Your social media feed is not a window onto reality but a mirror polished to reflect your worst impulses. You think you choose what you see, but you don’t. The system always picks whatever keeps you watching, and that’s how it makes money.
Recommendation systems are not content sanitizers. They don’t value truth or your mental wellbeing. They optimize a single metric: engagement. In technical jargon, watch time or session duration. Every action you take is a data point. Every pause, every scroll back up, every comment you write is a reward signal. The algorithm learns which combination of words, images and colors stops you. This is not a political conspiracy. It’s differential calculus applied to human psychology.
The model builds a high-dimensional vector representation of you and of the content. It computes predictions for the probability that you’ll click or watch a video for ten seconds. If it gets it right, it reinforces that digital neural pathway. If it fails, it discards it. The system is blind to moral context. All it understands is the correlation between stimulus and behavioral response.
The problem we have is that hate is profitable, because it’s biologically efficient. One study analyzed more than 500,000 posts by political actors on social media to measure the effect of moral-emotional language [1]. The methodology was rigorous: large-scale natural language processing combined with real interaction metrics. The result was blunt. Each moral-emotional word increased the probability of the post being shared by 20%. Four years later, another team ran the same play at a larger scale: 2.7 million posts by media outlets and members of Congress on Facebook and Twitter [2]. And they found something worse. Each additional word referring to the opposing political group increased shares by around 67%. Trashing the rival was the strongest predictor of virality in the entire study, far above praising your own side. Outrage sells. Contempt for the enemy sells more.
The algorithm doesn’t understand politics; all it cares about is that moral-emotional content captures your attention before you’ve decided whether it interests you. This has been tested in the lab with attentional capture paradigms: morally and emotionally charged words are detected faster and hold the gaze longer than neutral ones, and that attentional advantage predicts which messages go viral [3]. The outrageous stimulus outruns slow scrutiny; by the time your deliberative side shows up, you’ve already clicked. Moral outrage generates more comments, and those comments weigh heavily in the engagement optimization function. The system isn’t politically biased by design. It’s biologically biased by default. Anger and rage are the cheapest, most efficient fuel to keep the machinery spinning.
The popular narrative claims that algorithms create polarization from scratch. The scientific evidence tells a more nuanced and far more uncomfortable story. A randomized controlled trial switched tens of thousands of Facebook and Instagram users from an algorithmic feed to a chronological one for three months, in the middle of the 2020 election campaign [4]. The methodology is the gold standard: random assignment, control group, attitudes measured before and after. The result dismantles the simplification. The chronological feed reduced screen time and changed users’ information diet. Yet the effect on affective polarization was minimal.
Affective polarization is not disagreeing about taxes. It’s seeing the opposing political group as an existential threat to the nation. It’s the difference between “your policies are wrong” and “you are evil.” The algorithm feeds this, yes. But it is not the origin of the disease. It’s the amplifier.
To grasp the real magnitude, you have to look at the deactivation experiments. A monumental study randomized 19,857 Facebook users and 15,585 Instagram users to deactivate their accounts for six weeks before the 2020 United States presidential election [5]. The methodology was brutal in its simplicity and rigor. One group was paid to deactivate their account for one week. The treatment group was paid more to deactivate for six weeks. An instrumental variables analysis was used to correct for imperfect compliance, cross-referencing the data with administrative voting records and passive app tracking.
The findings are counterintuitive. Deactivation reduced online political participation. It slightly reduced belief in misinformation. But the effects on affective polarization, issue polarization, the perceived legitimacy of the election and electoral turnout were precisely estimated and close to zero. Turning off the app doesn’t make you less polarized. Social fractures exist independently of the code. Unfortunately, in the United States, polarization has for decades been growing fastest precisely in the demographic group that uses the internet least, those over 65 [6]. If the algorithm were the root cause, you’d expect exactly the opposite. Which means there are almost certainly other factors at work, such as the media, corruption or current politics.
It’s important to keep in mind that the studies above are a collaboration with Meta, and during the exact period of those studies the platform had 63 emergency measures switched on — they called them “break glass” measures — designed to reduce the visibility of news from untrustworthy sources after the election [7]. In other words: the “normal” algorithm everything was being compared against had already been tweaked to resemble its most harmless version. An independent team pointed out that this changes the control condition of the experiment and may explain part of its results [8]. The original authors responded defending their conclusions, and the dispute remains open. The null result on polarization probably survives; the idea that Facebook’s standard algorithm is harmless when it comes to misinformation is not so clear. When the lab is owned by the subject of the experiment, it pays to read the fine print.
What nobody told you
The evidence shows the authors building a theoretical platform model with a clear premise: the more weight the ranking gives to social signals — likes, shares, comments — the more engagement it generates, and the more misinformation and polarization it produces, because ideologically extreme users interact disproportionately and the system amplifies them. They then test the model against a natural experiment: Facebook’s 2018 algorithm change (“Meaningful Social Interactions”), which gave more weight to precisely those signals. Using survey data from the United States and Italy and a difference-in-differences estimation, they find that the change contributed to an increase in ideological extremism and affective polarization in both countries.
And it’s not the only natural experiment at international scale. Since 2016, Twitter kept 1% of its users — about two million people — on the old chronological feed, as a permanent control group. When its own researchers compared the two groups across seven countries, they found that the algorithmic feed amplified political content in general, and in six of the seven countries it amplified the political right more than the left [10]. Another analysis, this one covering 375 million tweets across nine countries — Canada, France, Germany, Italy, Poland, Spain, Turkey, the United Kingdom and the United States — found the same structure everywhere: polarized political interaction networks where mentions of the opposing group are systematically more toxic than internal ones [11]. With one inconvenient nuance for any simple narrative: those hostile interactions received fewer likes than the friendly ones. Hostility structures global political debate even when it doesn’t win the engagement auction.
Why does this happen? Because the algorithmic feed prioritizes novel, high-arousal content. In today’s media ecosystem, that kind of framing aligns disproportionately with populism and anti-establishment rhetoric. The algorithm is blind to ideology. But the content it rewards is not. The platform’s architecture shapes behavior without any need for explicit political intent from its engineers.
A randomized field experiment on the platform X (formerly Twitter) demonstrates this starkly [12]. An independent academic team — with no cooperation from the platform, which is exactly what makes the data valuable — randomly assigned almost 5,000 active users in the United States to the chronological “Following” feed or the algorithmic “For You” feed for seven weeks in 2023. The result was clear. The algorithmic feed significantly increased engagement and shifted political opinions toward more conservative positions, particularly on public policy priorities, perceptions of the criminal investigations into Donald Trump, and opinions on the war in Ukraine. The content analysis revealed the mechanism: the algorithm promoted conservative and activist posts, and demoted traditional media.
Two details of this study matter more than the headline. First: affective polarization and self-reported partisanship did not move. The algorithm shifted specific opinions; it didn’t turn anyone into a fanatic in seven weeks. Consistent with everything above: amplifier, not origin. Second, and this is the one you should remember: the effect was asymmetric and persistent. Turning the algorithm on moved attitudes; turning it off did not move them back. Users exposed to the algorithmic feed started following conservative activist accounts, and kept following them once the algorithm was gone. The feed switches off; the network of accounts it built for you does not. This also explains, in passing, why the deactivation experiments find nulls: they arrive too late. They don’t detect the algorithm’s effect because the algorithm had already done its work before the experiment began.
This doesn’t happen because the code is written by conservatives. My view is that it happens because conservative content, on average, currently tends to use more emotional exaltation and anti-institutional rhetoric, which the engagement metric automatically rewards. The system is a mercenary. It sides with whoever pays it in attention, whatever flag they fly.
The case of TikTok points in the same mercenary direction. Its “For You” feed is an almost purely algorithmic product; user agency is minimal. An audit using 323 sock-puppet accounts seeded with Democratic or Republican content, which collected more than 280,000 recommendations over the 27 weeks leading up to the 2024 US election, found a systematic, asymmetric skew: Republican-seeded accounts received 11.5% more content from their own party, while Democratic-seeded accounts were exposed to 7.5% more content from the opposing party, much of it anti-Democrat material [13]. Before anyone grabs a banner: the same team had documented a left-leaning skew in YouTube’s political recommendations in 2023. The direction of the bias depends on the platform and the moment. What doesn’t depend on anything is the existence of the bias: algorithmic neutrality is a myth manufactured by communications departments.
Now, if turning off the algorithm doesn’t depolarize, is the ranking then irrelevant to affective polarization? No. An experiment published in Science settled it with an elegant trick: instead of turning anything off, it reordered [14]. A browser extension, armed with a language model, detected in real time the X posts containing partisan hostility and antidemocratic attitudes — calling for prison or violence against the other party’s voters — and moved them down or up in the feed of 1,256 participants for ten days during the 2024 presidential campaign. Preregistered, independent, no permission from the platform. Downranking that content cooled animosity toward the opposing party by about two points on a hundred-point thermometer; upranking it heated it by the same amount. In both camps equally. Not a single post was deleted. Only the order changed. Two points in ten days is roughly what American partisan animosity takes about three years to grow on its own. The synthesis of a decade of contradictory results fits in three lines: removing the algorithm entirely does nothing measurable [4][5]. Turning it on shifts opinions and leaves a persistent mark [12]. And surgically reordering hostility moves affective polarization in a matter of days [14]. The problem was never that a ranking exists, but what comes out on top.
Today’s models generate synthetic, hyper-personalized content, designed in real time to maximize the engagement of one particular user. If the system can create content and a specific reality for each user, generating the perfect argument, with the exact tone and the precise cultural reference, keeping them trapped in their dopamine loop, what happens to the concept of shared reality? The reordering experiment [14] showed that a language model can cool down a feed at will. The same technology, with the sign flipped, can heat it up. We are moving from algorithmic content selection to the generation of made-to-measure reality tunnels. Is it possible to control an algorithm that adapts and evolves faster than any human legislative process? Or will we simply be billions of nodes, each living our own private reality, generated by the algorithm? Unable even to define the terms of a disagreement — or an agreement.
The evidence converges on a single, solid point. Social media exploits human neurobiology to maximize our time in front of the screen. It amplifies polarization because polarization is profitable, not because there’s some grand political conspiracy in a Silicon Valley basement. And its effects don’t undo themselves when you flip the switch, because the algorithm doesn’t just order what you see: it rebuilds who you follow, which topics matter to you and which version of the adversary you know. That part stays.
Deleting your Instagram account won’t heal a divided society. The cracks are there, with or without the app. But deactivating it will give you back several hours of your week in which you can think for yourself, choose your own sources and decide for yourself how much weight events deserve. It will free you from the outrage feedback loop. The real question is not how to fix the algorithm. The question is why you’re still willing to be its raw material.
References
[1] Brady, W. J., Wills, J. A., Jost, J. T., Tucker, J. A. & Van Bavel, J. J. (New York University, 2017). Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences, 114(28).
Label: Reliable. Peer-reviewed and widely cited, though later reanalyses have disputed the size of the effect depending on how the network structure is modeled (see note 1).
[2] Rathje, S., Van Bavel, J. J. & van der Linden, S. (University of Cambridge / New York University, 2021). Out-group animosity drives engagement on social media. Proceedings of the National Academy of Sciences, 118(26).
Label: Reliable. Peer-reviewed, based on 2.7 million real posts, with findings consistent across two platforms and conceptually replicated in later work.
[3] Brady, W. J., Gantman, A. P. & Van Bavel, J. J. (New York University, 2020). Attentional capture helps explain why moral and emotional content go viral. Journal of Experimental Psychology: General, 149(4).
Label: With reservations. A series of lab experiments from a single team; a plausible, measured mechanism, pending broad independent replication.
[4] Guess, A. M. et al. (Princeton University and the US 2020 Facebook & Instagram Election Study consortium, 2023). How do social media feed algorithms affect attitudes and behavior in an election campaign? Science, 381(6656), 398–404.
Label: Large-scale randomized trial and methodological gold standard, but conducted in collaboration with Meta and under active methodological dispute (see notes 2 and 3).
[5] Allcott, H., Gentzkow, M., Mason, W., Wilkins, A., Barberá, P. et al. (Stanford University / New York University and the US 2020 consortium, 2024). The effects of Facebook and Instagram on the 2020 election: A deactivation experiment. Proceedings of the National Academy of Sciences, 121(21), e2321584121.
Label: Reliable. The largest deactivation experiment conducted to date, with random assignment, passive tracking and administrative voting records. Note: it belongs to the same consortium with Meta and its time window overlaps with the platform’s emergency measures (see notes 2 and 3).
[6] Boxell, L., Gentzkow, M. & Shapiro, J. M. (Stanford University / Brown University, 2017). Greater Internet use is not associated with faster growth in political polarization among US demographic groups. Proceedings of the National Academy of Sciences, 114(40).
Label: Reliable. Long demographic series with multiple measures of polarization.
[7] Thorp, H. H. & Vinson, V. (2024). Context matters in social media. Science, 385(6716), 1393.
Label: Editorial (institutional document, no empirical rating).
[8] Bagchi, C., Menczer, F., Lundquist, J., Tarafdar, M., Paik, A. & Grabowicz, P. A. (University of Massachusetts Amherst / Indiana University / University College Dublin, 2024). Social media algorithms can curb misinformation, but do they? Science (technical comment).
Label: With reservations. Methodological critique under active dispute, with a published response from the original authors (see note 3).
[9] Germano, F., Gómez, V. & Sobbrio, F. (Universitat Pompeu Fabra / University of Rome Tor Vergata, 2026). Ranking for engagement: How social media algorithms fuel misinformation and polarization. Journal of Public Economics, 255, 105589.
Label: Recent. Peer-reviewed; theoretical model with empirical validation via difference-in-differences on survey data from the US and Italy. No independent replication yet.
[10] Huszár, F., Ktena, S. I., O’Brien, C., Belli, L., Schlaikjer, A. & Hardt, M. (Twitter / University of Cambridge, 2022). Algorithmic amplification of politics on Twitter. Proceedings of the National Academy of Sciences, 119(1).
Label: With reservations. Solid experimental design with millions of users over several years, but built on internal company data that cannot be audited by third parties.
[11] Falkenberg, M., Zollo, F., Quattrociocchi, W., Pfeffer, J. & Baronchelli, A. (City University of London / Ca’ Foscari University of Venice / Sapienza University of Rome / Technical University of Munich, 2024). Patterns of partisan toxicity and engagement reveal the common structure of online political communication across countries. Nature Communications, 15, 9560.
Label: With reservations. Observational, correlational analysis, though at massive scale (375 million tweets, nine countries).
[12] Gauthier, G., Hodler, R., Widmer, P. & Zhuravskaya, E. (Paris School of Economics / University of St. Gallen, 2026). The political effects of X’s feed algorithm. Nature. DOI: 10.1038/s41586-026-10098-2.
Label: Recent. Randomized field experiment independent of the platform, with almost 5,000 users over seven weeks. Peer-reviewed in a top-impact journal; pending replication in other contexts and time frames.
[13] Ibrahim, H., Jang, H. D., Aldahoul, N., Kaufman, A. R., Rahwan, T. & Zaki, Y. (New York University Abu Dhabi, 2026). Systematic partisan content skews in TikTok during the 2024 US elections. Nature. DOI: 10.1038/s41586-026-10447-1.
Label: Recent. Peer-reviewed algorithmic audit with 323 controlled experiments; limited to a specific electoral window (see note 4).
[14] Piccardi, T., Saveski, M., Jia, C., Hancock, J., Tsai, J. L. & Bernstein, M. S. (Stanford University / Johns Hopkins University / University of Washington / Northeastern University, 2025). Reranking partisan animosity in algorithmic social media feeds alters affective polarization. Science, 390(6776). DOI: 10.1126/science.adu5584.
Label: Recent. Preregistered field experiment independent of the platform; a clean causal effect, measured in the short term and at the peak of an election campaign.