Stop Blaming the Algorithm
The research blaming your feed mostly didn’t replicate.
In a county hospital in West Texas in the spring of 2025, a child died of measles. The first American to die of that disease since 2015. By July, the national case count had reached 1,333, more than four times the previous year, and 92 percent of those who caught it were unvaccinated or had no record of a shot (Centers for Disease Control and Prevention, 2025). Gaines County, near the center of the outbreak, had kindergarten measles coverage as low as 82 percent (Cousins, 2025). Community immunity needs about 95.
The same season, a national survey of roughly 5,800 voters found that 63 percent of Republicans still believed the 2020 election was stolen, four years after every recount, audit, and court said it wasn’t, including the audit run by a firm that went looking for bamboo fibers in the ballots after someone decided the fraudulent votes were flown in from China (Public Religion Research Institute, 2024).
Two different rooms. Same wallpaper. People holding, with total confidence, a belief that the evidence flattens, and not budging when you show them the evidence.
The story everyone tells about why this happens goes like this: social media broke our brains. The algorithm built a funhouse mirror, fed each of us a private reality, radicalized the lonely, and now your uncle thinks the vaccine has a tracking chip. It’s a tidy story. It has a villain with a logo. It implies a fix: tune the algorithm, add a label, and sanity returns.
I want to tell you that the tidy story is mostly wrong, and that the research it was built on largely fell apart when people tried to reproduce it. The thing actually wrecking the shared world is older than the internet, older than cable news, older than me, and it does not have a logo. It has a name from a 1960 psychology paper, and you have it. I have it. The reason this matters is not academic. It changes what we should even be trying to fix.
The Thing Is Old
In 1960, a British psychologist named Peter Wason sat people down with a number sequence, 2, 4, 6, and asked them to figure out the rule that generated it. They could test guesses by proposing their own triples. He’d tell them yes or no, then ask for the rule (Wason, 1960).
Almost everyone did the same thing. They guessed “even numbers going up by two,” then proposed 8, 10, 12. Yes. Then 20, 22, 24. Yes. Then 100, 102, 104. Yes. Confident now, they announced the rule. Wrong. The actual rule was “any three increasing numbers.” They never found it because they never tried to break their own guess. They only ever asked the question whose answer they already expected.
That’s confirmation bias, and it’s worth being precise about what it is, because the loose version (“people believe what they want”) doesn’t align with the research. Klayman and Ha cleaned it up in 1987: it’s a positive test strategy (Klayman & Ha, 1987). We probe our hunches by looking for cases that fit them. This is a perfectly good strategy when your hunch is roughly right. It fails badly, and invisibly, when the truth is wider than your hunch, since every confirming case feels like proof and tells you nothing.
Sixty-five years of work since then, cataloged in Nickerson’s enormous 1998 review and threaded through Kahneman’s fast-and-slow machinery, points the same direction (Nickerson, 1998; Kahneman, 2011). The mind runs cheap. Agreeable information slides in without a fight. Disagreeable information triggers an audit, and the auditor is not neutral: the defendant hired them.
The cruelest finding in this literature, the one that should bother you if you think education is the exit, comes from Dan Kahan. He gave people data to interpret. On a neutral topic, people who were better at math read the data more accurately. On a politically loaded topic, the people who were better at math were more polarized, not less (Kahan, 2013). They used the horsepower to build a better defense of the conclusion they already wanted. Smart does not save you. Smart buys you a nicer lawyer.
None of this needed Facebook. Wason ran his study on a chalkboard.
The Algorithm Did Less Than You Think
Here is the part that gets me in trouble at parties.
The strongest causal evidence we have on whether the feed itself polarizes people comes from a set of experiments that Meta let outside academics run on real Facebook and Instagram users during the 2020 election, published across Science, Nature, and PNAS (Guess et al., 2023; Nyhan et al., 2023; Allcott et al., 2024). These weren’t surveys. They reassigned what real people saw, for months, and measured what happened.
They switched people from the ranked algorithmic feed to a plain reverse-chronological feed for 3 months. It changed the feed a lot. It changed political attitudes almost not at all. They cut people’s exposure to like-minded sources. No measurable polarization effect. They deactivated people’s accounts entirely for six weeks before the election. Small dip in news knowledge, nothing measurable on polarization or who people voted for.
Brendan Nyhan, one of the lead researchers, put it as plainly as a careful scientist can: nobody is saying social media has no negative effects, but these were among the most-discussed interventions, and none of them measurably moved attitudes (Wagner, 2023).
It goes back further. The famous filter-bubble panic, the Eli Pariser idea that the algorithm seals each of us in a private echo chamber, never had the empirical goods. Bakshy and colleagues looked at 10 million Facebook users back in 2015 and found that the algorithm trimmed cross-cutting content a little, while people’s own clicking trimmed it much more (Bakshy et al., 2015). We are not trapped in the bubble. We are building it by hand, and we like it in here.
YouTube was supposed to be the radicalization pipeline, the auto-play conveyor belt to extremism. Then Hosseinmardi and colleagues looked at actual behavioral data and found that the people watching far-right content on YouTube were already seeking it everywhere else too, and mostly arrived through external links and searches, not the recommender, which “quickly forgets” extreme viewing if you go back to normal stuff (Hosseinmardi et al., 2021). The pipeline mostly carried people who’d already bought the ticket.
I’m not telling you the algorithm is fine. I’ll get to what it does. I’m telling you that the specific, popular, satisfying claim that the recommendation engine reaches into a normal person and manufactures a conspiracist is the part of the story with the weakest evidence under it. We wanted it to be true because it located the disease outside of us, in a server farm in Menlo Park, where it could be regulated. The data has been rude about that wish.
The Backfire Effect Was Mostly Wrong Too
As long as we are demolishing comfortable beliefs, here’s one that lived inside the fact-checking profession itself.
In 2010, Nyhan and Reifler reported that correcting a political falsehood could backfire: conservatives shown a correction about Iraqi weapons believed the false thing more afterward (Nyhan & Reifler, 2010). It was catnip. It explained everything. It launched a decade of communications strategy built on the premise that correcting people is counterproductive, so don’t.
Then people tried to reproduce it. Wood and Porter ran five experiments with more than 10,000 subjects and 52 of the most polarized issues they could find, precisely where backfire should show up if it were real. They found none (Wood & Porter, 2019). The original authors joined a collaborative effort to find it again and couldn’t reliably reproduce their own effect. One of the researchers began calling the search for it the ‘white whale’ (Nurse, 2019). The current read, from the meta-analysis, indicates that backfire is rare, not nonexistent, and not the normal response to correction (Swire-Thompson et al., 2020).
Sit with what that means. For roughly a decade, a chunk of the people whose job was fighting misinformation operated on a finding about misinformation that was, functionally, misinformation itself. The zombie idea that you shouldn’t correct people because it backfires shambled around long after its evidentiary skull was empty. Confirmation bias does not spare those who study it. That should make you humble, and it should make you suspicious of any version of this story, including a satisfying one, including this one, that you didn’t have to work to believe.
Corrections mostly nudge factual beliefs in the right direction. They just don’t move the attitude underneath, and they fade. That’s not a backfire. That’s something more boring and more durable, which is the actual subject.
What’s Actually Doing the Work
So if it isn’t the algorithm reaching in and reprogramming people, what is it?
Three things, locked together, none of them sufficient alone, which is the whole point and the reason the single-villain stories keep failing.
One: motivated reasoning, the old engine. Not “people believe what they want.” More precisely, people believe what they can build a passable argument for, and their desire decides which arguments get built and how hard each one gets checked (Kunda, 1990). The check is real. It’s just rigged. And it runs hardest exactly when the belief has become a uniform, a way of showing whose side you’re on. Whether the election was stolen stopped being a question about ballots a long time ago. It became a question about who you are. You don’t reason your way out of a hat.
Two: the partisan media ecosystem, which is neither new nor symmetric. The American right built an explicitly conspiratorial broadcast layer, OAN and Newsmax, with no real left-side equivalent at the same wattage. The cleanest causal estimate we have, from Martin and Yurukoglu exploiting the near-random spot where Fox landed in different cable lineups, shows that slanted cable doesn’t just preach to the choir; it moves votes at the margin (Martin & Yurukoglu, 2017). And the partisan gap on the stolen-election claim, 63 percent of Republicans to 4 percent of Democrats, is not a both-sides shrug. It’s a 59-point canyon with a donor network and elite voices shoveling on the far side of it (Public Religion Research Institute, 2024).
Three: engagement-based ranking. This is where the algorithm actually earns its indictment, and it’s narrower and nastier than the funhouse-mirror story. The single strongest predictor of whether a political post spreads is not how true it is, not how emotional it is, but whether it attacks the other side. Rathje and colleagues went through 2.7 million posts. Each additional word about the political out-group raised shares by 67 percent on Twitter and 45 percent on Facebook (Rathje et al., 2021). Brady’s team found that moral-emotional language travels the same way, mostly within the in-group, rather than across (Brady et al., 2017).
Read those two together, and you get the actual mechanism. The algorithm is not building you a private reality. The algorithm is running an auction, and the currency is contempt for the people you already dislike. It doesn’t need to manufacture the bias. The bias was there in 1960. It just pays a bounty on the ugliest expression of it, which trains the people producing content to bring more of that, which is fed to a mind that was already, by ancient design, going to wave the agreeable version through without an audit.
No layer does it alone. The cognitive engine without the partisan ecosystem is just ordinary human stubbornness. The ecosystem without the engine has nothing to grip. The ranking without either is an empty auction house. You need all three in the room, and almost every popular explanation, and every viral fix, picks one and ignores the other two. That’s why the fixes keep underperforming.
The Loop That Can’t Lose
You can watch all three layers click together if you look at the worst cases, since those are the ones where the belief has been engineered, accidentally, to never lose.
QAnon started in October 2017 as anonymous posts on an imageboard most people will never see, claiming a cabal of Satan-worshipping pedophiles ran the government and that Trump was secretly at war with them (Public Religion Research Institute, 2024). On its face, this should have stayed in the basement. It didn’t. It walked onto Facebook, YouTube, and Instagram, got recommended, gained followers, and then folded itself into the stolen-election movement so completely that by January 6, 2021, the iconography was at the Capitol. After the deplatforming, the brand fragmented, and the belief didn’t. According to polling, the share of Americans holding the central tenets increased between 2021 and 2023, not decreased.
Here is the engineering. Every prediction QAnon made failed. The “storm” never came. A normal belief takes damage when its predictions fail. This one survived because the framework had a built-in answer: the failure proves the enemy is powerful and the plan is deep. Disconfirmation got recycled as confirmation. The audit Wason’s subjects skipped, this belief structurally cannot perform, since every possible piece of evidence, including the evidence against it, has been pre-assigned the meaning “I’m right.”
The sovereign citizen movement runs the identical trick in a different register. Adherents believe, on idiosyncratic readings of commercial law, that the legal system has no authority over them except when they consent. Every time a court rules against one of them, the ruling is not evidence that the theory is wrong. It’s evidence that the illegitimate system is doing exactly what an illegitimate system does. The FBI has tracked them as a domestic terror threat since 2011, mostly since the loop tends to detonate at traffic stops (Public Religion Research Institute, 2024). You cannot argue someone out of a structure whose every exit has been pre-labeled as the enemy’s trap. The motivated-reasoning engine, handed a self-sealing frame and an ecosystem that rewards the most contemptuous version of it, doesn’t just resist correction; it actively resists it. It converts the correction into fuel.
This is why I keep insisting that no single fix touches it. You cannot fact-check a structure that eats fact-checks. You cannot redesign a feed out from under a belief that the person is now seeking on purpose, across every platform, now that it has become the load-bearing wall of who they understand themselves to be.
My Own Feed
Here is the part a reader on my own side gets to skip, and I let them skip it the first time.
Every example I just walked you through is one where the evidence-backed answer is the one a left-leaning reader already holds. Iraqi WMD. The stolen election. Vaccine denial. QAnon. A progressive can read all of that, nod, and close the tab feeling diagnosed and then told they’re healthy. That’s cotton candy. It also quietly guts the only sentence in this piece I actually care about, the one that says the bug is in you too. So let me spend the discomfort.
In March 2026, CNN’s As Equals team published an investigation into a horrifying thing: networks of men trading instructions for drugging and assaulting their partners, including a Telegram group, “Zzz,” with nearly a thousand members (CNN, 2026). Real reporting. German journalists had already documented a wider web of these groups, some with tens of thousands of members. The harm is not in dispute, and I want that on the record before the next sentence, since the next sentence is the uncomfortable one.
Buried in that same CNN piece was a traffic statistic: a pornography site connected to the story had roughly 62 million visits in a month. Site visits. Across more than a hundred content categories. On April 15, Shannon Watts, who founded Moms Demand Action, posted it to X, writing, “Over 62 million men attended in February alone.” That post drew about two million views. The number was wrong in every load-bearing way. It was visits to a website, not men, not attendance, not a seminar. The Telegram group at the actual center of the reporting had closer to a thousand users. Snopes had it corrected within days (Liles, 2026).
You already know what happened next, since it’s the same thing that happens on the other side. The false version traveled. The correction didn’t. There’s a 2018 study in Science that measured this directly: false stories reach far more people and move faster than true ones, and the gap is not small (Vosoughi et al., 2018). The retraction, when it comes, runs into the continued influence effect, the finding that a correction reduces but never fully removes the original belief (Lewandowsky et al., 2012). Watts’s post was still up weeks later, carrying a reader-added community note and no correction from her. About two million views of the false number. A fact-check that, by the structure of the thing, could never catch it.
I’m not dragging one activist. That’s the point I keep making and keep having to make to myself. The “62 million men” post is an out-group-animosity post. Remember the Rathje et al. (2021) number from earlier in this piece, the one where each word about the political enemy raised shares by sixty-odd percent. That mechanism doesn’t check your politics at the door. It paid a bounty on contempt-for-Republicans content, and it paid the identical bounty on contempt-for-men content, and the people sharing the second kind weren’t lying any more than those sharing the first were. The number confirmed a prior so cleanly that the audit nobody ran on the right is the same audit nobody ran here. Same machine. Same bug. Different jersey.
My editor on the first draft, a reader named Max, also raised the issue of femicide statistics and said they need more care, since it’s not the same kind of error. The UN’s 2024 figures put roughly 83,000 women and girls killed intentionally worldwide, and roughly 50,000 of those killed by a partner or family member (UNODC & UN Women, 2025), which of those is “the femicide number” depends entirely on whom you decide to count: only intimate-partner killings, all gender-motivated killings, or every woman killed for any reason. Roughly four in ten of those killings don’t carry enough information to be classified at all. That’s not a fabricated statistic. It’s a real disagreement over a definition, and both sides reach for the count that flatters the argument they already want to make. The technical name for that is motivated numeracy (Kahan et al., 2017), and I have to flag something to keep myself honest: the famous study behind motivated numeracy failed a careful preregistered replication in 2021 (Persson et al., 2021). If I’m going to spend this whole essay arguing that replication failures should make us hold our favorite findings loosely, I don’t get to exempt the finding that happens to be convenient here. The femicide case is softer than the 62 million case for exactly that reason. It’s contested counting, not a clean false claim. It belongs in the essay anyway, since the softer cases are the ones we’re most fluent at not noticing.
The Fixes Are All Small
I want this part to be honest, because the temptation in a piece like this is to end on a five-point plan that makes you feel better.
Fact-checking works a little. Real effect, small, decays, doesn’t touch the identity underneath (Walter et al., 2020). Prebunking, where you show people the manipulation trick before they encounter it in the wild, is the most promising thing we’ve got, but its effects fade, and they are modest (Roozenbeek et al., 2022). Accuracy nudges, a little prompt asking whether this is true before you share, genuinely shift sharing toward better stuff, and the effect got smaller the harder people looked at it (Pennycook et al., 2021; Roozenbeek et al., 2021). Platform redesign, per the Meta experiments above, did almost nothing in the short run. Media literacy education helps a bit, and carries a quiet risk: it assumes the reader wants to find the truth, when the entire problem is that the motivated reasoner is using exactly those skills to defend the conclusion they walked in with.
Add them up, and you do not get a cure. You get a set of small forces pushing against a very large one. The honest framing is that misinformation is not an outbreak you end. It’s a chronic condition you manage, like the thing in that Texas county that we used to manage and then, collectively, decided to stop believing in.
Why I Actually Care About This
I do health informatics. I build and study the systems that hand information to patients. So this is not a spectator sport for me.
The large language models now being wired into symptom checkers and patient portals, and the little chat box on the hospital website, were trained on the open internet, which is to say they were trained on the full corpus of human motivated reasoning. They can produce the vaccine-skeptical script fluently. They can write it in the calm voice of a doctor. They will hand a worried parent a confident, personalized, wrong answer with the same cognitive ease that lets a false thing slide into your head unaudited at Thanksgiving. The Center for Countering Digital Hate got one major system to generate misleading election images on every single one of 60 test prompts (Center for Countering Digital Hate, 2024).
We are about to install machines that exploit the 1960 bug at an industrial scale, into the one domain where the wrong answer comes with a body count, and a measles ward in West Texas is the early reading on what that costs. That is the part that should keep my field up at night, and it mostly doesn’t yet.
The Landing
Here is what makes me angriest, to borrow my own habit.
We spent a decade and a fortune building a story in which the problem was the machine, since a machine can be subpoenaed, regulated, redesigned, and blamed. The machine does something real, an auction that pays out in contempt, and that part deserves the fight. But the engine the auction is feeding has been running since before the transistor; it runs in numerate people and credentialed people and, demonstrably, in the researchers who study it, and it is running, right now, in you. At the same time, you decide whether this whole argument is correct based partly on whether it flattered a thing you already suspected.
The algorithm is not the disease. The algorithm is a pharmacy that fills prescriptions quickly. The disease is older, it’s ours, and it does not have a logo to sue.
For a deeper dive on this topic, read the spring semester final paper for one of my graduate classes right here.
Author Note
Grace Ann Hansen is an independent researcher and writer, and an MBA & PhD graduate student in health informatics and artificial intelligence. She is also a published author, a professional musician, a gymnastics coach, and a queer transgender woman living in Sioux Falls, South Dakota. She corrects all her papers and articles with Grammarly, because even though she has deep thoughts, she has shallow patience for punctuation. She uses Anthropic’s Claude in Research mode for source location and verification on cited factual claims; all interpretation, argument, and prose are her own. Correspondence concerning this article should be addressed to Grace Ann Hansen at grace@graceannhansen.com.