The Filter Bubble: They Switched Off the Algorithm for 23,000 People

The filter bubble is the idea that personalised feeds wall you into agreeable information and that this is what polarises people. It was named in 2011, it is intuitive, and it has been the organising metaphor for a decade of writing about social media.

In 2023 it was tested properly - with randomised experiments inside the platforms themselves, at a scale nobody outside them could achieve.

The results are not what the metaphor predicts, and the honest account of them is more interesting than either the original claim or the backlash to it.

What was actually done

A collaboration between independent academics and Meta ran a set of pre-registered randomised experiments during the 2020 US election, published simultaneously in Science and Nature.

Guess and colleagues replaced the algorithmic feed with a reverse-chronological feed for consenting Facebook and Instagram users for three months.

The intervention worked in the sense that it changed the experience enormously. Time on platform fell substantially. The composition of content changed. Users saw more content from sources they did not follow closely, and more political and untrustworthy content on Facebook.

It did not detectably change political attitudes, polarisation, or issue positions.

Companion papers tested other levers. Reducing exposure to reshared content changed what people saw and knew - it lowered news knowledge - without moving polarisation. Reducing exposure to like-minded content by roughly a third likewise failed to shift attitudes. González-Bailón and colleagues measured ideological segregation across Facebook and found it real and asymmetric, concentrated among a comparatively small set of sources, with the segregation on the supply side more than the feed algorithm side.

The pattern across all four is consistent: the algorithm shapes exposure a great deal and attitudes very little, at least over three months.

Why this is not the end of the argument

Three caveats have been raised, and two of them are serious.

The experiments ran during a specific and unusually intense political period - the final months of the 2020 US election, when attitudes were probably as fixed as they ever get. An intervention that fails to move opinion in October 2020 might move it in a quieter year.

Three months is short relative to the timescale on which the filter bubble is supposed to operate. Nobody claims a feed rewires you in a quarter.

Meta’s own algorithm continued running during the study period, and one published critique noted that platform-side changes during the election period complicate interpretation. This is a real limitation of doing experiments inside a company.

None of these rescues the strong version of the claim. But they do mean “the filter bubble is a myth” overstates what was shown. What was shown is that switching off personalisation for three months does not change what people think, which is narrower and more specific.

The evidence before 2023 was already mixed

Bakshy, Messing and Adamic (2015), in Science, examined exposure to ideologically cross-cutting content among Facebook users who reported a political affiliation. Their conclusion was that individual choice mattered more than the algorithm in determining what people clicked. The paper attracted substantial methodological criticism - over its sample restriction to users who self-identify politically, and over how it partitioned algorithmic from individual effects - and both the finding and the criticism are worth knowing.

Work outside the platforms pointed the same way. Flaxman, Goel and Rao found that search engines and social media were associated with higher ideological distance between individuals but also with greater exposure to opposing perspectives - the two are not contradictory, because both diversity and segregation can rise at once. Dubois and Blank found that people with diverse media diets and higher political interest were unlikely to be in an echo chamber at all, and that the echo chamber is a minority phenomenon rather than a general condition.

So the 2023 experiments did not overturn a settled consensus. They resolved, in one direction, a question the observational literature had been arguing about for eight years.

What the finding probably means

If personalisation shapes exposure strongly but attitudes weakly, the interesting question becomes what is doing the work instead.

The most parsimonious reading is that people select their own information environment, and the algorithm mostly accommodates them. González-Bailón’s segregation result points there: the asymmetry lives substantially in what sources exist and who chooses to follow them, not only in what the ranking does with them.

That is a less satisfying story than a machine building a bubble around you, because it removes the villain and puts the mechanism partly in the audience. It also predicts something the chronological-feed experiment found: give people an unsorted feed and they will still assemble roughly the same diet, because the accounts they follow have not changed.

The design lesson is not that ranking is harmless. It is that ranking is downstream of choice, and interventions aimed only at the ranking will underperform expectations - which is exactly what happened.

What this does not establish

That algorithms have no effects. They demonstrably changed exposure, time on platform, and in one experiment news knowledge. Those are effects.

That polarisation is not a problem. These experiments tested one proposed cause of it.

That the result generalises beyond three months, beyond the US, or beyond an election period. It does not, and the authors say so.

That the observational literature was wrong. It was mixed, and it remains mixed on questions the experiments did not address.

We have not printed effect sizes for the 2023 experiments. The pattern of findings is reported as the papers state it; the numeric estimates were spread across four papers and extensive supplementary materials, and we were unable to verify them at the level of precision this programme requires. Where we could not verify a number, we have described the direction instead - which is the claim the papers themselves foreground.

What follows for a reader

Be suspicious of the bubble metaphor as an explanation. It describes exposure well and attitude change badly, and it has been tested at a scale that few social-science claims ever reach.

Notice that “switch to chronological” is a solution aimed at the wrong variable. It was the intervention tested, it worked mechanically, and it produced no attitudinal benefit. If you have switched your feed to chronological for the sake of your politics, the evidence says you have changed your experience and not your mind. There may be other reasons to do it - less time on platform is one the experiment measured.

The lever that did something was subtraction of content, not reordering it. Reducing reshares changed what people knew. That is a stronger effect on the thing feeds actually do - moving information - than anything the reordering achieved.

And treat the confident version of this claim, in either direction, with caution. A decade of writing asserted the filter bubble on observational data. A single set of experiments, however large, in one country during one election, is not the last word either.

Algorithmic curation is one of several places where the design story and the measured effect diverge - see the persuasive design guide for the wider set, what your feed knows about you for what personalisation is actually built from, and infinite scroll for a design change that was deployed globally and never tested at all.