When disinformation spreads, the instinctive response is to correct it. Publish the fact-check, link the source, show the data. The assumption behind this response is simple: people believe false things because they lack the right information, and giving them that information will change their minds.

For some people, in some situations, this works. For a growing part of the online public, it does not, and it can even backfire. To understand why, we need to distinguish between two phenomena that are often confused under the same label: the epistemic bubble and the echo chamber. They look similar from the outside, but they fail in different ways, and they need different remedies.

The philosopher C. Thi Nguyen (2018) offers the clearest distinction between the two.

An epistemic bubble is an information network from which relevant voices have been left out by omission. Sometimes the omission is deliberate: we avoid views that make us uncomfortable. More often it is entirely unintended. Even if we never set out to avoid disagreement, our friends on Facebook tend to share our views and interests. When we use networks built for social reasons as our source of news, we end up missing opposing views and encountering an exaggerated degree of agreement.

An echo chamber is something different: a network in which other relevant voices have been actively discredited. A bubble simply leaves contrary views out; an echo chamber teaches its members not to trust them. In this respect, Nguyen argues, an echo chamber works much like a cult. It isolates its members by actively alienating them from any outside source. Outsiders are labelled as malicious and untrustworthy, and members' trust is narrowed and redirected towards a few voices on the inside.

The difference can be put in one sentence: in a bubble, other voices are not heard; in an echo chamber, other voices are actively undermined.

The two have very different exits. Breaking an epistemic bubble is relatively simple: expose its members to the information and arguments they have missed. An echo chamber is much harder to break, because its members have already been prepared to dismiss exactly that information (Nguyen, 2018).

Neither bubbles nor echo chambers are new. Communities of like-minded people have always reinforced their own beliefs, and an echo chamber in the general sense is simply a situation in which a person's opinions are strengthened by repeated exposure to information that confirms them (Kte'pi, 2021). What social media changes is the scale and the speed, because the selection of what each of us sees is no longer left to chance or to our own habits. It is engineered.

Personalised feeds rank content by predicted engagement: how likely we are to like, comment, share or keep watching. TikTok's recommendation system, for example, combines what an individual has interacted with in the past with patterns observed across millions of similar users. Over time, this tends to reinforce preferences that have already been observed, producing an increasingly narrow and personalised stream of content and reducing exposure to divergent or contradictory information (Zhou, 2024). Research on social media more broadly finds that users tend to prefer information that confirms their existing beliefs and to ignore contradicting information, which makes them more likely to cluster into groups of like-minded people and favours the formation of polarised communities (Cinelli et al., 2021).

This process builds bubbles almost automatically. What turns some of them into echo chambers is emotion. Expressing and sharing emotions online is one of the ways people signal which group they belong to, and emotions spread from person to person. In a large-scale study, Kramer, Guillory and Hancock (2014) showed that the emotions people see expressed on Facebook influence the emotions they express themselves, evidence of emotional contagion at massive scale. Goldenberg and Gross (2020) describe how this contagion works in digital settings, through exposure to others' emotional expressions and the social appraisals that come with them.

Put these mechanisms together and the picture becomes clear. Feeds that reward engagement amplify content that provokes strong emotions, such as outrage and fear, because those emotions keep people on the platform. Shared outrage strengthens group identity, and a strong group identity is exactly what makes outsiders easy to cast as enemies. The bubble closes, and then it hardens.

This is where fact-checking runs into trouble. A fact-check is designed for a bubble. It assumes the person has simply not seen the relevant evidence, and that once they see it, they will weigh it fairly.

Inside an echo chamber, that assumption fails. Nguyen (2018) makes an important point here: members of an echo chamber are not irrational. They reason consistently from what they have been taught, but they have been systematically misinformed about where and how to place their trust. When one political camp presents economic statistics to support its argument, the other camp may reject them not on the basis of their validity, but by attacking the source, claiming that the publication is biased or that the academic institution that produced the data is corrupt.

From inside the chamber, a fact-check from a mainstream outlet is not neutral evidence. It is exactly the kind of message from an untrustworthy outsider that members were warned about. The correction can even confirm the story the chamber tells about the outside world: they are trying to manipulate us. More information, delivered through channels the audience has already learned to distrust, does not open the chamber. It reinforces its walls.

The problem, in other words, is not a shortage of facts. It is a distortion of trust.

A distortion of trust is not only a problem for public debate; it is also an opportunity for anyone who wants to exploit it. Military and academic institutions now describe this under the name of cognitive warfare. Claverie and du Cluzel (2022) define it as the art of using technological tools to alter the cognition of human targets who, most of the time, are unaware of it. Unlike cyber warfare, which targets information systems, cognitive warfare targets the human mind itself: memory, perception, attention and the capacity to reason. Its goal, as they put it, is to wage war "on what an adversary community thinks, loves or believes, by altering perceptions" (p. 2).

Echo chambers make populations unusually vulnerable to this kind of operation. When malicious actors can identify large segments of a population that no longer trust traditional media, trust only certain alternative sources, and automatically reject information from outside, they can exploit that structure systematically. They know their content will be received without verification inside the chamber, and that any correction from outside will be dismissed. The coordinated campaign that preceded the annulment of Romania's 2024 presidential election is one of the most documented recent examples of such an operation running through a recommendation algorithm.

Seen this way, an echo chamber is not just a social phenomenon. It is an attack surface.

If the problem is trust rather than information, the remedies have to look different.

For bubbles, exposure still works. Making sure people encounter perspectives they would otherwise miss remains valuable, and it is something platforms could design for, for example by limiting how far personalisation narrows a feed over time.

For echo chambers, the messenger matters more than the message. A correction is far more likely to be heard if it comes from a source the audience still trusts, ideally someone from inside the community, than from an institution the chamber has already learned to dismiss. Fact-checking that ignores who is speaking will keep reaching only those who did not need it.

Upstream, the design matters most. Echo chambers are not only the result of individual preferences, but also of product design decisions. Choosing to prioritise metrics such as likes, shares or time spent on the platform directly determines which content becomes visible and which is excluded (Cinelli et al., 2021). Feeds that do not systematically reward outrage would make it harder for closed worlds to form in the first place, and harder for others to weaponise them.

Fact-checking is not useless. But it treats a structural problem as an information problem. As long as our information environments are built to reward what divides us, correcting individual falsehoods will remain a necessary, but never sufficient, response.