Social media algorithms are designed to keep people scrolling by showing them engaging posts, but the content they promote often contradicts the personal values of the people reading it. Research published in the Proceedings of the National Academy of Sciences reveals that the feed algorithm on the platform prioritizes content that conflicts with users’ explicit beliefs. The researchers suggest this dynamic may be driven by users replying to posts they dislike, a habit that appears to disproportionately affect left-leaning users.
Social media platforms rely on automated ranking systems to curate content for their users. These algorithms select posts from a larger pool of available content, including accounts a person follows, to populate a personalized feed. The primary goal of these systems is to maximize user attention by predicting which posts will generate the most interaction.
Platform designers generally operate on the assumption that interaction indicates preference. If a person likes, shares, or comments on a post, the system registers that action as a signal to show them more of the same. But researchers have raised questions about whether reactive behaviors online actually mirror what people care about in real life.
Sometimes, people interact heavily with content that makes them angry or defensive. If an algorithm simply chases interaction without understanding the context, it risks promoting divisive material while neglecting the core principles people say they hold.
Researchers Ziv Epstein of the Massachusetts Institute of Technology and Michael S. Bernstein of Stanford University led a team to investigate this gap. They sought to measure the relationship between the values people explicitly claim to hold and the values embedded in the content that algorithms choose to amplify. The team framed their research around a widely accepted psychological framework known as the Theory of Basic Human Values.
Developed by social psychologist Shalom H. Schwartz, this framework identifies nineteen foundational motivations recognized across cultures. These include a desire for societal security, a commitment to tradition, an openness to change, or a focus on caring for others. By mapping these principles onto social media posts, the researchers could track how a platform’s automated systems filter information based on moral themes.
To observe the algorithm in action, the research team recruited 715 American adults who actively used X, the platform formerly known as Twitter. Participants installed a custom browser extension that captured the posts appearing in two distinct feeds.
The first was the “Following” feed, which represents the raw, unranked inventory of posts from accounts the user chose to follow. The second was the “For You” feed, which represents the algorithm’s personalized selection of what the platform thinks the user should see.
The team used a specialized artificial intelligence model to scan the text of these social media posts. The system evaluated whether a post expressed specific values from the psychological framework, allowing the researchers to quantify the moral and social themes present in large datasets of text. By comparing the raw inventory to the curated feed, the team could calculate which types of content the platform systematically promoted or suppressed.
The analysis showed that users generally follow accounts that align with their personal values. When looking at the unranked inventory, the posts reflected the beliefs the participants had reported in surveys.
However, the algorithm’s curation flipped this dynamic. The researchers found an overall negative correlation between a person’s stated values and the values amplified by the For You feed. The automated system actively promoted content that was misaligned with the user’s explicit beliefs.
At the highest level of categorization, the algorithm showed a slight preference for individualist values over collectivist values. Certain types of content received consistent boosts across the platform. The algorithm amplified posts expressing a need for conservation and order, such as rule conformity, societal security, and tradition. It suppressed posts expressing self-transcendent values, like caring for others, dependability, and preservation of nature.
The researchers then explored how these algorithmic choices varied across political affiliations. They found that the algorithm promoted misaligned content to users on both sides of the political aisle, but the discrepancy was wider for Democrats.
To understand why the algorithm behaves this way, the researchers conducted a subsequent analysis focusing on engagement behaviors. They tracked how participants interacted with posts on X for a 24-hour period after the initial data collection. They categorized these interactions by type, looking separately at likes, reposts, and replies.
The researchers found that when people liked or reposted content, those posts generally matched their personal values. These actions represent an intentional endorsement of the material.
Replying to posts, however, told a different story. The act of replying was less aligned with users’ beliefs than liking or reposting, consistent with a form of confrontation. This pattern was especially pronounced among Democrats, whose replies were, on average, slightly misaligned with their stated values.
Because X’s algorithm weights replies heavily when deciding what to promote, the researchers propose that this confrontational behavior may create a feedback loop. When users argue with posts they dislike, the algorithm may register the high engagement and feed them more of the same frustrating material. This could flood the user’s feed with content that provokes a reaction rather than content they actually agree with.
In a final analysis, the team examined the overall, platform-wide engagement metrics for the posts in the participants’ feeds. They found that globally popular content tended to align more closely with Republican values. This suggests that right-leaning users encounter more content that naturally aligns with their beliefs without needing to engage in the confrontational reply loops that drive the feed for left-leaning users.
The study relies on observational data, meaning the researchers tracked natural user behavior without introducing controlled changes to the platform’s algorithm. Because the environment was not experimentally manipulated, the research cannot definitively prove that confrontational replies are the sole mechanism causing the algorithmic misalignment.
Future studies will need to test this dynamic in controlled settings to isolate the exact cause and effect. Researchers will also need to replicate the findings with larger sample sizes to verify the strength of the proposed feedback loop across different demographics.
These results raise questions about the design of social media ranking systems and what types of human behavior they should reward. If automated systems learn primarily from friction and argument, they will continue to produce environments that contradict the preferences of the people using them.
The study, “Value misalignments in X’s feed algorithm is a reflection of value tensions in engagement,” was authored by Ziv Epstein, Farnaz Jahanbakhsh, Tiziano Piccardi, Axel Peytavin, Isabel Gallegos, Shardul Sapkota, Dora Zhao, Johan Ugander, and Michael S. Bernstein.