ENSI Research · Built on a library of 46 primary documents
The argument, before the list
The public conversation about social media and trust has swung between two lazy positions. The first says the platforms broke democracy: algorithms radicalise, lies outrun the truth and trust collapses as a result. The second, which has gained ground since the large Meta election experiments of 2020, says the panic was overdone: take away the algorithm and nothing much changes, so the problem must lie somewhere else. Both positions treat social media as a cause that either is or is not doing the damage. The evidence in this library supports a third reading, and it is the spine of this report. Social media is an amplifier and a reward system laid over an older dynamic of grievance and blame. It did not create the dynamic. It changed the terms of trade.
Consider what the strongest studies actually measure. When a post from a US news outlet or a member of Congress names the political out-group, each such term raises the odds that the post is shared by 67% (Rathje et al. 2021). Out-group language beats negative emotion by a factor of about 4.8 and moral-emotional language by about 6.7. The single most rewarded thing a political actor can do online is not to be angry in general. It is to point at them. That is blame, in the precise sense that Malle and colleagues give it: a judgement about an agent, aimed at a target and seeking an audience (Malle et al. 2014). The engagement economy does not reward emotion as such. It rewards the attribution of fault to an out-group.
The same reward then trains the people who post: more likes for outrage today means more outrage tomorrow (Brady et al. 2021). And engagement-ranked feeds amplify out-group animosity beyond what users chose to follow, while users rate the political posts so selected as worse than their chronological feed (Milli et al. 2025). The platform selects for content its own users, asked reflectively, say they do not want.
So why do the large field experiments find so little? A three-month switch to chronological feeds for some 44,000 Facebook and Instagram users in 2020 did not measurably change polarisation, knowledge or trust (Guess et al. 2023). Four weeks off Facebook did not significantly reduce affective polarisation (Allcott et al. 2020). Across twelve OECD countries, online news use is unrelated to the trend in polarisation, and US polarisation began rising decades before Facebook (Boxell et al. 2020). These results are real. But they test a narrow question — what happens to one individual for a few weeks when their feed changes while everyone else’s stays the same. They cannot see what the incentive structure does to what politicians, outlets and creators produce for everybody. Guess and colleagues say so themselves.
The reconciliation is therefore not “it is all the algorithm” or “it is nothing”. The grievance, the identity sorting, the elite polarisation and the distrust were already there. What social media did was change the economics of acting on them. It lowered the cost of blaming (no face, no retaliation, a crowd to hide in), enlarged the audience for the blamer (networks of hundreds instead of a dinner table) and made the reward visible and immediate (a counter that goes up). Crockett put it in one line: digital media inflate the triggering stimuli of outrage, reduce some of its costs and amplify many of its personal benefits (Crockett 2017). The consequence for trust follows the in-group and out-group split. Outrage raises the expresser’s standing and perceived trustworthiness inside the group while eroding trust across groups — through dehumanisation, false polarisation and the misperception of how hostile the other side really is (Brady, Crockett & Van Bavel 2020; Finkel et al. 2020).
This has a direct implication for what to fix. If social media were the root cause, the remedy would be to remove it or clean up the content. If it is an amplifier and a reward system, the remedy is to change what is rewarded and how content is ranked — and to address the grievances it amplifies at their source. Content moderation chases the output; incentive and ranking design changes the production function. Field reranking experiments support this (Piccardi et al. 2024).
The stakes are not only American. Social and video networks have overtaken news sites as the main route to online news across 48 markets, and trust in news met there sits well below trust in news overall (Reuters Institute 2026); 70% of people across 28 countries are hesitant or unwilling to trust someone different from them (Edelman 2026). For a mid-sized EU state with a fragmented party system and a news market already migrating to platforms, the question is not whether social media “causes” distrust. It is whether the national information environment pays a dividend to people who blame, and what a state can do about the price.
Summary of the main points
Out-group blame is the engagement currency: each out-group term raises sharing odds by 67%, far more than emotional language (Rathje et al. 2021).
Engagement ranking over-supplies division: out-group animosity +0.24 SD, political anger +0.75 SD, yet lower user value for political picks (−0.18 SD) (Milli et al. 2025); reranking moves out-party warmth by ±2–2.5° in a week (Piccardi et al. 2024).
Outrage is learned through feedback and network norms (Brady et al. 2021).
Platforms lower the cost and raise the reward of outrage (Crockett 2017; Brady, Crockett & Van Bavel 2020).
The blame machine changes who can blame whom — stranger, audience-directed blame without warrant checks (Malle et al. 2014).
Moral contagion is real but modest: +13% shares per moral-emotional word, not +20% (Brady et al. 2025).
Human attention amplifies too: falsehood is 70% more likely to be retweeted (Vosoughi et al. 2018); each negative headline word lifts clicks 2.3% (Robertson et al. 2023).
Ranking amplifies tone, not extremes, and favours the mainstream right in 6 of 7 countries (Huszár et al. 2022).
The end state is sectarianism: US out-party hate now exceeds in-party love (Finkel et al. 2020).
Trust is going insular and platform-mediated: news trust 37%, on social media 22%, in AI answers 20% (Reuters Institute 2026).
What the evidence does not show: that platforms caused the rise in polarisation, that short feed swaps change attitudes, that cross-cutting exposure heals, or that misinformation is the core problem.
Verdict: an amplifier and reward system for an existing grievance-and-blame dynamic. Fix incentives and ranking, not only content.
How this is organised
The ten mechanisms below are ranked by three criteria taken together. The first is the strength of the evidence — pre-registered, large-sample and causal designs rank above correlational ones and above frameworks. The second is proximity to trust and blame — how directly the mechanism bears on whom people blame and whom they trust. The third is leverage — whether the mechanism sits in something that can be designed or regulated, such as a ranking objective or a reward signal, rather than in human nature. Each mechanism is set out in three parts: what the evidence shows (with exact figures and samples), stakes for trust, and what to watch. A full section then sets out what the evidence does not show, before the verdict.
1. Out-group animosity is the strongest engagement predictor
What the evidence shows. Rathje, Van Bavel and van der Linden analysed 2,730,215 posts from US news media and members of Congress on Facebook and Twitter (Rathje et al. 2021). The media datasets held 599,999 Facebook posts and 227,229 tweets; the Congress datasets 1,078,562 tweets and 825,424 Facebook posts. The headline is stark: “Each individual term referring to the political out-group increased the odds of a social media post being shared by 67%.” Posts about the out-group were shared about twice as often as posts about the in-group.
An internal meta-analysis across eight datasets puts the comparison in one line · out-group words: exp(b) = 1.67 · in-group words: 1.05, not significant · negative affect: 1.14 · moral-emotional words: 1.10 · positive affect: 0.95. The out-group effect is about 4.8 times the size of negative affect and 6.7 times that of moral-emotional language. Posts containing only out-group terms, compared with posts containing only in-group terms, were associated with a 132% increase in sharing.
Three further findings sharpen the point. First, elites amplify more: the out-group effect was exp(b) = 1.99 in the Congress datasets against 1.41 for media, and per-word effects across Congress datasets ranged from 65% to 180%. Second, the effect was not moderated by political orientation or platform. Third, out-group language predicted “angry” reactions (meta exp(b) = 2.19) and “haha” reactions — mockery — while in-group language predicted “love” (1.57). Anger was the most popular Facebook reaction in seven of the eight datasets. The authors draw the uncomfortable conclusion that people say they do not want leaders to express partisan animus, yet this is the content that receives the most engagement. They also note Facebook’s internal finding, reported in the press, that posts rated “bad for the world” drew more engagement and that down-ranking them cut engagement — and that the company “ultimately chose not to approve the feature”.
Stakes for trust. This is the single most important finding in the library for the question of blame. The strongest engagement signal is not how emotional a post is but who it blames. In-group warmth earns little; out-group derogation earns a great deal. A political actor who wants reach is paid to name an enemy. Since political elites capture the largest premium, the incentive acts on exactly the people whose rhetoric shapes followers’ norms. Rathje and colleagues add a point that runs against much of the policy debate: “the severity of online echo chambers appears to be a less important issue than the kind of content that tends to surface at the top of one’s feed”. The problem is less who people are connected to than what wins the competition for their attention.
What to watch. The share of political and party communication that names an out-group, and its engagement premium over in-group or policy content · whether elite accounts’ out-group premium widens in election periods · whether platforms publish the engagement performance of content their own raters judge harmful, as Facebook’s internal research did · non-US replications, since all of this evidence is American and from before 2020.
2. Engagement ranking amplifies divisive content users say they do not want
What the evidence shows. Milli and colleagues ran a pre-registered audit of 806 US Twitter users in February 2023 (Milli et al. 2025). They compared the first ten tweets in each user’s engagement-ranked timeline with the ten most recent tweets from accounts the user followed — holding constant who the user had chosen to follow — and surveyed users about each tweet. Across 26 outcomes, every result significant at p < 0.05 survived a false discovery rate of 0.01.
Relative to the chronological baseline, the algorithm amplified partisanship (+0.24 SD) and out-group animosity (+0.24 SD). Readers felt worse about their out-group (−0.17 SD) and better about their in-group (+0.08 SD). Author anger rose by 0.47 SD overall; in political tweets “anger was by far the predominant emotion amplified”, with author anger up 0.75 SD and reader anger up 0.37 SD. The decisive result concerns satisfaction. Overall, users rated the engagement-ranked tweets very slightly higher on wanting to see them (+0.06 SD). But for political content, the algorithm’s selections delivered significantly lower user value than the chronological feed (−0.18 SD, p = 0.005). The authors’ reading is that engagement ranking serves users’ revealed preferences — what they linger on and react to — but not their stated preferences.
Two field and lab experiments move this from audit to cause. In a pre-registered ten-day experiment on X during the 2024 US campaign, Piccardi and colleagues used a browser extension to rerank the real feeds of 1,256 participants (Piccardi et al. 2024). Cutting posts that expressed anti-democratic attitudes and partisan animosity from about 10% of the feed to about 1% warmed feelings toward the out-party by +2.11° on a 100-point thermometer. Raising their share from about 10.6% to 13.4% cooled them by −2.48°. The effects did not differ significantly by party, and the authors equate them to roughly three and 3.6 years of the US polarisation trend. In-feed, reduced exposure lowered anger by about 5 points and increased exposure raised it by a similar amount. In a pre-registered lab study of 1,380 US partisans, Jia and colleagues found that both downranking and a plain chronological feed reduced partisan animosity relative to an engagement-ranked feed (d ≈ −0.20 to −0.25), and concluded that “status quo engagement-based ranking may pose the greatest risks to democratic attitudes” (Jia et al. 2023).
Stakes for trust. This mechanism converts the abstract complaint about “the algorithm” into a specific, testable charge: the engagement objective systematically over-supplies hostile content relative to what users endorse on reflection. That is a trust problem twice over. It cools feelings toward the other side, and it erodes trust in the platform itself as a faithful agent of the user. A system that people experience as feeding them what they would rather not see invites the cynical “they are manipulating us” reading that spreads to institutions generally. The Piccardi result also shows that reranking does not require removing speech. Hostile posts were down-weighted, not deleted, and 74.2% of participants noticed no impact on their experience.
There is a business tension: reduced exposure cut time on X by 4.71 minutes a day (Piccardi et al. 2024), and bridging-oriented ranking is hardest to sustain under competition for attention (Ovadya & Thorburn 2023). And ranking by stated preference alone reduced hostility mainly by showing fewer out-group tweets — trading outrage for an echo chamber — whereas a tie-breaker that down-weights out-group animosity avoided that cost (Milli et al. 2025).
What to watch. The gap between revealed-preference (engagement) and stated-preference (survey) metrics in platform reporting · whether ranking objectives include an explicit down-weight on out-group animosity · independent, platform-free reranking experiments of the Piccardi type outside the US · the engagement cost platforms report when testing such changes, since that number decides whether they survive without regulation.
3. Outrage is rewarded — and learned
What the evidence shows. Brady, McLoughlin, Doan and Crockett combined two pre-registered observational studies of Twitter with two pre-registered experiments (Brady et al. 2021). The observational studies covered 7,331 users and 12.7 million tweets: 3,669 politically engaged users tweeting around the Kavanaugh hearings (6.1 million tweets) and 3,669 less political users tweeting around the United Airlines passenger-removal incident (6.6 million tweets). Outrage was measured with a classifier trained on 26,000 tweets that matched expert raters and was distinct from generic negative sentiment. The definition matters for this report: outrage is anger, disgust or contempt in response to a perceived moral violation, associated with blaming people, events or things, holding them responsible, or wanting to punish them. Outrage, as measured here, is blame expressed in public.
The core result is a reinforcement-learning signature. Outrage-specific feedback (likes and retweets) on one day predicted more outrage the next day (b = 0.03 in Study 1 and 0.02 in Study 2, both p < .001). In plain terms, a 100% increase in feedback yields an expected 2–3% increase in outrage expression the following day. The effect weakened with longer tweeting histories and responded to positive prediction errors, as learning models predict, and it was specific to outrage rather than generic negativity.
The second result is norm learning. Users embedded in more ideologically extreme networks expressed more outrage, controlling for their own ideology (b = 0.13 and 0.31). In the simulated-platform experiments (N = 120 each), participants exposed to an outrage norm chose outrage posts on the first trial with odds ratios of 4.94 and 5.38. Norms and reinforcement interacted: in more extreme networks, feedback mattered less, as if the norm had already been internalised. The authors spell out the long-run implication: norm and reinforcement learning “might encourage more moderate users to become less moderate over time”. Because expression can become decoupled from felt emotion, this “could result in a form of pluralistic ignorance … whereby people falsely believe their peers are more outraged than they actually are”.
The limits are real and the authors state them. The large studies are observational; network composition may be self-selected; the experiments used only liberal participants; and expression is not the same as feeling.
Stakes for trust. This is where the reward system becomes a training system. Platforms do not merely surface outrage that already exists; their feedback loops shape what users and elites learn to produce. Brady and colleagues put the governance point bluntly: newsfeed algorithms “can influence users’ moral behaviors by exploiting their natural tendencies for reinforcement learning”, and “crucial inputs to the learning process are shaped by corporate interests”. The 2–3% effect looks trivial for an ordinary user and is not trivial for a political leader with millions of followers, whose daily feedback can double many times over. The trust damage comes through misperception: if visible outrage outruns felt outrage, each side sees the other as angrier and more hostile than it is, and calibrates its own distrust accordingly.
What to watch. Visible reward counters on political content and whether platforms test hiding them · trends in outrage expression among high-follower political accounts relative to their felt or stated positions · meta-perception surveys measuring how hostile each side believes the other to be versus how hostile it is · whether the norm-cascade effect replicates outside the US and outside Twitter.
4. Platforms rewrite the economics of outrage
What the evidence shows. Crockett’s 2017 comment is a framework, not a large study, but it supplies the economic logic that the empirical work later tests (Crockett 2017). Outrage runs from stimuli to responses to outcomes, and digital media change all three. On the stimulus side, offline life is a poor supplier of moral violations: in the everyday-morality dataset Crockett reanalysed (1,252 North American adults signalled five times a day for three days), fewer than 5% of reported daily experiences involved directly witnessing or experiencing immoral acts. Online is different. Participants learned about immoral acts online more than in person or via traditional media, and acts encountered online evoked more outrage than those met in person or through print, TV or radio. Because platforms compete for attention, their algorithms promote what is most likely to be shared, “regardless of whether it benefits those who share it — or is even true”, and selection forces may favour “supernormal” stimuli.
On the cost side, expressing outrage online is cheap. Echo chambers reduce the risk of retaliation; people “hide in a crowd”; targets appear as “two-dimensional icons”. As Crockett puts it, “It’s a lot easier to shame an avatar than someone whose face you can see.” On the benefit side, outrage signals moral quality, and online it “instantly advertises your character to your entire social network and beyond”. Feedback arrives on an unpredictable reinforcement schedule, the classic recipe for habit, so that “a habitual online shamer might express outrage without actually feeling outraged”.
The MAD model of Brady, Crockett and Van Bavel generalises this into three channels — Motivation, Attention and Design (Brady, Crockett & Van Bavel 2020). Motivation: simply logging on can act as a group-identity manipulation, and moral emotions serve group-image goals — outrage and contempt derogate the out-group, while shame is “linked to blaming others, distancing oneself from the negative event, and denial”. Attention: moral and emotional words capture attention within about 250 milliseconds, and words with higher attentional capture predicted more retweets. Design: algorithms act as an “external attentional filter”; quantified likes and shares produce reinforcement; and fast response options favour quick emotional moral judgements. Scale amplifies all of it: online networks of 350–500 contacts on Facebook and about 700 on Twitter, against about 150 offline; and people are seven times more likely to share content when they perceive others are sharing it.
Crockett names the upside — online outrage lets “traditionally disempowered groups … check the behaviour of more powerful interests” — and the costs: messages “ricochet” inside echo chambers, the line between “the truly heinous” and “the merely disagreeable” blurs, and cheap expression substitutes for costlier action.
Stakes for trust. The MAD model identifies a mechanism that is easy to miss: outrage builds trust inside the group even as it destroys trust across groups. People who express outrage by punishing unfairness “are viewed as more trustworthy by others”, and moral-emotional displays raise status “as long as those reactions are well aligned with the value system of their social network” (Brady, Crockett & Van Bavel 2020, citing Jordan et al. 2016). That is the private return on public blame. The cost lands elsewhere — on intergroup trust, through dehumanisation and through the false polarisation that makes differences “appear worse online than they are in reality”. The substitution effect matters too: if online outrage replaces costly civic action, the accountability function that justifies outrage decays while its divisive function grows.
What to watch. The ratio of online outrage to offline costly action (donations, volunteering, complaints filed through formal channels) on salient issues · design features that lower the cost of expression — quote-posting, pile-on mechanics, anonymous replies — and whether friction experiments raise that cost · evidence on the proposed but largely untested interventions in the MAD model: hidden like counts, batched notifications, empathy prompts and superordinate identity cues.
5. The blame machine: who can blame whom, at what cost, with what reward
What the evidence shows. The library’s theory of blame comes from outside the social-media literature, and that is what makes it useful here. Malle, Guglielmo and Monroe define blame as a moral judgement that is “both cognitive and social, regulates social behavior, fundamentally relies on social cognition, and requires warrant” (Malle et al. 2014). Warrant is the key. One can say “it’s just wrong, I can’t tell you why”, but it is socially unacceptable to say “he deserves blame, but I can’t tell you why”. Because blame must be justified, “the community can fact-check the blamer’s warrant”. Ordinary blame, on this account, is largely evidence-tracking: in studies with notable outcome effects, information about mental states explained six times more variance in blame than outcomes did.
The machinery fails in a specific setting. Second-person blame — said to the offender’s face — aims at repair and works best with “low emotional intensity and high thoughtfulness”. Third-person blame, voiced to an audience without the offender present, “sometimes represents a first step toward socially excluding the offender”. Between strangers, “there is little pressure to communicate, persuade, repair, and find common ground”; moral criticism “becomes akin to road rage, an episode of Jerry Springer, or hateful anonymous comments on the internet”. The decisive line is that people “are far less likely to be called on for the lack of warrant in their judgments”. Malle and colleagues add that values can be “preset” by group membership — “the rival always intentionally harms us” — which is the formal hook for scapegoat-style blame.
Read against the platform evidence, these pieces assemble into a single mechanism. Social media converts a large share of blame from second-person to third-person, from known to stranger, and from private to performed. Three parameters change at once · Cost of blaming: falls — no face, low retaliation risk, a crowd to hide in (Crockett 2017) · Audience for the blamer: grows by an order of magnitude, from a handful of listeners to networks of hundreds or thousands (Brady, Crockett & Van Bavel 2020) · Reward: becomes visible, quantified and immediate, and is largest when the target is the out-group (Rathje et al. 2021; Brady et al. 2021). Meanwhile the brake — the demand for warrant — weakens, because the audience is self-selected for agreement and nobody asks the blamer to justify the charge.
This also changes who can blame whom. Before platforms, broadcasting blame to a mass audience required a newspaper, a broadcaster or a party apparatus, each with editorial and legal warrant checks. Now any account can address a mass audience, and any public figure, institution or group can be the target. That democratisation has a genuine upside — the disempowered can name the powerful (Crockett 2017) — but it also opens a channel for those Petersen and colleagues describe as chaos-seekers. Across eight US surveys (combined N = 10,921), a one-SD rise in “need for chaos” raised intentions to share hostile political rumours by about 10 percentage points; the motives were to mobilise against disliked groups and to amuse friends, while sharing because the story is true was negatively related (Petersen et al. 2023). The authors’ warning is that a few individuals with extraordinary negative sentiments “may still have a large reach”.
Stakes for trust. Blame that is cheap, mass-audienced and rewarded without warrant changes the trust equilibrium. Institutions face a constant stream of accusations whose cost to the accuser is close to zero and whose reach can exceed the institution’s own. Out-groups are cast as intentional agents of harm, because preset intentionality is the quickest path to a satisfying blame judgement. And blame shifts from a tool of repair to a tool of exclusion. Finkel and colleagues identify the narrative structure that results: “The other side cheats, so our side would be foolish to adhere to long-standing democratic norms” (Finkel et al. 2020).
What to watch. The share of political content that attributes generalised blame to a group — Ovadya and Thorburn list “attributes blame to a group using plural ‘you’ / ‘they’ pronouns” as an operational content label (Ovadya & Thorburn 2023) · whether platforms build “warrant” features — context, sourcing, notes rated helpful across divides — into blame-heavy content · the reach of high-grievance minorities relative to their size · whether institutions under attack respond with second-person, repair-oriented communication or with counter-blame.
6. Moral contagion is real — but smaller than first thought
What the evidence shows. The claim that moral-emotional language spreads further became one of the founding results of the field. The original 2017 study, as summarised in the MAD model, found that each moral-emotional word added to a tweet was associated with a roughly 20% increase in sharing, and Donald Trump gained an expected 27% more retweets per moral-emotional word against 2% for Hillary Clinton (Brady, Crockett & Van Bavel 2020). Brady, Rathje, Globig and Van Bavel then did what too few fields do: they tested their own result (Brady et al. 2025). A pre-registered direct replication used 849,266 tweets on five new US topics, and a meta-analysis pooled 27 studies from five labs with a combined N of 4,821,006.
The meta-analytic estimate is an incidence rate ratio of 1.13 (95% CI 1.06–1.20): each additional moral-emotional word is associated with 13% more shares, not 20%. The replication varied by topic · Kavanaugh: 1.05 · border-wall shutdown: 1.07 · transgender military ban: 1.66 · COVID-19 masks: 1.12 · George Floyd: 0.97, a reversal. The combination matters: distinctly emotional words alone had a mean IRR of 1.00 and distinctly moral words 1.03; it is moral and emotional language together that predicts sharing. The robustness checks are reassuring — no detectable publication bias (Egger’s test P = 0.171) and evidential value on the p-curve — but the measurement caveat is serious: dictionary and LLM measures of moral-emotional content correlate at only r = 0.27. The authors are explicit that the analysis “cannot establish a causal link”.
Stakes for trust. This mechanism is ranked here for a reason: it calibrates the others. Moral contagion is real, robust and modest, and its sign depends on context. It is dwarfed by the out-group effect — Rathje’s per-word out-group premium of 67% against a moral-emotional premium of about 10% in the same data — which suggests that part of what looked like contagion of moral emotion is contagion of blame aimed at the other side. The George Floyd reversal, and the observation that positive emotions spread most after the 2015 same-sex marriage ruling (Brady, Crockett & Van Bavel 2020), show that online moral emotion is not inherently corrosive. Outrage spreads most where derogating the out-group is already the norm. That matters for policy: it argues against treating emotional or moral language as the target, and for targeting the out-group-hostile, blame-laden subset.
What to watch. Whether future work separates out-group reference from moral-emotional language in the same models · contexts in which moral emotion is shared positively, and what distinguishes them · measurement convergence between dictionary, LLM and human coding, since the size of the problem currently depends on the instrument used to measure it.
7. Human attention is an amplifier in its own right: falsehood and negativity travel further
What the evidence shows. Not all amplification is algorithmic. Vosoughi, Roy and Aral studied every fact-checked rumour cascade on Twitter from 2006 to 2017 — about 126,000 stories tweeted by about 3 million people more than 4.5 million times, with veracity judged by six fact-checking organisations that agreed 95–98% of the time (Vosoughi et al. 2018). Falsehood “diffused significantly farther, faster, deeper, and more broadly than the truth in all categories of information”, and most of all for false political news. The truth rarely reached more than 1,000 people; the top 1% of false cascades routinely reached between 1,000 and 100,000. It took the truth about six times as long to reach 1,500 people. Controlling for account features, falsehoods were 70% more likely to be retweeted. False spreaders had fewer followers and were less active and less often verified: falsehood travelled further “despite these differences, not because of them”. Bots accelerated true and false news at the same rate, so the difference came from humans. False rumours were more novel, and replies to them expressed more surprise and disgust — though the authors caution that they “cannot claim that novelty causes retweets”.
Robertson and colleagues provide the cleanest causal evidence on negativity (Robertson et al. 2023). Using the Upworthy archive of 22,743 randomised headline tests — about 105,000 headline variants, 5.7 million clicks and more than 370 million impressions — they found that for a headline of average length, each negative word raised click-through by 2.3% and each positive word lowered it by 1.0%. The effect was strongest for government and economy stories. Anger words, surprisingly, had no significant effect on clicks; sadness raised them. The authors’ explanation is important for this report: clicking is “a private behaviour”, while sharing is “public, curated and influenced by a myriad of social factors (for example, signalling group identity…)”.
Stakes for trust. Two different selection pressures operate. Privately, people click on negativity. Publicly, they share what signals their group — and anger and out-group hostility are that currency (Rathje et al. 2021; Milli et al. 2025). Both pressures pre-date algorithms; ranking systems learn them and scale them. The trust consequence is a skewed information diet: an environment where alarming, negative and novel claims about public life outcompete accurate and calm ones, especially on government and the economy. That diet need not change minds directly to change trust. It changes the background sense of what public life is like.
What to watch. The negativity premium in headlines on government and economy stories in national media markets · the share of fact-checked claims that reach large cascades before correction · whether behavioural interventions on sharing (labels, prompts, friction) reduce the human differential, as Vosoughi and colleagues recommend over a narrow focus on bots.
8. Algorithmic amplification favours tone, not extremes — and is asymmetric
What the evidence shows. The strongest platform-wide test of political amplification comes from Twitter’s own long-running holdback experiment, in which 1% of accounts were randomly kept on a reverse-chronological feed from 2016 onward (Huszár et al. 2022). Comparing the reach of 3,634 legislators’ tweets in seven countries between April and August 2020, the authors found that all parties were amplified by personalised ranking, in some cases by more than 200%. In six of seven countries, “the mainstream political right enjoys higher algorithmic amplification than the mainstream political left”; Germany was the exception. The largest gaps were in Canada (Liberals 43% vs Conservatives 167%) and the UK (Labour 112% vs Conservatives 176%). Two findings cut against popular belief. Far-left and far-right parties — VOX, Die Linke, AfD, LFI, RN — were generally amplified less than moderate parties. And at the level of individual politicians, amplification ranged from below 0% to 400% within parties, with no significant association between party and amplification. For US news sources, partisan outlets were amplified more than centrist ones under one media-bias rating scheme, with results sensitive to the labels used.
The design has limits: network spill-overs mean the estimates are not unbiased causal quantities, the treatment algorithm changed over time, and the authors were Twitter employees or consultants. The mechanism is unknown; the authors suggest differing communication strategies.
Stakes for trust. Read alongside Milli and colleagues, Huszár’s results resolve an apparent paradox. If ranking amplified extremism, the far parties would be boosted most. They are not. What is amplified is a tone — partisan framing, anger, out-group hostility — that mainstream actors can adopt as easily as fringe ones. Ovadya and Thorburn report that political messaging staff describe “needing to use inflammatory language in order to be competitive” (Ovadya & Thorburn 2023). Asymmetric amplification also matters for trust directly. When one side’s elites gain more reach from the same system, the other side’s perception that the system is rigged against it has some basis, and that perception corrodes trust in platforms and, by extension, in the information environment.
What to watch. Independent audits of party-level amplification in European election periods, including in mid-sized EU states · whether the amplification premium tracks communication style (out-group references, anger) rather than ideology · publication of holdback-experiment results by platforms, since without randomised holdbacks amplification cannot be measured at all.
9. Sectarianism: when out-party hate exceeds in-party love
What the evidence shows. Finkel and colleagues gave the end state a name: political sectarianism, “the tendency to adopt a moralized identification with one political group and against another” (Finkel et al. 2020). It has three ingredients · othering: seeing opposing partisans as essentially different or alien · aversion: disliking and distrusting them · moralisation: seeing them as iniquitous. When the three converge, “political losses can feel like existential threats”. In US national election survey data, warmth toward the in-party held steady at 70–75 degrees, while warmth toward the out-party fell from 48 degrees in the 1970s to 20 today. Out-party hate now exceeds in-party love. Boxell and colleagues measure the same trend: the US in-party minus out-party gap rose from 27.4 points in 1978 to 56.3 in 2020, an increase of 1.08 standard deviations (Boxell et al. 2020).
Misperception is central. Republicans estimate that 32% of Democrats are LGBT when the figure is 6%; Democrats estimate that 38% of Republicans earn more than $250,000 a year when the figure is 2% (Finkel et al. 2020). Finkel and colleagues attribute sectarianism to three main causes — identity sorting into a “mega-identity”, partisan media, and elite polarisation — with social media “intensifying” it through popularity-based algorithms and moralised content that is “especially likely to be retweeted within rather than between partisan networks”. The consequences reach institutions: the correlation between district and presidential vote rose from 0.54 in the 1970s to 0.94 by the 2010s, so that majority-party candidates in most House districts could win despite openly violating democratic principles. The COVID-19 pandemic, a shared threat that might have united the country, did the opposite: 77% of Americans believed the nation had grown more divided, 2.8 SD above the mean of 13 other nations.
Stakes for trust. Sectarianism is where the blame dynamic and the trust dynamic fuse. Its narrative is a blame narrative — the other side cheats — and it licenses norm violation in response, which then supplies fresh evidence that the other side cheats. Social media sits inside that loop as an intensifier, not as its origin. Finkel and colleagues are explicit that the trend predates the platforms. Their remedies are telling: correct misperceptions of out-party hostility; build ranking signals from crowdsourced trust and accuracy rather than raw popularity; and beware that “simply tweaking algorithms to show partisans more content from the opposition may aggravate sectarianism”.
What to watch. Out-party thermometer trends and meta-perception gaps in European panels · whether shared shocks unify or divide, as a threshold indicator of sectarianism · tolerance of norm violations by one’s own side · ranking experiments that target misperception rather than exposure.
10. Insular trust and the platform turn: news moves to feeds, creators and AI answers
What the evidence shows. The 2026 survey data show where trust is going. The Reuters Institute’s Digital News Report covers 97,520 respondents in 48 markets (Reuters Institute 2026). For the first time, social media and video networks (54%) are the most used way of accessing online news, ahead of news sites and apps (51%). Among 18–24-year-olds, 52% name social, video or AI as their main source. AI chatbots are used for news by 10%, up from 7%, and by 16% of under-35s. Trust in news fell to 37%, the lowest since measurement began in 2015, declining in 29 of 48 markets. Trust varies sharply by channel · news overall: 37% · search results: 32% · news on social media: 22% · AI chatbot answers: 20%. The report draws the arithmetic conclusion: “trust overall is prone to go down as more people’s consumption tends towards using social media and video networks (and AI chatbots)”. Part of the fall is therefore a composition effect. News creators are rated more entertaining and relatable but “less trustworthy and less impartial”; 27% get some news from news-focused creators. Concern about what is real and fake online rose to 62%, yet “convenience appears to trump concern”. And a highly engaged minority that prefers like-minded news shares more, comments more and pays more, creating an incentive “rewarding the production of partisan news”.
Edelman’s 2026 barometer of 33,938 respondents in 28 countries describes the psychological counterpart (Edelman 2026). Seventy percent are hesitant or unwilling to trust someone different from them; only 39% get information weekly from sources with a different political leaning, down 6 points; and among those whose trust shifted after a societal event, net trust in national government leaders fell by 16 points while trust in neighbours, family and friends rose by 11. Insularity tracks grievance: 61% of the insular report moderate or high grievance, against 45% of the open. The OECD’s 2026 trust survey adds the institutional anchor: trust in news media stands at 40% and in political parties at 26%, while 47% use social media for information (OECD 2026).
Stakes for trust. Trust is retreating to the proximate and the personal — to people who look like us and to creators who speak like a friend — while the institutions that used to adjudicate shared facts lose ground. The platforms are now the main gateway, and trust in what arrives through them is the lowest of any channel. AI answers, the newest gateway, start from a similarly low base. This is not simply a failure of journalism: the Reuters data show trust in individual brands falling much less than trust in “news” generically. It is a change in the architecture through which people meet information, and in who is paid for supplying it.
What to watch. The share of under-35s whose main news source is social, video or AI · trust in AI-generated news answers as AI gateways scale · the size and engagement premium of the partisan-preferring minority · cross-cutting exposure (Edelman’s 39%) as a leading indicator of insularity · the conditional-trust gap — whether people would trust an institution led by someone different from them, where the insular score 28–30 points lower than the open (Edelman 2026).
What the evidence does NOT show
The mechanisms above are real, but the strongest causal tests in the library set hard limits on how far they can be pushed.
1. It does not show that social media caused the long rise in polarisation. Boxell, Gentzkow and Shapiro assembled 149 surveys from 12 OECD countries over roughly four decades (Boxell et al. 2020). US affective polarisation rose steadily from 1978 — long before social media. Over the same period it fell in six countries (Japan, Australia, Britain, Norway, Sweden and Germany), all of which also went online. Across the 12 countries, the trend in the share of people getting news online was negatively and insignificantly correlated with the trend in polarisation (rank correlation −0.255, p = 0.450). The strongest correlate was elite polarisation (0.782, p = 0.011), followed by the non-white population share (0.609, p = 0.052). Twelve data points cannot settle causation, but a technology that spread everywhere cannot by itself explain a trend that diverged by country.
2. It does not show that switching off the algorithm changes attitudes in the short run. The 2020 Facebook and Instagram election study is the largest test (Guess et al. 2023). Moving 23,391 Facebook and 21,373 Instagram users to a reverse-chronological feed from late September to late December 2020 changed what they saw substantially. On Facebook, content from untrustworthy sources rose from 2.6% to 4.4%, uncivil content fell from 3.2% to 1.8%, and content from moderate or mixed sources rose from 22.6% to 30.9%. Yet the chronological feed “did not significantly alter levels of issue polarization, affective polarization, political knowledge, or other key attitudes”, and had no distinguishable effects on trust in traditional or social media, confidence in political institutions, or belief in the legitimacy of the election. The study was powered to detect affective-polarisation effects as small as d = 0.032. The authors write that their findings “rule out even modest effects, tempering expectations that social media feed-ranking algorithms directly cause affective or issue polarization in individuals”. Companion studies from the same programme — on removing reshared content and on reducing exposure to like-minded sources — are cited in the library but not held in it; Jia and colleagues and Piccardi and colleagues describe them as finding similarly little attitude change (Jia et al. 2023; Piccardi et al. 2024).
Two qualifications keep this from being the last word. First, Meta’s ranking in 2020 was not simply “worse” than chronological order: it suppressed untrustworthy and uncivil content, which complicates any blanket “engagement algorithm equals misinformation” story. Second, the design cannot see general-equilibrium effects. Users in the control group still lived in an ecosystem shaped by the algorithm, and politicians, outlets and creators continued to produce for it. Guess and colleagues point to “the incentives created by the advertising model of social media” as one of the things their experiment could not test.
3. It does not show that leaving Facebook transforms politics or well-being. Allcott, Braghieri, Eichmeyer and Gentzkow paid users to deactivate Facebook for four weeks around the 2018 US midterms (impact sample 1,661; compliance above 90%) (Allcott et al. 2020). Deactivation freed about 60 minutes a day, cut news knowledge by 0.19 SD, and lowered an overall polarisation index by 0.16 SD — driven mainly by reduced exposure to congenial news and by issue polarisation. It “did not statistically significantly reduce affective polarization” (point estimate −0.06 SD). Subjective well-being rose by 0.09 SD — about a third of what cross-sectional correlations would have suggested. The authors’ verdict is balanced: the effects are “large enough to be real concerns, but also smaller in many cases than what one might have expected”. Social media made people better informed and somewhat more polarised at the same time.
4. It does not show that exposure to the other side heals division. Bail and colleagues paid 1,652 frequent Twitter users to follow a bot that retweeted 24 messages a day for a month from opposing-ideology elites and opinion leaders (Bail et al. 2018). Republicans became more conservative (intention-to-treat 0.12, p = 0.008; an increase of between 0.11 and 0.59 SD depending on compliance). Democrats moved slightly and insignificantly in the liberal direction. The authors found “no evidence that exposing Twitter users to opposing views reduces political polarization”. The exposure came from high-profile elites more polarised than the electorate, and the mechanism was not identified, but the warning has been absorbed across the library. Ovadya and Thorburn, Schirch, and Finkel and colleagues all warn that indiscriminate exposure diversity can make things worse (Ovadya & Thorburn 2023; Schirch 2025; Finkel et al. 2020). Bridging is not the same as cross-cutting exposure. Who delivers the opposing view, and in what tone, decides whether it bridges or inflames.
5. It does not show that misinformation is the main problem. Altay, Berriche and Acerbi argue that “alarmist narratives about online misinformation continue to gain traction despite evidence that its prevalence and impact are overstated” (Altay et al. 2023). Misinformation makes up about 0.15% of the American media diet and 0.16% of the French; 61% of French participants consulted no unreliable sources at all over 30 days. The Vosoughi result itself, they note, studied contested, fact-checked claims and left aside the large volume of uncontested news that is extremely viral. Their causal reading reverses the usual arrow: conspiracy belief and vaccine refusal may share a common cause, low trust in institutions, and misbeliefs are often post-hoc rationalisations of pre-existing distrust. Hence their prescription: “enhancing trust in reliable sources should be a priority over fostering distrust in unreliable sources”, and “people are more likely to be uninformed than misinformed”. They also warn that alarmism can itself fuel mistrust of the media. This is consistent with the Reuters finding that trust in news on social media is far lower than trust in news overall: the deficit is in trust in the reliable, not only in belief in the false.
6. It does not show a uniform effect of digital media on democracy. The largest synthesis in the library, a pre-registered systematic review of 496 articles by Lorenz-Spreen, Oswald, Lewandowsky and Hertwig, finds a mixed picture (Lorenz-Spreen et al. 2023). Digital media use is associated with more political participation and information consumption — often in autocracies and emerging democracies — and with declining political trust, rising populism and growing polarisation, more strongly in established democracies. Trust is the clearest negative: associations with trust in institutions “were overwhelmingly pointing in directions detrimental to a functioning democracy”. But only 24 of the 496 articles reported causal evidence, the literature is heavily American, and context reverses meaning — in authoritarian regimes, reduced trust in the political system can be the start of democratic change. Northern European media systems with high journalistic standards appeared to blunt the far-right effect of right-leaning digital media.
7. It does not measure trust directly as often as it seems. Most of the outrage and amplification studies measure sharing, engagement, emotion or thermometer warmth, not trust. The link to declining trust is usually inferred through polarisation, dehumanisation and misperceived norms. Almost all of the causal evidence is American, election-period and short — four weeks to three months. The honest summary is that the mechanisms are well established, the individual short-run attitude effects are small, and the long-run, society-wide effects remain the largest gap in the evidence.
The verdict: an amplifier and a reward system, not the sole cause
Put the two halves of the evidence together and a coherent picture emerges. The grievance, the sorting, the elite polarisation and the distrust were already present and in some countries rising before platforms existed (Boxell et al. 2020; Finkel et al. 2020). Individual exposure to a different feed for a few weeks barely moves deep attitudes (Guess et al. 2023; Allcott et al. 2020). And yet the platforms are not neutral. They pay a measurable dividend for out-group blame (Rathje et al. 2021), they train users and elites to produce more of it (Brady et al. 2021), their ranking objectives over-supply it relative to what users say they want (Milli et al. 2025), and changing the share of hostile content in real feeds does change how warmly partisans feel about one another (Piccardi et al. 2024).
Social media is an amplifier and a reward system for an underlying grievance-and-blame dynamic, rather than its sole cause. Its distinctive contribution is to the economics of blame. It changes who can blame whom — any account can now address a mass audience, and any person, group or institution can be a target. It changes at what cost — the face, the retaliation risk and the demand for warrant all fall away. And it changes with what reward — the payoff is public, quantified, immediate and largest when the target is the out-group. Under those terms, the rational strategy for anyone seeking reach is to blame, and the people most responsive to the incentive are those with the largest audiences.
What this implies: fix the incentives and the ranking, not just the content
If the problem were the content, moderation would be the answer. Because the problem is the reward structure, moderation chases output that the system keeps paying to produce. The evidence points to six implications.
Change the ranking objective. Move political ranking from raw engagement toward stated preferences, with an explicit down-weight on out-group animosity — the tie-breaker that cut hostile content without creating echo chambers (Milli et al. 2025). The reranking experiments show this can be done without deleting speech (Piccardi et al. 2024; Jia et al. 2023).
Reward bridging, not reach. Diverse-approval signals work in practice: Community Notes selected by a bridging algorithm reduced agreement with misleading tweets by about 26% and cut likes and retweets by 25–34%, with no significant difference in effect between Democrats and Republicans (Wojcik et al. 2022).
Restore warrant to public blame. Context, sourcing and cross-validated notes reintroduce the justification demand that stranger-to-stranger blame has lost (Malle et al. 2014).
Avoid visible paternalism. Content warnings backfired with conservatives and sharply raised perceived threats to free speech (d ≈ 0.62–0.77), while silent reranking did not (Jia et al. 2023). Interventions that look like imposed authority recreate the distrust they aim to cure.
Make amplification measurable. Randomised holdbacks, published experiment results and researcher access are the preconditions for any accountability, a point the governance blueprint in the library makes explicit (Schirch 2025; Huszár et al. 2022).
Address the grievance at its source. The people who share hostile rumours to “burn it all down” are driven by personal status loss and isolation, and fact-checks and nudges will not reach them (Petersen et al. 2023). The largest trust gradient in the OECD data is political voice: 69% of those who feel they have a say trust their government, against 22% of those who do not (OECD 2026).
For a mid-sized EU state, the practical first move is not a national platform or a content regime. It is a standing capability to measure its own information environment — the out-group share and engagement premium of political communication, the amplification of national parties, cross-cutting exposure and trust by channel — and to use EU-level instruments to require the holdback data that make such measurement possible. The dividend that social media pays for blame is a design choice. Design choices can be priced, audited and changed.
Sources from the library cited in this report
Allcott, Braghieri, Eichmeyer & Gentzkow 2020 — The Welfare Effects of Social Media — 07 - Social Media, Polarisation and Declining Trust
Altay, Berriche & Acerbi 2023 — Misinformation on Misinformation: Conceptual and Methodological Challenges — 08 - Misinformation and Algorithmic Amplification
Bail et al. 2018 — Exposure to Opposing Views on Social Media Can Increase Political Polarization — 07 - Social Media, Polarisation and Declining Trust
Boxell, Gentzkow & Shapiro 2020 — Cross-Country Trends in Affective Polarization — 07 - Social Media, Polarisation and Declining Trust
Brady, Crockett & Van Bavel 2020 — The MAD Model of Moral Contagion — 06 - Moral Outrage and Emotional Contagion Online
Brady, McLoughlin, Doan & Crockett 2021 — How Social Learning Amplifies Moral Outrage Expression in Online Social Networks — 06 - Moral Outrage and Emotional Contagion Online
Brady, Rathje, Globig & Van Bavel 2025 — Estimating the Effect Size of Moral Contagion in Online Networks — 06 - Moral Outrage and Emotional Contagion Online
Crockett 2017 — Moral Outrage in the Digital Age — 06 - Moral Outrage and Emotional Contagion Online
Edelman Trust Institute 2026 — 2026 Edelman Trust Barometer: Trust Amid Insularity — 03 - Measuring Trust
Finkel et al. 2020 — Political Sectarianism in America — 07 - Social Media, Polarisation and Declining Trust
Guess et al. 2023 — How Do Social Media Feed Algorithms Affect Attitudes and Behavior in an Election Campaign? — 07 - Social Media, Polarisation and Declining Trust
Huszár et al. 2022 — Algorithmic Amplification of Politics on Twitter — 08 - Misinformation and Algorithmic Amplification
Jia, Lam, Mai, Hancock & Bernstein 2023 — Embedding Democratic Values into Social Media AIs via Societal Objective Functions — 10 - Prosocial Platform Design and Bridging Systems
Lorenz-Spreen, Oswald, Lewandowsky & Hertwig 2023 — A Systematic Review of Worldwide Causal and Correlational Evidence on Digital Media and Democracy — 07 - Social Media, Polarisation and Declining Trust
Malle, Guglielmo & Monroe 2014 — A Theory of Blame — 04 - Blame, Scapegoating and Victimhood
Milli et al. 2025 — Engagement, User Satisfaction, and the Amplification of Divisive Content on Social Media — 08 - Misinformation and Algorithmic Amplification
OECD 2026 — Drivers of Trust in Public Institutions 2026 Results — 03 - Measuring Trust
Ovadya & Thorburn 2023 — Bridging Systems: Open Problems for Countering Destructive Divisiveness — 10 - Prosocial Platform Design and Bridging Systems
Petersen, Osmundsen & Arceneaux 2023 — The “Need for Chaos” and Motivations to Share Hostile Political Rumors — 05 - Complexity, Uncertainty and Conspiracy
Piccardi et al. 2024 — Reranking Partisan Animosity in Algorithmic Social Media Feeds Alters Affective Polarization — 10 - Prosocial Platform Design and Bridging Systems
Rathje, Van Bavel & van der Linden 2021 — Out-group Animosity Drives Engagement on Social Media — 06 - Moral Outrage and Emotional Contagion Online
Reuters Institute 2026 — Digital News Report 2026 — 03 - Measuring Trust
Robertson et al. 2023 — Negativity Drives Online News Consumption — 08 - Misinformation and Algorithmic Amplification
Schirch 2025 — Blueprint on Prosocial Tech Design Governance — 10 - Prosocial Platform Design and Bridging Systems
Vosoughi, Roy & Aral 2018 — The Spread of True and False News Online — 08 - Misinformation and Algorithmic Amplification
Wojcik et al. 2022 — Birdwatch: Crowd Wisdom and Bridging Algorithms Can Inform Understanding and Reduce the Spread of Misinformation — 10 - Prosocial Platform Design and Bridging Systems