In the early days of social media, the prevailing assumption was that the internet would be a broadly democratising force for information. The barriers to publishing were collapsing. Access to knowledge was expanding. The diversity of voices available to any individual reader had never been greater. The optimism was genuine and not entirely misplaced — those things were all true.
What the optimism did not account for was curation. Not all information, it turned out, would reach all people equally. Between the individual and the vast landscape of available content stood something new: systems designed to select, rank, and surface material according to optimisation targets that had nothing to do with accuracy, balance, or social benefit. They had to do with engagement — the clicks, shares, watch time, and reactions that translated, through advertising markets, into revenue.
The consequences of this arrangement have taken two decades to become fully visible. They are now difficult to look away from.
These figures are not from advocacy groups. They are from platform internal research, academic studies published in peer-reviewed journals, and investigative reporting based on leaked documents. The platforms themselves, in many cases, have known for years that their systems produce outcomes they cannot fully defend. The gap between what they have known and what they have disclosed is itself a significant part of the story.
01
Engagement optimisation amplifies outrage
Content that provokes strong emotional responses — particularly anger and moral indignation — consistently outperforms neutral content on engagement metrics. Algorithms trained to maximise engagement therefore systematically surface more of it, regardless of its accuracy or social cost. This is not a bug or an oversight. It is the direct consequence of optimising for the wrong objective.
What the research shows
A 2021 study in Science found that false news is 70% more likely to be retweeted than true news, and reaches people six times faster. Emotional novelty — not source credibility — is the primary driver.
02
Filter bubbles narrow the information diet
Personalisation systems infer user preferences from behaviour and serve progressively more of what already aligns with demonstrated interests. The effect is gradual and largely invisible: the range of perspectives, sources, and counterarguments available to a user narrows over time, not because anything is deliberately hidden, but because the system has learned what keeps this person engaged and optimises accordingly.
What the research shows
The evidence on filter bubbles is more nuanced than popular accounts suggest — cross-cutting exposure still occurs. But research by
Eli Pariser and subsequent work shows that algorithmic curation meaningfully reduces exposure to challenging viewpoints relative to chronological or random feeds.
03
Recommendation pathways escalate toward extremity
Perhaps the most concerning mechanism: the documented tendency of recommendation systems to progressively suggest more extreme content as users follow a topic. A user who watches mainstream political commentary is recommended increasingly partisan content; a user interested in alternative health information is nudged toward misinformation. The system is not malicious — it is optimising for the next click, and more extreme content tends to produce more certain engagement than moderate content.
What the research shows
Internal Facebook research, surfaced in the 2021 whistleblower disclosures, found that its own algorithm led users to increasingly extreme groups — a dynamic its engineers described as the platform "taking you down rabbit holes."
The question of what to do about algorithmic influence on belief is genuinely difficult, in part because the mechanisms operate at a level of complexity and scale that outpaces simple regulatory frameworks. Three categories of response have shown varying degrees of promise.
None of these responses is sufficient on its own, and none addresses the fundamental tension at the heart of the problem: the business models that fund these platforms at their current scale depend on engagement metrics that are structurally misaligned with the public interest in accurate, balanced information. Until that tension is resolved — through regulation, through the emergence of alternative models, or through some combination — the systems will continue to optimise for what they were built to optimise for, with outcomes that are now well-documented and not particularly surprising.
What is perhaps most striking about the current moment is not the scale of the problem but the clarity with which it is now understood. The mechanisms are no longer opaque. The evidence is extensive. The question of what happens next is less a matter of knowledge than of collective will — which is, of course, also being shaped, in part, by algorithms.
The systems will continue to optimise for what they were built to optimise for. The question of what happens next is less a matter of knowledge than of collective will — which is, of course, also being shaped, in part, by algorithms.
How do recommendation algorithms influence what people believe?
Recommendation algorithms influence belief by systematically amplifying content that generates strong emotional responses — particularly outrage and moral indignation — because that content drives higher engagement metrics. Over time, this narrows the range of perspectives a user encounters and can progressively escalate exposure toward more extreme content. The effect is not deliberate manipulation but the predictable outcome of optimising for engagement in an information environment where emotional content outperforms accurate content.
What is the filter bubble effect?
The filter bubble effect refers to the way personalisation algorithms gradually narrow a user's information diet by surfacing more of what they have already engaged with. The result is that challenging viewpoints, diverse sources, and contradictory evidence become progressively less visible — not because they are blocked, but because the algorithm has learned that they generate less engagement. Research shows algorithmic feeds produce measurably less exposure to cross-cutting perspectives than chronological feeds.
Does social media spread misinformation faster than true information?
Yes. A landmark 2018 study published in
Science by Vosoughi, Roy, and Aral found that false news spreads approximately 3.6 times faster than true news on social platforms, reaches more people, and penetrates deeper into social networks. The primary driver is novelty and emotional resonance — false news tends to be more emotionally novel than accurate reporting — rather than differences in who is spreading it.
What regulations exist for recommendation algorithms?
The EU's Digital Services Act (DSA), which came into full effect in 2024, is currently the most comprehensive regulatory framework for recommendation algorithms. It requires very large online platforms to conduct algorithmic risk assessments, provide access to independent auditors, and offer users at least one recommendation option not based on profiling. Regulatory frameworks in Australia, the UK, and the US are less advanced, though legislative attention to algorithmic accountability is increasing across all three jurisdictions.
