. Reducing dislike is easy; reducing willingness to break the rules to win is a different problem with different causes.

The interventions that did move antidemocratic attitudes worked mostly by correcting misperceptions, telling people the truth about how few of their opponents actually support undemocratic measures. Which means the disease and the treatment are the same substance. What is making us dangerous is a factual error about each other, and the error is manufactured continuously, for free, by systems that profit from it.

And the effects faded. A follow-up with 9,000 participants two weeks later found the depolarizing effects had waned significantly.

So the honest answer to "what would it take" is not a message. Messages wash out in fourteen days while the feed runs every day forever.

It takes changing what gets rewarded. Remember the detail buried in Facebook's own documents: when the company finally set the angry-reaction weight to zero, users saw less misinformation and less graphic violence, and engagement did not drop. The dial exists. It works. It is not being turned, because nobody has made it expensive not to turn it.

It takes rebuilding the shared local fact. Not as nostalgia. As infrastructure. Fifty million Americans with limited or no access to local news are fifty million Americans whose only available picture of their own country is a national one, and the national one is optimized.

It takes cross-cutting identities, which cannot be legislated and can only be lived. Mason's cross-pressured partisans, the ones who could not get fully enraged because their own memberships pulled in different directions, were the shock absorbers of American politics. We sorted them out of existence. Anything that puts people in genuine, repeated, non-optional contact across the sort, which is to say churches, unions, service, sports leagues, town councils, jobs, rebuilds them.

And it may take giving up on the word.

"One true America" was never a real thing. There was no year when we agreed. There were only years when the disagreements were arranged differently, cutting across each other instead of stacking up into two piles. The goal is not unity. Unity is a coalition signal, and we already know what those do.

The goal is a country where you can be wrong about your neighbor's politics without being wrong about your neighbor.

Right now the entire apparatus, the wiring, the optimizer, the ad market, the foreign services renting time on it, and the vacuum where the local paper used to be, is arranged to prevent exactly that.

None of it requires anyone to be evil. That is the part I cannot get past. Every actor in this chain can be behaving rationally inside their own incentive, and the aggregate output is still a nation that cannot tell the truth about itself.

The wiring kept us alive for a hundred thousand years. It is not going anywhere. The only variable is who gets to hold the other end of it. Right now, tech companies and corrupt administrations hold the reigns, because we handed them over. The power still lies with the people, but social media and the manipulation of algorithms made us forget how much we still have in common.

Sources

The Grounded stands behind every factual claim it publishes with a $100 accuracy guarantee. Find an error of fact and tell us. If you are right, we pay.