Dr Matt Kneale

The Machine That Ate Democracy

The evidence is now overwhelming: social media’s business model is structurally hostile to democratic health. The question is what, if anything, we intend to do about it.

In March 2018, three researchers at MIT published a paper in Science that should have ended the debate about whether social media is good for democracy. Soroush Vosoughi, Deb Roy and Sinan Aral analysed every verified true and false news story that spread on Twitter from the platform’s founding in 2006 through to 2017. They examined 126,000 stories shared by three million users more than 4.5 million times. The fact-checking was verified by six independent organisations with 95 to 98 per cent agreement on classifications.

Their findings were brutal. False news was 70 per cent more likely to be retweeted than true stories. It reached 1,500 people approximately six times faster. The top one per cent of false news cascades reached between 1,000 and 100,000 people; the truth rarely exceeded 1,000. False political news spread more than three times faster than any other category.

The researchers then removed bot activity from the dataset. The results did not change. Human psychology, not automation, was driving the spread. We like novel information. False news, being more surprising, triggers stronger emotional reactions. People who share novel information are seen as being in the know. The algorithm rewards what the algorithm measures, and what it measures is engagement.

This was not an aberration. It was the architecture working exactly as designed.

The algorithm’s fingerprints

Twitter’s own internal research, published in PNAS in December 2021, confirmed what the MIT study had found. Researchers studied millions of tweets across seven countries and found that in six out of seven nations, the algorithmic timeline amplified right-leaning political content over left-leaning content compared to a chronological baseline. In some cases, amplification exceeded 200 per cent.

Then came Frances Haugen.

In September and October 2021, the former Facebook product manager disclosed internal company documents showing that Facebook understood exactly what its algorithms were doing. The company’s “Meaningful Social Interactions” algorithm weighted angry emoji reactions five times more than likes until 2020. An internal memo acknowledged “unhealthy side effects on important slices of public content, such as politics and news.” A 2016 internal study found that 64 per cent of all extremist group joins were due to recommendation tools.

Most damning: Facebook’s own teams projected that removing the MSI model would reduce misinformation by 30 to 50 per cent. The company declined to implement the changes. Mark Zuckerberg personally rejected proposed modifications in April 2020.

A pre-registered experiment by UC Berkeley researchers, published in PNAS Nexus in 2024, added further confirmation. Of political tweets selected by Twitter’s algorithm, 62 per cent expressed anger (versus 52 per cent in a chronological feed) and 46 per cent contained out-group animosity (versus 38 per cent). Users felt worse about political opponents after viewing algorithmic content. Notably, users did not prefer the political content the algorithm selected for them. They were being served outrage they never asked for.

The correction problem

If false news spreads faster, surely corrections catch up eventually?

They do not. A meta-analysis by Walter and Tukachinsky in 2020, aggregating 32 studies, found that correction does not entirely eliminate the effect of misinformation. A comprehensive review by Ecker and Lewandowsky in Nature Reviews Psychology in 2022 confirmed what researchers call the “continued influence effect”: misinformation continues to shape reasoning even after correction.

Meanwhile, the “illusory truth effect” compounds the problem. A meta-analysis by Dechêne and colleagues in 2010 found a medium effect size for repetition increasing perceived truth. Research by Pennycook, Cannon and Rand in 2018 demonstrated that a single prior exposure to fake news headlines increased perceived accuracy. Repeat a lie often enough and it starts to feel true. Social media repeats lies at industrial scale.

The Twitter revolution that wasn’t

Remember when social media was going to save democracy?

The 2011 Arab Spring generated breathless coverage about Twitter revolutions and Facebook uprisings. Western media treated the protests as validation of Silicon Valley’s emancipatory promises. The narrative has not aged well.

Marc Lynch, a political scientist at George Washington University who studies the region, called the claims “hugely overrated.” Evgeny Morozov, author of The Net Delusion, described the narrative as “an intellectually impoverished, lazy way to study the past.”

The statistics explain why. During the 2011 Egyptian protests, only 7 to 10 per cent of the population used Facebook and approximately one per cent used Twitter, concentrated in Cairo. Twitter had no Arabic interface during the revolution. A study by Wolfsfeld, Segev and Sheafer in the International Journal of Political Communication in 2013 found that a significant increase in social media use was much more likely to follow protest activity than precede it. Politics came first.

Zeynep Tufekci’s analysis in Twitter and Tear Gas offered the most nuanced assessment. Social media created “tactical freeze”: movements could mobilise quickly but lacked infrastructure to adapt when challenged. When Mubarak shut down the internet on 28 January 2011, it actually strengthened the revolution by forcing people into the streets.

Research by Philip Howard and Muzammil Hussain noted that only 10 to 20 per cent of those in the Middle East and North Africa had easy internet access. But here is the crucial finding: countries with the lowest internet penetration, Yemen and Libya, still experienced successful uprisings, while Bahrain, with 88 per cent internet penetration, saw its protests suppressed through Saudi military intervention.

Technology was not determinant. Military decisions were.

Foreign interference, documented

The Mueller Report stated that Russian interference in the 2016 US election occurred “in sweeping and systematic fashion.” The Internet Research Agency conducted a social media campaign favouring Donald Trump while seeking to “provoke and amplify political and social discord.”

The numbers are staggering. 126 million Americans may have been served IRA content on Facebook between 2015 and 2017. IRA accounts posted 8.5 million tweets that were retweeted more than 31 million times. IRA Instagram posts garnered 185 million likes and four million comments. The IRA’s monthly budget reached $1.25 million by September 2016.

The Oxford Computational Propaganda Project identified 20 distinct audience segments targeted by IRA operations. African American politics and culture received the most ads: 841 ads with 13.6 million impressions. The strategy was explicit: campaign for African American voters to boycott elections or follow wrong voting procedures, while encouraging extreme right-wing voters to be more confrontational.

The UK’s response was rather different. The Intelligence and Security Committee Russia Report in 2020 found the government “actively avoided looking for evidence that Russia interfered” in the Brexit referendum, a “stark contrast” with the US response. The committee identified that “insufficient attention has been paid to Russian state-linked infiltration in British politics.” It could not even identify which government department carries operational responsibility for defending democratic processes from disinformation.

The collapse of civic infrastructure

While attention flowed to platforms, civic institutions withered.

Trade union membership in the UK peaked at 13.2 million in 1979. By 2024, it had fallen to 6.4 million, 22 per cent of employees, the lowest rate on record. Private sector density stands at just 11.7 per cent, also a record low. In the United States, union membership reached a record low of 9.9 per cent in 2024, down from 20.1 per cent in 1983.

Political party membership has collapsed. The Conservative Party’s membership fell from approximately three million in 1953 to around 121,000 by 2024, a loss of nearly 200,000 members since 2001 alone. Only approximately one per cent of the UK population now belongs to a political party, the lowest level in Europe.

Local media, the essential infrastructure of democratic accountability, has been hollowed out. At least 293 UK local newspapers have closed since 2005. Frontline journalists in the UK fell from approximately 23,000 in 2007 to 17,000 in 2017-18, a 26 per cent decline. The top three local publishers employ roughly 3,000 journalists today versus 9,000 in 2007, a 67 per cent decline.

In the United States, the Medill State of Local News Report 2025 identifies 213 “news desert” counties where 50 million Americans have limited to no access to local news. Since 2005, more than 3,200 print newspapers have vanished. A George Mason University study found newspaper closure correlates with a 6.9 per cent increase in corruption charges and 6.8 per cent increase in indicted defendants in affected areas.

No local newspaper means no one watching the council meetings. No one watching means no one is caught.

The democratic recession

Freedom House has documented 19 consecutive years of global freedom decline, from 2006 to 2024. In 2024, 60 countries experienced deterioration while only 34 showed improvement. The number of electoral democracies fell to 106 of 195 countries, down from 110 in 2023.

The V-Dem Institute’s Democracy Report 2025 is starker still. The Liberal Democracy Index has declined to 1996 levels by country averages and 1985 levels when weighted by population. The “third wave of autocratisation” has lasted 25 years. 45 countries are now autocratising, up from 12 countries two decades ago. For the first time in over 20 years, autocracies (91) outnumber democracies (88) worldwide. Liberal democracies (29) are now the least common regime type since 1990. 72 per cent of the world’s population lives in autocracies, the highest since 1978.

The Economist Intelligence Unit Democracy Index 2024 recorded its lowest global score since tracking began in 2006. France was downgraded from “full democracy” to “flawed democracy.” Only 25 countries qualify as full democracies, representing just 6.6 per cent of the world’s population.

The public broadcasting exception

Not all democracies have declined equally. The pattern is instructive.

European Broadcasting Union research found strong correlations between robust public service media and democratic health: higher press freedom, higher voter turnout, lower corruption, lower right-wing extremism. In countries where public broadcasters have high market share, citizens are much less likely to think of authoritarian leadership as a good way of governing.

The Nordic countries demonstrate the pattern. All five rank in the top seven globally on the EIU Democracy Index: Norway first with a score of 9.81, Iceland third, Sweden fourth, Finland fifth, Denmark sixth. All have strong, well-funded public broadcasters. Finland maintains the highest media trust globally at 69 per cent according to the Reuters Institute.

The Carnegie Endowment noted that “the American media environment is indeed much more ideologically polarised than that of other Western democracies, perhaps partly due to the absence of a centrist public broadcasting organisation that captures both left- and right-wing voters.”

Trust in freefall

The Edelman Trust Barometer 2025 found media remains the least trusted institution globally at approximately 50 per cent, distrusted in 15 of 28 countries. The UK Trust Index stands at 43, among the lowest of developed nations. UK media trust fell to 32 per cent in 2024. The “mass-class divide” identified in 2012 has widened significantly, with a 30-point trust gap between high-grievance and low-grievance populations.

Pew Research documents that US trust in the federal government has collapsed from 73 per cent in 1958 to 17 per cent in September 2025, near the historic low. Trust in national news organisations fell from 76 per cent in 2016 to 56 per cent in 2025. The partisan divide is stark: Republicans’ trust in national news dropped 30 percentage points since 2016. Adults under 30 now trust social media (52 per cent) nearly as much as national news (56 per cent).

Eurobarometer data shows significant rural-urban and class-based divides. Working class respondents show 39 per cent trust in the EU versus 77 per cent among upper class respondents. The Reuters Institute Digital News Report 2024 found UK news trust at 36 per cent, down from 51 per cent in 2015, with 41 per cent actively avoiding news.

Monopoly confirmed

Google’s market dominance has been confirmed by US courts. On 5 August 2024, District Judge Amit Mehta ruled “Google is a monopolist, and it has acted as one to maintain its monopoly” in violation of the Sherman Act. The 277-page opinion found Google holds 89.2 per cent of the general search market, 94.9 per cent on mobile. In April 2025, Judge Brinkema found Google unlawfully monopolised publisher ad server and ad exchange markets. The Department of Justice compared the dominance to “Goldman Sachs or Citibank owning the New York Stock Exchange.”

Meta generated over $160 billion in advertising revenue in 2024, 98 per cent of total revenue. Google, Amazon and Meta are projected to capture 55 per cent of global advertising spend outside China in 2025. Google is forecast to become the first company to exceed $200 billion in annual ad revenue in 2025.

The business model is now clear: extract attention, monetise surveillance, externalise the costs to democracy.

The leader-to-follower pipeline

Social media has enabled a new form of political mobilisation that bypasses parties entirely.

Donald Trump had 88.9 million Twitter followers when permanently suspended in January 2021, having grown from 2.98 million when announcing his 2015 campaign. The White House declared his tweets “official statements by the President.” Trump used the platform to fire cabinet secretaries, announce policy changes without consulting party leadership, and threaten foreign nations.

Jair Bolsonaro’s 2018 Brazilian campaign operated through 120 million WhatsApp users, 90 per cent of Brazilian internet users. Some 87 per cent of respondents claimed to have received fake news via WhatsApp during the election. Bolsonaro spent only $0.01 per vote due to social media efficiency. The Oxford Internet Institute described it as “a marked departure from previous political campaigns.”

Narendra Modi maintains over 35 million Instagram followers. The BJP’s social media operation includes the “Association of Billion Minds” and spent Rs 7.8 crore ($1.08 million) on over 18,000 Facebook ads in 10 weeks before the 2019 election. 400 million Indians use WhatsApp, the platform’s largest market globally.

Viktor Orbán has combined traditional media capture, with approximately 80 per cent of Hungary’s media market controlled by aligned entities, with aggressive social media spending. Hungary spent more than any other EU nation on Meta political advertising: $4.8 million in 30 days in mid-2024, outspending Germany despite having one-eighth the population. During COVID-19, Orbán announced government decisions via Facebook videos, bypassing independent media entirely.

The pattern is consistent: social media enables leaders to build direct relationships with followers, weakening party discipline, institutional constraints and editorial gatekeeping simultaneously.

The attention economy’s logic

Tim Wu’s The Attention Merchants traced how the advertising industry’s “epic scramble to get inside our heads” created the conditions for manipulation. His subsequent The Curse of Bigness warned that concentrated attention platforms threaten democracy itself. Shoshana Zuboff’s The Age of Surveillance Capitalism described how companies “monitor, archive, analyse, and market as much personal information as they can siphon from users,” trading in what she calls “human futures.”

The structural logic is clear. Advertising-driven optimisation creates incentives for maximum engagement, not user wellbeing or democratic deliberation. As internal Facebook documents showed, the company understood its algorithm amplified misinformation and divisive content but judged changes would involve “material tradeoff with MSI impact.” The engagement metrics were more important than the democratic externalities.

Where this leaves us

The evidence base for concern about Web 2.0 architecture is now substantial. It spans peer-reviewed research, court findings, whistleblower disclosures and official government reports. The structural incentives of the attention economy, engagement optimisation, algorithmic amplification of emotional content, advertising-driven surveillance, have created an information environment measurably hostile to the conditions democratic governance requires.

The platforms did not set out to damage democracy. They set out to maximise engagement and sell advertising. Democracy was simply collateral damage in the pursuit of quarterly earnings.

The question now is whether we are prepared to do anything about it. The Nordic model suggests strong public broadcasting and robust media regulation can provide some insulation. The court rulings confirm the antitrust tools exist. The research makes the diagnosis clear.

What remains absent is the political will to act on what we already know.