The Misinformation Machine — How It Actually Works

The World Economic Forum's Global Risks Report 2025 ranked misinformation and disinformation as the number one global risk over a two-year horizon — for the second consecutive year ✓ Established Fact [2]. Over 900 experts from 136 countries placed it ahead of extreme weather events, armed conflict, and societal polarisation. Over a ten-year horizon, it remains in the top five. This is not a problem that experts expect to resolve itself.

The numbers describing public exposure are staggering. Across 25 nations surveyed by Pew Research Center, a median of 72 per cent say the spread of false information online is a major threat to their country [11]. An estimated 86 per cent of global citizens have been directly exposed to misinformation ◈ Strong Evidence. Researchers estimate that approximately 40 per cent of content shared on social media is false — a figure that makes the platform experience closer to a coin flip than a reliable information channel [11].

The economic cost is not hypothetical. A joint study by cybersecurity firm CHEQ and the University of Baltimore estimated that online misinformation costs the global economy $78 billion annually ◈ Strong Evidence [14]. This includes stock market disruptions caused by false financial reporting, public health costs from medical misinformation, and reputational damage to businesses targeted by fabricated narratives. The figure is almost certainly an undercount — it was calculated before the generative AI era dramatically reduced the cost of producing convincing falsehoods.

A UNESCO/Ipsos survey across 16 countries — encompassing nations with 2.5 billion voters facing elections in 2024 — found that 85 per cent of people are apprehensive about the repercussions of online disinformation ✓ Established Fact [5]. Meanwhile, 97 per cent of respondents in a separate global survey said they believe misinformation is harmful to society [11]. The public is not unaware of the problem. It simply has no structural means of solving it.

What makes the current moment different from historical precedents — propaganda has existed since antiquity — is the combination of speed, scale, and economic incentivisation. A false claim can now reach millions of people within minutes, at zero marginal cost, through systems explicitly designed to maximise the spread of content that triggers emotional reactions. The printing press took decades to reshape information dynamics. Social media accomplished it in under ten years.

The MIT study's most significant finding was not about bots or algorithms — it was about people. False news spread primarily through human sharing behaviour, driven by the novelty and emotional arousal that false content generates [1]. Falsehood diffused significantly farther, faster, deeper, and more broadly than truth in every category of information — but the effects were most pronounced for false political news. The misinformation machine runs on human psychology. Technology merely removes the friction.

The UK Parliament's Science and Technology Committee stated it plainly: the business models of social media platforms «incentivise the spread of harmful speech (misinformation, disinformation, polarising content, hate speech)» ✓ Established Fact [13]. This is not an allegation from activists or academics. It is the official finding of a parliamentary inquiry into the structural relationship between advertising revenue and information quality.

The mechanism is straightforward. Platforms sell advertising based on user attention. The more time users spend on a platform, the more advertising they see. Content that generates strong emotional reactions — outrage, indignation, fear, surprise — holds attention longer and spreads further than content that is merely accurate. A study on engagement ranking published in the Journal of Public Economics found that social media algorithms designed to maximise engagement systematically amplify sensational and divisive content [13]. The platform does not need to intend misinformation — the economic incentive produces it automatically.

Frances Haugen, the former Facebook product manager who disclosed tens of thousands of internal documents in 2021, told the US Congress that the company «consistently chose to maximize its growth rather than implement safeguards» [3]. The Wall Street Journal's «Facebook Files» investigation, based on these documents, revealed that the company's own researchers had identified the harms and proposed mitigations — which were rejected because they would reduce engagement metrics [3].

This is not unique to Meta. The structural incentive applies to every advertising-funded platform. Content that is calm, nuanced, and accurate generates less engagement than content that is sensational, divisive, or false. The platform's revenue model creates an invisible subsidy for misinformation — false content is not merely tolerated; it is functionally rewarded by the same system that generates the platform's revenue.

The advertising-driven model also creates a structural dependency for the organisations theoretically positioned to counter misinformation. Fact-checking organisations, news outlets, and investigative journalists all compete for the same advertising revenue that flows overwhelmingly to platforms. The International Fact-Checking Network reported that approximately 60 per cent of fact-checking organisations participated in Meta's third-party fact-checking programme [8]. On average, these organisations received 45 per cent of their revenue from Meta [8]. The entities tasked with checking falsehoods were financially dependent on the company whose algorithms spread them.

Then, in January 2025, Meta ended the programme entirely in the United States ✓ Established Fact [9]. CEO Mark Zuckerberg announced the company would replace third-party fact-checking with a Community Notes model similar to X's, effective April 7, 2025. The company simultaneously eliminated content policies around immigration, gender, and other politically sensitive topics, and reversed changes that had reduced political content in user feeds [9].

The economic logic is unambiguous. Misinformation is not an externality of the attention economy — it is a product of it. The same systems that generate hundreds of billions of dollars in annual advertising revenue also generate the conditions under which falsehood systematically outperforms truth. Reforming content moderation without addressing the underlying business model is the equivalent of treating symptoms while prescribing the disease.

The relationship between algorithmic amplification and misinformation is now extensively documented. A systematic review published in Frontiers in Communication in 2025 synthesised evidence on social media's influence on news judgment, audience development, and the amplification of polarisation and misinformation [13]. The review found that algorithmic curation does not merely reflect user preferences — it actively shapes them, creating echo chambers in which users are predominantly exposed to viewpoints that confirm their existing beliefs.

Research published in Information Systems Research demonstrated that recommendation algorithms on platforms like X systematically skew exposure toward high-popularity accounts, with both left- and right-leaning users seeing amplified exposure to content aligned with their political stance while exposure to opposing viewpoints is reduced [12]. The algorithm does not discriminate between true and false information; it discriminates between engaging and non-engaging content. Since false content is consistently more engaging — the MIT study confirmed this across every category [1] — the algorithm becomes an accelerant for falsehood.

The mechanism operates through what researchers describe as a «human-algorithm interaction» loop. Humans have documented cognitive biases toward novel, emotionally arousing, and morally charged content. Algorithms detect and reinforce these biases by promoting content that generates the highest engagement signals — likes, shares, comments, time-on-page. The content that maximises these signals is disproportionately sensational, divisive, or false [1]. The algorithm then surfaces more of the same type of content to the same user, creating a reinforcing cycle that progressively distorts their information environment.

The echo chamber effect is particularly concerning for political information. Research on X/Twitter during the 2024 US presidential election found that the platform's algorithm amplified exposure to politically aligned content while reducing exposure to opposing viewpoints [12]. Users on both sides of the political spectrum were shown a distorted version of political reality — not through deliberate censorship, but through optimisation for engagement. The result is that users with different political inclinations increasingly inhabit different informational realities, each reinforced by algorithmic curation.

The emergence of generative AI has added a new dimension to algorithmic amplification. AI-enabled tools — particularly large language models and engagement-optimisation algorithms — are now capable of producing misinformation at industrial scale with minimal human input [15]. While deepfakes attracted the most public attention, the more significant AI-driven threat is the ability to generate vast quantities of plausible-sounding text, flood comment sections with synthetic opinions, and create coordinated inauthentic behaviour at a fraction of previous costs.

Platform responses to algorithmic amplification have been limited and often performative. Instagram and Facebook introduced «reduce» labels that deprioritised flagged content — but the threshold for flagging required either user reports or fact-checker intervention, both of which lag significantly behind the speed of viral content. YouTube adjusted its recommendation algorithm in 2019 to reduce «borderline content» — but a 2024 audit found that inflammatory and misleading content continued to appear prominently in recommendations for politically active users [13].

The fundamental challenge is that algorithmic amplification of misinformation is not a bug — it is a feature of engagement-maximisation design. Fixing the algorithm to deprioritise false content would require platforms to sacrifice engagement metrics and, by extension, advertising revenue. The structural incentive runs directly counter to the public interest in accurate information. Until that incentive changes, algorithmic amplification will continue to function as the primary accelerant of the misinformation machine.

The Medill/Northwestern State of Local News 2025 report documents a crisis that has moved beyond decline into structural collapse. More than 130 newspapers closed in the past year alone [4]. The number of complete news desert counties — counties with no local news source whatsoever — rose to 208, up from 204 in 2023 ✓ Established Fact [4]. But the true scale is worse: more than half of the nation's 3,143 counties have little to no local news coverage. Nearly 55 million Americans live in communities with limited or no access to local journalism [4].

The journalist-to-population ratio tells the story with brutal clarity. The United States had approximately 40 journalists per 100,000 residents less than 25 years ago. That number has fallen to 8.2 — a decline of nearly 80 per cent ✓ Established Fact [4]. In 39 states, fewer than 1,000 journalists remain. More than one thousand US counties — one in three — lack the equivalent of even a single full-time local journalist. These are not minor gaps. They are structural absences in the information infrastructure that democratic self-governance requires.

The financial driver is clear. US newspaper advertising revenue peaked at $49.4 billion in 2005. By 2024, it had fallen to approximately $10.5 billion — a 79 per cent decline ✓ Established Fact [4]. That revenue did not disappear — it migrated to digital platforms. The same advertising ecosystem that funds algorithmic amplification of misinformation extracted the financial foundation from the institutions designed to produce accurate local reporting.

The geographic pattern of news deserts is not random. Rural communities, smaller cities, and economically disadvantaged areas are disproportionately affected. While more than 300 local news start-ups have launched in the past five years — 80 per cent of which are digital-only — the vast majority are concentrated in metropolitan areas [4]. The communities most vulnerable to misinformation are precisely those least served by the emerging digital news ecosystem.

Research has consistently linked news deserts to increased civic disengagement, lower voter turnout, higher municipal borrowing costs, and greater vulnerability to corruption. When no one is watching, accountability evaporates. When accurate local information is unavailable, residents turn to national partisan media and social media — channels that prioritise engagement over local relevance and accuracy.

The collapse of local journalism and the rise of misinformation are not parallel trends — they are causally linked. The advertising revenue that once funded newsrooms now funds the platforms that amplify falsehood. The information vacuums created by newsroom closures are filled by unverified content distributed through engagement-maximising algorithms. The misinformation machine did not merely replace local journalism — it consumed the economic foundation that made local journalism possible.

The Romanian case is the starkest illustration to date of misinformation's capacity to subvert democratic processes. The election of November 24, 2024, saw an unexpected surge in support for ultra-nationalist candidate Călin Georgescu, who had polled in single digits but finished first in the initial round [10]. Subsequent investigation revealed a coordinated campaign involving approximately 25,000 TikTok accounts and an estimated €1 million in covert funding, allegedly linked to Russian state interests [10].

Romanian authorities documented more than 85,000 cyber intrusion attempts targeting election infrastructure before and during the first round — a scale and sophistication that pointed to state-sponsored actors [10]. The Constitutional Court initially confirmed the results, then reversed its decision on December 6 after President Iohannis declassified national security intelligence detailing the foreign influence operation. The European Commission subsequently opened a formal investigation into TikTok's role in the affair [6].

The year 2024 was the largest election year in human history, with more than 2.5 billion voters across dozens of countries going to the polls [5]. In India — where the WEF ranked the risk of misinformation highest — nearly 80 per cent of first-time voters were bombarded with fake news on social media [2]. Concerns centred on AI-generated deepfakes of political opponents and fabricated audio recordings. In Brazil, the Electoral Court ordered the removal of thousands of pieces of false content and temporarily suspended messaging applications that failed to control misinformation [2].

However, the anticipated «deepfake apocalypse» of 2024 did not materialise as feared. An analysis by Recorded Future found that less than 1 per cent of all fact-checked election misinformation was AI-generated ◈ Strong Evidence [15]. «Cheap fakes» — crudely edited images, out-of-context videos, and misleading screenshots — were used seven times more frequently than sophisticated AI deepfakes [15]. The misinformation threat remains primarily a human production problem, not an AI production problem — at least for now.

Established democracies with robust institutional ecosystems — independent judiciaries, professional media, active civil society — proved more resilient to disinformation campaigns than newer or more fragile democracies [2]. The EU's legislative framework, particularly the Digital Services Act and Digital Markets Act, forced platforms to conduct systemic risk assessments and provided regulatory tools that were deployed during the Romanian crisis [6]. Countries without such frameworks were significantly more vulnerable.

The weaponisation of information is not limited to foreign interference. Domestic political actors in the United States, the United Kingdom, India, Brazil, and Turkey have all deployed misinformation as a deliberate electoral strategy [15]. Turkey's President Erdoğan used a deepfake to link an opposition leader to terrorist groups. Political parties in Brazil and Pakistan deployed fabricated audio for negative campaigning. The distinction between «foreign interference» and «domestic misinformation» is increasingly blurred — and the platforms that host both profit regardless of the source.

The 2024 election cycle demonstrated that the misinformation machine is not a future threat — it is an operational present-day weapon. The question is no longer whether information warfare can influence elections but whether existing institutional defences are adequate to withstand it. Romania's answer was annulment — a democratic safeguard of last resort that implicitly acknowledged the failure of all preceding defences.

The Digital Services Act represents the most comprehensive regulatory framework yet developed for platform accountability. Since entering full enforcement in February 2024, it has mandated risk assessments for systemic platforms, required transparency in algorithmic decision-making, and established the legal basis for significant financial penalties [6]. The €120 million fine against X in December 2025 — for deceptive design practices, advertising transparency violations, and failure to provide researcher data access — marked the first time a major platform faced a nine-figure penalty specifically for DSA non-compliance [6].

A significant development came in February 2025, when the European Commission and the European Board for Digital Services endorsed the integration of the Code of Practice on Disinformation into the DSA's co-regulatory framework [6]. From July 2025, the previously voluntary code — signed by Google, Meta, Microsoft, and TikTok among others — acquired formal legal standing. The first full reporting cycle, covering July to December 2025, included submissions from platforms, fact-checkers, researchers, and civil society bodies [6].

The Romania crisis provided the DSA's first major operational test. After the annulment of the presidential election in December 2024, the European Commission triggered the DSA's transparency and content-moderation requirements against TikTok, leading to a formal investigation [6]. In a separate case, a Berlin court ruled in favour of German civil rights organisations and ordered X to provide data access to researchers for monitoring misinformation ahead of Germany's elections [6]. These enforcement actions demonstrated that the DSA provides meaningful legal tools — but they also highlighted the reactive nature of the framework: action occurs after harm, not before it.

The Community Notes model — now adopted by both X and Meta — represents a fundamentally different approach to content verification: crowdsourced, decentralised, and ostensibly neutral. Research on X's implementation has produced mixed results. Posts receiving visible Community Notes saw reposts decline by 46 per cent and likes drop by 44 per cent ◈ Strong Evidence [12]. Authors were 32 per cent more likely to delete their posts when a public note was attached [12].

But the limitations are severe. On average, it takes 15 hours for a Community Note to be published ◈ Strong Evidence [12]. By the time a note appears, a post has typically reached 80 per cent of its total audience. Furthermore, 91 per cent of proposed notes never reach «helpful» status and are therefore never displayed [12]. The system requires agreement between contributors with different political perspectives — a design feature that prevents partisan abuse but also prevents timely correction of clear falsehoods when political agreement is unattainable.

The global fact-checking infrastructure itself is under severe strain. The IFCN's State of the Fact-Checkers Report 2024 found that financial pressures and harassment were the top concerns for fact-checkers worldwide [8]. Meta's withdrawal removed a funding pillar. The Duke Reporters' Lab counted 443 active fact-checking projects globally — down from 451, a 2 per cent decline — but the trajectory suggests further contraction as alternative revenue sources prove difficult to secure [8]. Latin American fact-checkers have been particularly affected by the concurrent loss of journalism grants and the freeze on USAID funding for international media organisations [8].

The «algorithms cause misinformation» position has substantial evidence behind it. Facebook's own internal research — disclosed by Frances Haugen in 2021 — demonstrated that the company's engagement-maximisation algorithms actively favoured outrage and misinformation [3]. The UK Parliament concluded that social media business models «incentivise the spread of harmful speech» [13]. The structural argument is compelling: remove the algorithmic accelerant and misinformation loses its primary distribution mechanism.

But the MIT study complicates this narrative significantly. Vosoughi, Roy, and Aral found that false news spread primarily through human sharing behaviour — not through bots or algorithmic promotion [1]. People shared false content because it was novel and emotionally arousing. «False news is more novel than true news,» the researchers wrote, «which suggests that people were more likely to share novel information» [1]. The implication is uncomfortable: even without algorithmic amplification, human psychology alone may drive the preferential spread of falsehood.

A comprehensive review published in 2024 attempted to reconcile these positions, concluding that «existing evidence suggests that algorithms mostly reinforce existing social drivers» of misinformation [13]. This framing — algorithms as accelerants rather than originators — has significant implications for policy. If algorithms are the primary cause, platform reform can address the problem. If they are accelerants of pre-existing human tendencies, platform reform is necessary but insufficient.

The trust crisis adds another dimension. Gallup's 2025 survey found that only 28 per cent of Americans trust mass media — the lowest figure in the poll's 50-year history and the first time trust has fallen below 30 per cent ✓ Established Fact [7]. The partisan divide is extreme: Republican trust stands at 8 per cent, Democratic trust at 51 per cent [7]. When institutional media is not trusted, the audience for misinformation expands — not because people are gullible, but because they have lost confidence in the institutions that were supposed to provide reliable alternatives.

The honest assessment is that both positions contain significant truth. Human psychology creates the demand for misinformation — novelty, emotional arousal, confirmation of existing beliefs. Algorithmic amplification creates the supply side — a distribution mechanism that can deliver false content to millions within minutes, at zero cost, with no editorial gatekeeping. The interaction between human vulnerability and algorithmic amplification produces the misinformation machine. Addressing one without the other will not solve the problem.

The debate also has a strategic dimension. Platform companies have consistently emphasised the «human behaviour» explanation because it externalises responsibility. If misinformation is a human nature problem, platforms cannot be expected to solve it — they are merely neutral conduits. The internal documents disclosed by Haugen suggest this framing is self-serving: Facebook knew its algorithms amplified misinformation and chose not to fix the problem because doing so would reduce engagement [3]. The causation debate is not merely academic — it is the arena in which accountability is contested.

Consider the asymmetry in economic terms. The digital advertising market is worth €625 billion and growing [13]. US newspaper advertising revenue has fallen from $49.4 billion in 2005 to $10.5 billion in 2024 [4]. The fact-checking sector — 443 organisations globally — is losing its primary funding source as Meta withdraws [8]. The ratio of resources dedicated to producing misinformation versus verifying information is not merely unequal — it is diverging exponentially.

The temporal asymmetry is equally stark. Misinformation travels at the speed of algorithmically amplified human emotion. Corrections travel at the speed of editorial process, regulatory procedure, and crowdsourced consensus. Community Notes take an average of 15 hours to appear [12]. EU enforcement actions unfold over months. Congressional investigations take years. By the time institutional responses materialise, the misinformation has already shaped public perception, influenced elections, and caused measurable harm.

Generative AI compounds the asymmetry. The cost of producing convincing misinformation has collapsed. A single operator with access to a large language model can now generate thousands of unique, contextually appropriate false narratives per hour. The cost of verifying a single claim — requiring human expertise, source evaluation, and contextual analysis — remains essentially unchanged. The production-verification cost ratio has shifted from unfavourable to catastrophic [15].

The institutional trust deficit further amplifies the asymmetry. With only 28 per cent of Americans trusting mass media [7], corrections issued by journalistic institutions are dismissed by a majority of the population before they are even read. When Republican trust in media stands at 8 per cent, fact-checks published by mainstream outlets are not merely ignored by that audience — they are actively interpreted as evidence of bias, further entrenching the misinformation they seek to correct. The fact-checking model assumed a shared baseline of institutional credibility that no longer exists.

What would a structurally adequate response look like? The evidence points to several necessary — though politically difficult — interventions. First, the advertising business model itself must be addressed. As long as platforms generate revenue by maximising engagement regardless of accuracy, the economic incentive for misinformation will persist. Options include digital advertising taxes, mandatory algorithmic transparency, or requirements for platforms to share revenue with the news organisations whose content drives engagement on their systems.

Second, public investment in journalism must be treated as infrastructure, not subsidy. The information needs of a democratic society do not self-fund in an attention-economy market. Countries like Denmark, Norway, and Finland, which have maintained robust public media systems, consistently rank lowest in vulnerability to misinformation. The correlation between public investment in journalism and resistance to falsehood is not coincidental — it is structural.

Third, the temporal asymmetry must be addressed through systemic design, not post-hoc correction. Regulatory frameworks that act after viral misinformation has already spread are inherently insufficient. Pre-distribution verification, algorithmic friction for unverified claims, and mandatory source labelling would shift the burden from reactive correction to proactive prevention.

The evidence is clear. The misinformation machine is not a content problem solvable by better moderation. It is a structural problem embedded in the economic architecture of the modern information ecosystem — an architecture in which truth is disadvantaged by design. The advertising model rewards engagement over accuracy. The journalism sector that once provided a counterweight has been economically hollowed. The regulatory response, while developing, operates at institutional speed against a threat that moves at algorithmic speed. The structural asymmetry is not a solvable glitch. It is the defining feature of the information age — and addressing it will require structural solutions that match the scale of the machine itself.