The Attention Machine

← Symptoms

Radicalisation Pipelines

The attention machine was not built to radicalise anyone. It was built to hold attention. Extreme content holds attention well. That overlap is where this harm lives.

How to read this page. Tap any underlined word to see the precise academic term and a short definition. Expand any "Deeper" box for the evidence and the parts that are still contested. The main text works on its own — you can skip both and still get the whole argument.

The same machine, very different beliefs

The clearest sign that this is about the machine, and not about any one ideology, is that the same pattern sits behind movements that hate each other. The beliefs are opposites. The mechanics are identical.

Three things the machine reliably hands to anyone building a following:

Reach. Among millions of people, a recruiter only needs the few who are receptive. The same targeting that finds customers for a product finds the lonely, the angry, and the searching for a cause.

Belonging. A group that makes someone feel chosen is the oldest pull there is. Online, that group is available instantly, at any hour, to anyone who feels they have none.

Escalation. There is always something one step further on, and a system that sorts content by what holds attention keeps offering it. Mild grievance is more engaging than calm; sharp grievance is more engaging than mild. The gradient points one way.

Islamist recruitment: the early template

The first version of this story to reach the public was the so-called Islamic State. J.M. Berger and Jonathon Morgan's 2015 census for Brookings identified at least 46,000 Twitter accounts used by ISIS supporters in late 2014 — a coordinated, networked operation that used ordinary mainstream platforms to project reach and pull recruits toward private channels. When the platforms eventually cracked down, the operation did not stop; it migrated to more closed apps. The lesson others took from it was not the ideology. It was the method.

Far-right violence: Christchurch to Buffalo

The same method was turned to the opposite politics. In 2019 the Christchurch mosque attacker, active in far-right online spaces, livestreamed his attack and posted a written manifesto designed to spread — documented by New Zealand's Royal Commission of Inquiry (2020).

Three years later, the New York Attorney General's investigative report into the 2022 Buffalo shooting traced the pathway in unusual detail. It concluded that fringe platforms radicalised the shooter, that livestreaming services were weaponised to publicise the attack and invite copycats, and — the detail that matters most here — that he had watched the Christchurch footage on 4chan. The pipeline crossed platforms, and it crossed cases: one attack became the template for the next. That is the machine working exactly as designed, pointed at an end no engineer intended.

Incel violence: the same shape again

The pattern repeats in the manosphere's most extreme corner. The wider movement attributes male difficulties primarily to women and feminism; at its fringe, that grievance has preceded real-world violence. The content shares nothing with the two cases above. The mechanics — reach, belonging, escalation — are the same again.

Does the algorithm do it?

The popular version of this story is the "rabbit hole": the recommendation engine takes an ordinary person and walks them, click by click, to extremism. The honest answer is that the evidence is genuinely split. One large study found users migrating toward more extreme content over time (Ribeiro and colleagues, 2020). A second, tracking more than 300,000 people on and off the platform, found little evidence that the algorithm itself drives this — demand from the user, and links arriving from elsewhere, mattered more (Hosseinmardi and colleagues, 2021).

The careful reading sits between the headlines. The machine is the ground this grows on and the accelerant on top of it: it lowers the cost of finding extreme content, of finding each other, and of escalating. It is not a simple cause that turns ordinary people into killers. As with the mental-health debate elsewhere on this site, the correlation and the mechanism are clearer than the simple causal arrow — and saying so plainly is what keeps the argument honest.

Why the antidote is the same

Because the mechanism is blind to ideology, the defence is the same for all of them: understanding how it works. That is the whole argument of this site, applied to its darkest case. The response then splits three ways — safer platform design, which runs through every page here; law enforcement, because the recruitment and the violence are deliberate crimes; and, most immediately, the awareness of the people around someone being drawn in. The page on how networks recruit the young and the what-you-can-do guide are the practical companions to this one.

The contested part — does the recommendation engine radicalise? +

This is the most over-claimed area in the whole debate, in both directions, so it is worth being precise.

The case that it does. Ribeiro and colleagues (2020) audited hundreds of YouTube channels and tens of millions of comments and found that users consistently migrated from milder to more extreme communities, and that more extreme channels were reachable through recommendations. They were careful, though, to say their work showed that migration happens, not why — it offers "little insight" into the cause.

The case that it mostly doesn't. Hosseinmardi and colleagues (2021) followed a representative panel of over 300,000 people across the whole web, not just on YouTube. They found radical-content consumption was small and stable, dominated by user preference and by links arriving from other sites, with "little evidence that the YouTube recommendation algorithm is driving attention" to it. A related strand (Chen and colleagues, 2023) found the system serves more extreme content mainly to people already showing resentment — amplifying an existing demand rather than manufacturing it.

What this leaves standing. The strong claim — that the algorithm converts ordinary users into extremists on its own — is not well supported. The weaker, sturdier claim is: the platforms lower every cost involved in radicalisation (finding the content, finding the group, escalating, and broadcasting the result), and they preferentially feed extreme material to those already leaning that way. Kevin Roose's 2019 New York Times account of one man's path is a vivid illustration of the mechanism, not proof of its prevalence. Holding both halves at once is the only honest position.

Why this page is careful with the cases +

The factual claims here rest on official investigations and peer-reviewed work, not on the attackers' own material. Christchurch is documented by a New Zealand Royal Commission of Inquiry (whose remit was state agencies, not the platforms — so we use it for the established facts of his online activity, not for a finding about platform liability). Buffalo is documented by the New York Attorney General's investigative report. The ISIS figures come from a Brookings analysis paper.

We describe these only at the level the investigators chose to make public, and we deliberately do not reproduce manifestos, name the perpetrators, or set out methods — naming and detailing the techniques helps the people who use them and re-traumatises survivors, the same restraint we apply on the Recruited to Harm page. And we make no claim about why any single individual acted: the population-level mechanism is the subject here, not a diagnosis of one person.

Symptoms: Xenophobia & Violence →
Symptoms: Recruited to Harm →
The Attention Machine: State Actors →