The feed is not neutral: how a Virginia Tech study exposes the bias baked into social media timelines

On 20 July 2026, researchers at Virginia Tech published a finding that lands awkwardly for two decades of platform-criticism orthodoxy: people are simply drawn to excitement, and the feeds that follow them reflect it. The study, summarised by Phys.org and authored by faculty in the university's College of Science, argues that human bias toward stimulating, novel and aspirational content shapes what surfaces on social media timelines more than the platforms' own curation logic is willing to admit.

The finding cuts against a popular narrative in which recommendation engines are the primary authors of what users see. The Virginia Tech work suggests the relationship runs the other way too: the platforms are following users as much as users are following the platforms, and the result is a feed cluttered with international travel, skydiving, fitness transformations and other high-arousal content that crowds out quieter material.

What the researchers actually measured

The Virginia Tech team framed the project around a deceptively simple question: when users are given a feed, do they choose excitement, or does the feed choose excitement for them? Their answer, drawn from behavioural observation rather than internal platform data, leans toward the user. Participants gravitated to content depicting novel experiences, extreme sports, scenic travel and visible status markers. The same pattern held across age cohorts, though younger users showed stronger pull toward adrenaline-coded imagery.

This matters because the dominant critique of social platforms has run the other way. Critics, regulators and a thicket of academic papers have argued that the recommendation layer itself is the engine of radicalisation, of eating-disorder content, of political polarisation. The Virginia Tech paper does not refute those claims outright. It does suggest that a non-trivial share of the variance in what a user encounters is driven by where the user's attention already wants to go.

The counter-narrative: the algorithm is still the gatekeeper

The plausible alternative reading is that platforms have spent fifteen years training their users to want exactly the content the algorithm is best at serving. Recommendation systems optimise for engagement, defined narrowly as clicks, watch time and shares. They have, in effect, conducted a vast unsupervised experiment in shaping user preference. The Virginia Tech researchers may be measuring a population that has already been conditioned.

That is the harder claim to test, because it requires either internal platform data, which firms guard as trade secrets, or longitudinal studies of users before and after algorithmic exposure, which are ethically and logistically fraught. The Phys.org summary does not claim to settle the chicken-and-egg problem. It does claim to show that preference and feed reinforce one another, which is a narrower finding than either camp usually admits.

What the platforms say, and what they do not

None of the major platforms have commented on the Virginia Tech study specifically. Their standard line, rehearsed in congressional hearings and in SEC filings, is that recommendation systems are neutral tools that reflect user interest. That line is increasingly hard to defend on its own terms. Engagement-optimised ranking does not operate on a clean signal of interest; it operates on a noisy one, amplified by network effects and by the platform's own commercial interest in keeping users scrolling.

At the same time, the platforms are correct that user agency is non-zero. People can close the app. They can unfollow. They can search for less stimulating content. The data behind the Virginia Tech study suggests that, on average, they do not.

What changes, and what does not

The structural picture is this: social platforms are caught between two constituencies they cannot reconcile. Regulators, on one side, want demonstrable control over what the algorithm surfaces. Users, on the other, vote with their attention for the most arousing possible content. The platforms' business model requires keeping the second constituency happy, which makes the first constituency's job harder.

The Virginia Tech finding, read carefully, does not exonerate the platforms. It does relocate some of the responsibility. If users reliably prefer excitement to substance, then any system designed to maximise engagement will produce feeds that look like a tourism brochure stitched together with a Red Bull sponsorship. The harder question, which the study does not answer, is whether platforms have a duty to surface less stimulating, more informative content even when users will scroll past it. The answer to that question is now firmly in the regulatory lane.

The research is one paper, not a verdict. What it does establish is that the standard platform-criticism formula, in which users are passive recipients of algorithmic output, is incomplete. The picture is messier, and more interesting, than the critics usually admit. The platforms will, predictably, treat the study as vindication. It is not. It is an invitation to a harder argument about what feeds are for.

This publication framed the study as a complication of the standard critique rather than as a vindication of either side; the wire treatment emphasised novelty, while the structural question of who bears responsibility for what users see remains unsettled.