Main

Social media platforms have fundamentally transformed human lives: a large and growing share of the global population connects with others, gets entertained and learns about the world through social media2. These platforms have also become increasingly important for political news consumption. A quarter of US adults report social media as their primary news source, and one half say they at least sometimes get news from these platforms3. Typically, platforms use feed algorithms to select and order content in personalized feeds for each user4. Before algorithms were introduced, users saw a simple chronological feed that displayed posts from followed accounts, with the most recent posts appearing at the top.

Public intellectuals and scholars have raised concerns about the potential adverse effects of social media, particularly feed algorithms, on social cohesion, trust and democracy5,6,7,8. These concerns arise from the spread of misinformation9,10,11, the promotion of toxic and inflammatory content12,13,14 and the creation of ‘filter bubbles’ with increasingly polarized content15,16,17,18. There is substantial rigorous quantitative evidence that internet access and social media indeed have important negative effects19,20,21,22. Research on search engine rankings also shows that the order in which information is presented can influence user behaviour and political beliefs23. However, previous literature on the effects of social media feed algorithms reports zero political effects. A large study of Facebook and Instagram, conducted by academics in cooperation with Meta during the 2020 US election, found that experimentally replacing the algorithmically curated feed with a chronological feed did not lead to any detectable effects on users’ polarization or political attitudes, despite causing a substantial change in political content and lowering user engagement with the platforms1. Similarly, studies on Google’s search engine and YouTube algorithms found little evidence of filter bubbles24,25,26,27. Studies of Meta platforms linking content to user behaviour and attitudes also found no impact, despite prevalent like-minded content and amplified political news28,29,30.

Yet, the fact that switching off a feed algorithm does not affect users’ political attitudes does not mean that algorithms have no political impact. If the initial exposure to the algorithm has a persistent effect on political outcomes, switching off the algorithm might show no effects despite its importance. For instance, this could happen because people start following accounts suggested by the algorithm and continue following them when the algorithm is switched off. In addition, different platforms may have different effects, for instance, due to different informational environments or the different objectives of their owners31,32,33,34.

We study the effects of X’s feed algorithm and find that switching the algorithm on substantially shifts attitudes on policies and current political news towards more conservative opinions without significantly affecting polarization or partisanship. We conducted a randomized experiment involving actual X users in the United States over a 7-week period in the summer of 2023.

Our study departs from the previous literature in several ways. First, we leverage the feature on X that allowed users in 2023 to choose between a chronological feed (the ‘Following’ tab) and an algorithmic feed (the ‘For you’ tab), where content was both added (showing posts from accounts not followed by the user) and reordered (prioritizing some posts while hiding others), compared with the chronological feed setting (as confirmed by our data). This feature enabled us to conduct two experiments simultaneously: examining the effects of switching the feed algorithm on for users who previously used the chronological feed and of turning it off for those who were on the algorithmic feed before the experiment.

Second, this feature allowed us to conduct the experiment independently, without cooperation from X. Hence, we avoid potential concerns specific to studies conducted in partnership with platforms35,36. Recently, another independent study used a browser extension to show that re-ranking content within X’s algorithmic feed by moving up or down posts expressing anti-democratic attitudes and partisan animosity influenced affective polarization37. In contrast, we randomize users’ exposure to X’s algorithmic feed as designed and implemented by X or to a chronological feed. We then quantify which posts the algorithm promotes and demotes, estimate its causal effects on users’ attitudes and behaviour, and leverage behavioural responses to provide evidence on the mechanism.

Finally, in addition to affective polarization and partisanship, we also study outcomes such as policy priorities and attitudes towards current political events, which may be less rigid and, therefore, more easily influenced by exposure to different social media content.

Our experiment took place more than 6 months after Elon Musk’s acquisition of Twitter, a few months following the publication of the platform’s source code and shortly after Linda Yaccarino assumed the role of CEO, yet about 1 year before Musk’s public endorsement of Donald Trump in July 2024. An earlier study examined changes in the content of users’ feeds on Twitter, when the platform introduced the feed algorithm in 2016, well before Musk’s takeover, and found that the algorithm already prioritized right-wing content38, despite different platform ownership. In addition to content analysis, our study focuses on the effects on real users’ behaviour and political attitudes38.

Design

We conducted an experiment with X users during July and September of 2023. This research resulted from a collaboration of academics independent of X. We obtained ethical approval from the Ethics Committee of the University of St. Gallen, Switzerland (see Supplementary Information section 1.1 for a detailed discussion of ethical considerations and the measures we implemented to uphold ethical integrity). We pre-registered the experiment with the American Economic Association’s registry for randomized controlled trials (AEARCTR-0011464).

The experiment had several phases: recruitment, a pre-treatment survey (collecting baseline characteristics), randomization into feed settings, a treatment phase (with participants using their assigned feed setting) and a post-treatment survey to gather self-reported outcomes. We also collected data on the content of users’ feeds and data on users’ online behaviour. Participants were recruited through the survey company YouGov, drawing from US-based registered members of the YouGov panel. Participation was voluntary and compensated.

YouGov contacted 13,265 participants to enter the pre-treatment survey. Of these, 3,434 were screened out as they were not active X users. By design, we only admitted participants who self-reported being active on X at least ‘several times a month’. A total of 8,363 participants provided informed consent and 6,043 completed the entire pre-treatment survey. During the survey, each participant was assigned randomly to a feed setting, which they were paid to stay on until completing the post-treatment survey. Admission to the pre-treatment survey occurred on a rolling basis in July 2023. At the end of August 2023, all participants were invited for the post-treatment survey, completed by the third week of September. The treatment phase duration varied slightly among participants, with the average time between entering and leaving the study being 7 weeks and 1 day, with s.d. of 3 days. During the treatment phase, participants received two reminders to stick to their assigned feed setting. A total of 4,965 participants completed the post-treatment survey, forming our main sample. Extended Data Fig. 1 presents the flowchart for the experiment, describing its structure and sample size at each stage. Supplementary Information sections 1.2–1.8 provide further details about the structure of the experiment, all collected data, compliance and attrition at each stage.

In the pre-treatment survey, we inquired about which feed setting the participants used before the experiment. Initially, 76% of participants were using the algorithmic feed, the default on X; and 24% were using the chronological feed. For the duration of the study, participants were assigned randomly with equal probability to either the algorithmic or the chronological feed.

The pre-treatment survey included questions about X usage, including purpose and frequency, as well as questions on life satisfaction, partisanship and feeling thermometers for Democrats and Republicans, which we use to calculate affective polarization. YouGov provided data on participants’ socio-economic backgrounds. All pre-treatment characteristics were fully balanced, except for a slight imbalance in the initial feed setting. The share of participants initially on the algorithmic feed setting was two percentage points higher among those assigned randomly to the algorithmic feed (77% versus 75%; two-sided t-test: P = 0.08; Supplementary Information section 1.7). In our analysis, we always control for the initial feed setting. Participants who had to switch their settings in either direction as a result of the experiment were more likely to drop out (22.7% versus 20.9%; two-sided t-test: P = 0.077). As demonstrated in Methods, selective attrition during the treatment phase does not drive our results.

In the post-treatment survey, we collected several groups of outcomes. First, we repeated questions about X usage, partisanship, affective polarization and life satisfaction. Furthermore, we asked participants to rank policy areas by priority and to express their views on current political events, specifically their assessments of the criminal investigations into Donald Trump and the war in Ukraine, using a series of questions. For the baseline analysis, we aggregate these outcomes into seven composite measures by topic: user engagement, partisanship and affective polarization, conservative policy priorities, belief that the investigations into Trump are unacceptable, pro-Kremlin attitudes towards the war in Ukraine and life dissatisfaction. We also construct an overall measure that combines all questions related to policy attitudes and current political news. In the Supplementary Information, we present results for each individual outcome.

At the end of both surveys, we offered participants the option to install and run a Google Chrome extension on their desktop computer for additional compensation. This extension, designed for the experiment, captured the first 100 posts seen under each feed setting and saved them to a file on the participant’s device, which they were asked to upload. A total of 784 participants ran the extension at the end of the pre-treatment survey, and 599 did so during the post-treatment survey.

For participants using the Google Chrome extension, we observed their actual X feed setting directly and, hence, their compliance with the assigned setting. The observed compliance rate in this sample was 89.15%. In addition, we asked all participants in the post-treatment survey whether they complied with the assigned feed setting during the study period. The self-reported compliance rate, that is, the share responding that they complied most of the time or always with the assigned setting, was 85.38% in the full sample (Extended Data Fig. 2).

We asked participants whether they agreed to share their X handles with us and grant permission to scrape their publicly available data, specifically, the names of the accounts they followed. All participants were informed about the purpose of our use of their X handles, and their decision to participate did not affect their compensation. We scraped the list of accounts followed by 2,387 participants after the post-treatment survey. Supplementary Information section 1.3 describes the construction of all variables and reports summary statistics.

Our experimental design allows us to evaluate the impact of switching the feed algorithm on and off. Intuitively, for respondents who were initially on the chronological feed, we compare the outcomes between those randomized to remain on the chronological feed and those randomized to switch to the algorithmic feed. Similarly, for respondents who were initially on the algorithmic feed, we compare the outcomes between those who remained randomly on the algorithmic feed and those who switched randomly to the chronological one. Our empirical strategy for baseline results is detailed in Methods, which also summarizes our analyses of attrition and compliance (discussed in detail in Supplementary Information section 1.8).

Impact of feed algorithm on attitudes and behaviour

Figure 1 illustrates the main results by summarizing raw data means by treatment status for selected individual outcomes in a series of bar charts (all tests are two-sided, similar graphs for all individual outcomes are presented in Supplementary Information section 2.1). The top row presents data for participants who were initially on the chronological feed. Among these participants, those assigned to switch to the algorithmic feed were 5.2 percentage points less likely to reduce their X usage than those who remained on the chronological feed (95% confidence interval (CI): 0.7, 9.7; P = 0.024). They were 4.7 percentage points more likely to prioritize policy issues considered important by Republicans, such as inflation, immigration and crime (95% CI: 0.7, 8.7; P = 0.023). They were also 5.5 percentage points more likely to believe that the investigations into Trump are unacceptable, describing them as contrary to the rule of law, undermining democracy, an attempt to stop the campaign and an attack on people like themselves (95% CI: 0.8, 10.2; P = 0.022). They were 7.4 percentage points less likely to hold a positive view of Ukrainian President Volodymyr Zelensky (95% CI: 1.8, 13.0; P = 0.009). Finally, they were 3.7 and 2.3 percentage points more likely to follow any conservative account (95% CI: 0.5, 7.0; P = 0.025) and any political activist account (95% CI: −0.1, 4.8; P = 0.061) on X, respectively. The bottom row of Fig. 1 presents the data summaries of the same outcomes for participants who were initially on the algorithmic feed. In sharp contrast to the top row, the outcome means in this group were not significantly affected by treatment assignment.

Fig. 1: Raw data summary for the main outcomes by initial feed setting and treatment.
figure 1

Means for selected outcomes by assigned treatment feed in the two samples: top, summarized outcomes for the sample of users initially on the chronological feed; bottom, summarized outcomes for the sample of participants initially on the algorithmic feed. The difference in the height of the two bars in each chart represents the unconditional ITT effect. Error bars, 95% CI for the ITT effect relative to the control mean (Control mean + ITT ± 1.96 × s.e.). The title of each chart presents the treatment-effect magnitude in percentage points (pp) and its P value (all tests are two-sided). The unit of observation is respondent. Sample sizes for each subgroup are reported below the bars. All outcomes are indicator variables except ‘Shares Republican priorities’, for which the minimum, first quartile, median, third quartile and maximum are: 0, 0, 0.33, 0.67, 1 for initial-Chronological to Chronological subsample; 0, 0, 0.33, 0.67, 1 for initial-Chronological to Algorithmic; 0, 0, 0.33, 0.67, 1 for initial-Algorithmic to Algorithmic; and 0, 0.33, 0.33, 0.67, 1 for initial-Algorithmic to Chronological. Supplementary Information section 2.1 presents similar bar charts for all individual outcome variables. Supplementary Information section 2.4 presents regression results for all outcome variables.

Source data

Full size image

Regression results using aggregate outcomes are presented in Fig. 2, which displays the results of regression analysis for the composite survey outcomes. Average intention-to-treat (ITT) effects are reported (Methods). On the left, we show the standardized effects of switching the algorithm on for participants initially using the chronological feed; on the right are shown the effects of switching it off for those initially using the algorithmic feed. For each outcome, we report two sets of regression estimates: unconditional, controlling only for the initial feed setting and the pre-treatment value of the outcome variable when available; and conditional, controlling for all collected pre-treatment covariates, with the specification chosen using generalized random forests (GRFs)39.

Fig. 2: ITT estimates of feed-setting changes on engagement and political attitudes.
figure 2

ITT effect estimates of switching the algorithm on and off (in s.d.). Left, effect of moving from the chronological to the algorithmic feed for users initially on the chronological feed. Right, effect of moving in the opposite direction for users initially on the algorithmic feed. For each outcome, the results of two specifications are reported. Blue, unconditional estimates with robust s.e., controlling only for the initial feed setting and, where applicable, pre-treatment outcome levels. Orange: conditional estimates, controlling for pre-treatment covariates using GRFs; 90% and 95% CIs are reported. Numerical effect sizes and P values correspond to the conditional estimates (all tests are two-sided). The unit of observation is respondent. From top to bottom, sample sizes are n = 4,965, n = 3,337, n = 4,965, n = 4,965, n = 4,596, n = 4,596 and n = 4,850. Tests are described in Methods. Supplementary Information Table 2.16 reports the exact numerical point estimates, s.e., CIs and sample sizes for every specification. All outcomes are standardized. Additional results are presented in Supplementary Information section 2. PCA, first principal component from principal component analysis.

Source data

Full size image

First, we confirmed that the algorithmic feed is more engaging. Among participants initially on the chronological feed, the first principal component of the user engagement measures increased by 0.14 s.d. for those assigned to the algorithmic feed, compared with those who remained on the chronological feed (95% CI: 0.03, 0.25; P = 0.014). For participants initially on the algorithmic feed, engagement declined by 0.06 s.d. for those assigned to the chronological feed, but this effect is not significant at the 5% level (95% CI: −0.12, 0.01; P = 0.081). Second, we find precisely estimated null effects of the algorithm on partisanship and affective polarization, whether the algorithm was switched on or off.

Outcomes related to attitudes towards policies and current political events were affected significantly by turning the algorithm on. Participants who switched from the chronological to the algorithmic feed prioritized a conservative policy agenda over a liberal one by 0.11 s.d., compared with those who remained on the chronological feed (95% CI: 0.02, 0.20; P = 0.016). The experiment also influenced views on the criminal investigations into Trump: those who switched to the algorithmic feed were 0.08 s.d. more likely to consider them completely unacceptable (95% CI: 0.01, 0.16; P = 0.026). This effect is larger and more precisely estimated for aspects of this belief held by a smaller share of participants in the control group (as shown in Supplementary Information Fig. 2.4), indicating that the effect is concentrated among more loyal Trump supporters. Regarding Russia’s invasion of Ukraine, for participants initially on the chronological feed, exposure to the algorithmic feed increased pro-Kremlin attitudes by 0.12 s.d. (95% CI: 0.03, 0.21; P = 0.007). Finally, we address multiple hypothesis testing concerns by aggregating all questions concerning policies and current political news into a single index. We find that this index is also 0.12 s.d. higher for those who were randomized to switch from the chronological to the algorithmic feed, compared with those who stayed on the chronological feed (95% CI: 0.04, 0.21; P = 0.004).

Remarkably, none of the political attitudes were affected by the reverse switch from the algorithmic to the chronological feed, that is, by switching the algorithm off. This is consistent with the earlier findings of the Meta study1. We also found no significant impact of the experimental treatments on subjective wellbeing. As demonstrated by the similarity between the two specifications, unconditional versus conditional on pre-treatment covariates, the significant impact of switching the algorithm on, and the absence of any effect from switching it off, are robust to the choice of covariates and model specification. Supplementary Information Fig. 2.9 presents results for all individual outcomes in the same format as in Fig. 2. We also verify that the ITT estimates are very similar to the local average treatment effects (LATE) on self-reported compliers, presented in Extended Data Fig. 3 and discussed in detail in Methods.

The sample size is insufficient to reliably test for heterogeneities in these effects. However, the split of the sample by self-reported pre-treatment partisanship into Democrats versus Republicans and Independents (or other political parties) indicates strongly that the results are driven by Republicans and Independents, whereas Democrats’ views are largely unaffected by the experiment (Extended Data Fig. 4).

The initial feed setting is correlated with participants’ socio-demographic characteristics, as it may reflect individual preferences and because selecting the chronological feed requires an active choice: those on the chronological feed are more likely (based on two-sided t-tests; Extended Data Table 1) to be female (50% versus 48%; P = 0.15), older (51.6 versus 50.1 years; P < 0.001) and white (82% versus 77%, P < 0.001); to have a college education (63% versus 56%; P < 0.001); to use X more frequently (5.04 versus 4.85 on an ordinal frequency scale; P < 0.001) and to use it more for hard news consumption (73% versus 68%, P < 0.001). We verify that the difference in effects between switching the algorithm on and off is not driven by these socio-demographic differences. First, we show that the results are robust to re-weighting observations so that the socio-demographic characteristics of participants initially on the chronological feed match those of participants initially on the algorithmic feed, and vice versa (Extended Data Fig. 5). Second, we predict the probability of initially using the chronological feed based on pre-treatment covariates and show that there is no significant treatment heterogeneity with respect to the predicted initial feed setting, in contrast to the actual feed setting (details of these exercises are provided in Supplementary Information section 2.5). We also verified that the results are not driven by selective attrition by presenting Lee bounds on our estimates (Extended Data Fig. 6 and Supplementary Information section 1.8.2).

Although it is reassuring that the asymmetry between the political effects of turning the algorithm on and off is not explained by selection into the initial feed setting predicted by socio-demographic characteristics, selection on unobserved factors may contribute to this asymmetry. For example, conditional on socio-demographics, users with different traits—such as autonomy or desire for control—could both choose different initial feed settings and react differently to similar content changes.

Alternatively, the asymmetry of the effects could arise directly from an asymmetry in the changes to the content shown when feed settings are altered. In what follows, we examine how changes in feed settings affect the content shown to users and, in turn, how users adjust their online behaviour, while acknowledging that unobserved confounders, such as personality traits, cannot be ruled out.

To study the content, we used data from a Google Chrome extension capturing the first 100 posts each participant would see under both feed settings. We used natural language processing to annotate political leanings and content type (Supplementary Information section 1.5 describes the annotation process and human validation). Figure 3 presents raw summaries of the content shown to users under the two feed settings and reports estimates of the average content difference (\(\Delta \)) from Poisson pseudo maximum likelihood and ordinary least squares regressions with individual and survey-wave fixed effects, together with P values from two-sided t-tests (Supplementary Information section 2.6 presents specifications and robustness to an alternative machine-learning classifier). First, the algorithm promotes more engaging content. On average, posts shown in the algorithmic feed receive 13,212 likes, 1,836 reposts and 722 comments, compared with 2,781 likes, 453 reposts and 145 comments for posts in the chronological feed. The conditional estimates of \(\Delta \) indicate increases of 480%, 408% and 508% in likes, reposts, and comments, respectively (all P < 0.001).

Fig. 3: Content the algorithm promotes.
figure 3

Average content shown to users in each feed setting: grey, chronological feed, red, algorithmic feed; 95% CIs around subsample means (mean ± 1.96 × s.e.) are reported. The unit of observation is X post. The number of observations for each feed setting is n = 268,532 for all variables, except for ‘Conservative’ among ‘Political content’, for which n = 129,790 X posts in each feed setting. All variables are indicator variables except Likes, Reposts and Comments, which are heavily right-skewed. For the algorithmic feed, the minimum, first quartile, median, third quartile and maximum are 0, 174, 933, 8,197 and 1,000,000 for Likes; 0, 9, 137, 1,363 and 316,700 for Reposts and 0, 5, 56, 384 and 247,200 for Comments. For the chronological feed, the minimum, first quartile, median, third quartile and maximum are 0, 4, 40, 342 and 1,700,000 for Likes; 0, 1, 8, 66 and 572,000 for Reposts and 0, 0, 3, 21 and 256,100 for Comments. Corresponding summary statistics are presented in Supplementary Information Table 1.8. On the left, for each outcome, \(\Delta \) (the estimated difference in content between the algorithmic and chronological feed from regressions with user and survey-wave fixed effects) and the corresponding P value are reported. For Likes, Reposts and Comments, \(\Delta \) is expressed in percent and estimated using a Poisson pseudo maximum likelihood model; for all other outcomes, \(\Delta \) is in percentage points and estimated using linear ordinary least squares. Because fixed effects are included, \(\Delta \) does not exactly match the raw ratio or difference in means across feed settings shown in the bar chart. See Supplementary Information section 2.6 for the specification, full regression results and robustness; all tests are two-sided with s.e. values clustered at the respondent level.

Source data

Full size image

Second, the algorithm promotes political content and, within that category, prioritizes conservative content. Regression results yield that posts annotated as conservative are 2.9 percentage points (19.9%) more likely to appear in the algorithmic feed (P < 0.001), whereas posts annotated as liberal are 1.0 percentage point (3.1%) more likely (P = 0.043). Restricting to political posts, conservative content is 2.5 percentage points (8%) more likely in the algorithmic feed (P < 0.001).

Third, the algorithm demotes accounts of traditional news media and promotes those of political activists. Posts from news organizations appear 15.5 percentage points (58.1%) less often in the algorithmic feed (P < 0.001), whereas posts from political activists appear 5.9 percentage points more often (27.4%; P < 0.001). Fourth, posts from entertainment accounts are 9.1 percentage points (21.5%) more likely in the algorithmic feed (P < 0.001).

Supplementary Information Fig. 2.21 summarizes feed content separately for self-declared Democrats and for self-declared Republicans and Independents. As expected, partisans see substantially less cross-partisan than co-partisan content in their feeds. However, the overall share of conservative content among political content is significantly higher under the algorithmic feed than under the chronological feed for both groups (9.3% versus 5.4%, P < 0.001 for Democrats; 54.2% versus 52.4%, P = 0.004 for Republicans and Independents). This pattern does not hold for content generated by political activists: the algorithm promotes co-partisan activist content more strongly for both groups (Supplementary Information section 2.6).

Next, we use data collected from participants’ accounts to analyse the effect of the feed setting on user behaviour. Figure 4 presents the regression results; raw data summaries are presented in Supplementary Information Fig. 2.7. Consistent with the results on political attitudes, switching the algorithm off did not lead to significant differences in the types of account users follow. In contrast, switching the algorithm on did affect the typology of followed accounts among participants initially on the chronological feed. Those exposed to the algorithmic feed as a result of the experiment were more likely to follow conservative accounts (by 0.17 s.d.; 95% CI: 0.03, 0.31; P = 0.015), political activist accounts (by 0.13 s.d.; 95% CI: 0.01, 0.25; P = 0.036) and, especially, conservative political activist accounts (by 0.18 s.d.; 95% CI: 0.05, 0.32; P = 0.010). The entire effect on political activist accounts was driven by conservative political activists. There are no significant differences in the likelihood of following liberal accounts between treatment groups. Following news outlet accounts was also unaffected by the experimental change in the feed setting.

Fig. 4: ITT estimates of feed-setting changes on accounts that users follow.
figure 4

ITT effect estimates of switching the algorithm on and off (in s.d.). Left, effect of moving from the chronological to the algorithmic feed for users initially on the chronological feed. Right, effect of moving in the opposite direction for users initially on the algorithmic feed. For each outcome, the results of two specifications are reported. Blue, unconditional estimates with robust s.e. values, controlling only for the initial feed setting and, where applicable, pre-treatment outcome levels. Orange, conditional estimates, controlling for pre-treatment covariates using GRFs. Point estimates and 90% and 95% CIs are reported. Numerical effect sizes and P values correspond to conditional estimates (all tests are two-sided). The unit of observation is respondent. The sample size is n = 2,387 for all estimates. Tests are described in Methods. Supplementary Information Table 2.16 reports the exact numerical point estimates, s.e. values, CIs and sample sizes for every specification. All outcomes are standardized. Additional results are presented in Supplementary Information section 2.

Source data

Full size image

Finally, we provide suggestive evidence on how changes to the feed setting, and the resulting choices of which accounts to follow, shape the content shown to users. Extended Data Fig. 7 reports ITT estimates for the average feed characteristics of the 599 participants who ran the Chrome extension during the post-treatment survey; these results should be interpreted with caution given the small sample. The share of conservative content rose by 0.35 s.d. (95% CI: 0.06, 0.64; P = 0.018) when the algorithm was switched on, but did not change significantly when it was switched off, whereas liberal content was unaffected by either switch. This pattern reflected shifts in both activist and news posts. Switching the algorithm on increased posts by conservative activists (0.37 s.d.; 95% CI: 0.05, 0.68; P = 0.024) and reduced posts from liberal news outlets (−0.43 s.d., 95% CI: −0.71, −0.16; P = 0.002), whereas turning the algorithm off reduced liberal activists’ posts (−0.23 s.d.; 95% CI: −0.38, −0.08; P = 0.004) and increased posts from both liberal (0.50 s.d.; 95% CI: 0.32, 0.67; P < 0.001) and conservative news outlets (0.28 s.d.; 95% CI: 0.10, 0.46; P = 0.002).

These treatment effects on feed content reflect both what the algorithm promotes (Fig. 3) and changes in the accounts users follow (Fig. 4). Users may be particularly responsive to the latter because these changes arise from their own preferences and behaviour. To isolate the effect of followings, we compare the content users saw, or would have seen, on their chronological feeds depending on treatment assignment in Extended Data Table 2 (methodological details in Supplementary Information section 2.7). Controlling for the intensity of X use before the experiment, the chronological feeds of those exposed to the algorithmic feed setting as a result of the experiment contain 9.0 percentage points (60%) more posts by conservative accounts (95% CI: 2.6, 15.5; P = 0.006) and 6.1 percentage points (28%) more posts by conservative political activists (95% CI: 1.0, 11.3; P = 0.020) compared with the chronological feeds of those participants who were randomized to stay on the chronological feed. Meanwhile, we found no significant effects on these content characteristics of chronological feeds after switching the algorithm off. We also ran a placebo test, verifying that, in the pre-treatment survey before users were exposed to new content or could follow new accounts, there was no relationship between chronological feed content and treatment assignment (Supplementary Information section 2.7).

Discussion

In addition to promoting entertainment, X’s feed algorithm tends to push more conservative content to users’ feeds. Seven weeks of exposure to such content in 2023 shifted users’ political opinions in a more conservative direction, particularly with regard to policy priorities, perceptions of the criminal investigations into Trump and views on the war in Ukraine. The effect is asymmetric: switching the algorithm on influenced political views, but switching it off did not reverse users’ perspectives on policy priorities or current political issues.

We show that exposure to algorithmically curated content led users to follow conservative activist accounts. In contrast, when the algorithmic feed was switched off, users continued to follow the accounts they had engaged with previously. This indicates that exposure to feed algorithms has a lasting impact on users’ feeds and their political attitudes.

This mechanism helps explain why previous research found no political effects of switching the algorithm off on Meta platforms, even though it altered political content1. Once users follow accounts promoted by the algorithm that they find engaging and persuasive—such as those of conservative political activists—they continue to see posts from these accounts after turning the algorithm off. Hence, users’ political attitudes may remain unchanged, as the content they engage with most persists despite observed shifts in overall political content. Earlier work also showed that users’ feeds are shaped largely by the accounts they follow, even under algorithmic curation40. Our findings demonstrate that the algorithm significantly influences whom users choose to follow, indicating that algorithms shape content exposure more than believed previously.

This mechanism may also help to explain why social media deactivation experiments, which completely disconnect users from such content, produce political effects, unlike merely switching off the algorithm while keeping users active on the platform. Facebook deactivation in 2018 reduced political polarization41, whereas deactivation during the 2020 US elections reduced self-reported net votes for Trump, although the effect did not meet the authors’ pre-registered significance threshold42. The latter finding aligns with the political effects observed in our experiment. It is important to note that potential treatment heterogeneity between respondents who initially selected different feed settings may also contribute to the asymmetry in switching the algorithm on and off. Although we control for observed socio-demographics, these groups may still differ along unobserved dimensions.

Additional considerations help clarify the broader implications of our findings. In contrast to policy attitudes and views on current events, partisanship and affective polarization were unaffected by our experimental treatments, consistent with the Meta Study1. This may reflect the rigidity and salience of these measures, as they are less susceptible to change over short periods43. Despite this rigidity, researcher-designed re-ranking of feeds can affect affective polarization37.

Politicized content was a large part of the public conversation on Twitter even before it became X38 and, in 2024, X stood out among social media platforms as a place people turn to for politics44. Further evidence is needed on the effects of algorithm activation on other platforms. Our study has limitations common to research on social media and feed algorithms. The effects are specific to the platform and the time period. For example, they depend on the platform owner’s preferences. We find that, relative to the chronological feed, X’s algorithm promotes conservative content to both Democrats and to Republicans and Independents, with suggestive evidence that persuasion is stronger for Republicans and Independents, who are more positively predisposed to conservative messages. This indicates that, within the context studied, X’s algorithm affects users mainly on one side of the political spectrum—a result that may not hold for algorithms that amplify content from both ideological sides. Another limitation, common to most studies, is that platforms can modify their algorithms at any time, including during the study, without researchers being aware of such changes33,38,45. Finally, because we focus on active X users, effect sizes may not generalize to the broader user base, as they are probably smaller for irregular or inactive users, who experience lower treatment intensity.

We conducted the experiment independently of the platform, which helps address additional external validity concerns35. However, relative to what could be achieved with the platform’s cooperation, our study’s independence came at the cost of relying on self-reported compliance for most of the sample, which limited our ability to measure treatment intensity directly, and of working with a smaller sample, which constrained our ability to test finer-grained treatment heterogeneity. Nevertheless, the sample was sufficiently large to detect significant and robust political effects from switching the algorithm on.

Overall, we conclude that social media feed algorithms can play an important role in shaping political opinions and online behaviour.

Methods

Participants

Participants in our experiment were recruited by YouGov—a professional online polling firm with a large panel of respondents in the United States. All study participants gave their informed consent to participate. Participants were included in the study only if they self-reported using X at least several times a month. The sample is 78% white, 52% male and relatively well educated, with 58% having completed at least 4 years of university. In terms of political affiliation, 46% identify as Democrats and 21% as Republicans. Among the participants, 66% use X at least once a day and 94% at least once a week. As for posting activity, 27% post at least once a day and 53% at least once a week. See Supplementary Information section 1.4 for detailed summary statistics and the comparison of our sample with Twitter users from the nationally representative Social Media Study by the American National Election Studies.

Experimental design

Participants were randomized into an algorithmic or chronological feed for the duration of the experiment in exchange for compensation. Supplementary Information section 1.2 provides information on the compensation scheme and on how different elements of the experiment appeared on the X platform.

The randomization procedure was effective: participants assigned to different feed conditions exhibited no systematic differences in demographic characteristics or social media usage beyond what would be expected by chance. The only notable imbalance relates to the initial feed setting. The proportion of participants who were already using the algorithmic feed at baseline was two percentage points higher among those assigned to the algorithmic group (77% compared with 75%). All analyses account for the initial feed setting. Supplementary Information section 1.7 provides further details on balance and reports summary statistics by treatment group and initial feed setting.

Data on outcomes

Outcome data on attitudes come from the post-treatment survey. The coding of the survey-based outcome variables is presented in Supplementary Information section 1.3. To analyse the content of users’ feeds and the accounts they follow, we applied natural language processing methods. The Llama 3-based classification of feed content, collected using a Google Chrome extension, categorizes posts by political leaning—either conservative or liberal—and by type, distinguishing between posts from political activists, entertainment accounts and news media outlets. Details are provided in Supplementary Information section 1.5. The Llama 3-based classification of followed accounts is based on data collected through users’ X handles. This procedure is described in Supplementary Information section 1.6.

We conduct several validation exercises for the Llama 3 annotations in Supplementary Information sections 1.5 and 1.6. These include comparisons with machine-learning classifiers based on word frequencies, as well as evaluations conducted by human annotators.

ITT effects estimates

Given the experimental design, the ITT effect estimates were obtained by comparing mean outcomes between respondents assigned randomly to the algorithmic feed and those assigned to the chronological feed, conditional on their initial feed setting. Our main specification estimates the effect of switching the algorithm on for users who initially had it off, and switching it off for users who initially had it on. Specifically, we estimate the following model:

$$\begin{array}{l}{Y}_{i}=\alpha +{\beta }_{1}{\rm{I}}{\rm{n}}{\rm{i}}{\rm{t}}{\rm{i}}{\rm{a}}{\rm{l}}\,{{\rm{C}}{\rm{h}}{\rm{r}}{\rm{o}}{\rm{n}}{\rm{o}}}_{i}\times {\rm{T}}{\rm{r}}{\rm{e}}{\rm{a}}{\rm{t}}{\rm{m}}{\rm{e}}{\rm{n}}{\rm{t}}\,{{\rm{A}}{\rm{l}}{\rm{g}}{\rm{o}}}_{i}+{\beta }_{2}{\rm{I}}{\rm{n}}{\rm{i}}{\rm{t}}{\rm{i}}{\rm{a}}{\rm{l}}\,{{\rm{A}}{\rm{l}}{\rm{g}}{\rm{o}}}_{i}\\ \,\times {\rm{T}}{\rm{r}}{\rm{e}}{\rm{a}}{\rm{t}}{\rm{m}}{\rm{e}}{\rm{n}}{\rm{t}}\,{{\rm{C}}{\rm{h}}{\rm{r}}{\rm{o}}{\rm{n}}{\rm{o}}}_{i}+{\beta }_{3}{\rm{I}}{\rm{n}}{\rm{i}}{\rm{t}}{\rm{i}}{\rm{a}}{\rm{l}}\,{{\rm{A}}{\rm{l}}{\rm{g}}{\rm{o}}}_{i}+{{\bf{X}}}_{i}^{{\prime} }{\boldsymbol{\Gamma }}+{\varepsilon }_{i}\end{array}$$
(1)

where i indexes respondents. Yi represents outcome variables (for example, post-treatment political attitudes of respondents). Initial Algoi is a dummy variable for the initial feed setting; it equals one for respondents who had been using the algorithmic feed before our intervention and zero for those who had been using the chronological feed. Treatment Algoi is a dummy variable for being assigned to the algorithmic feed setting during the treatment phase; it equals 1 for respondents assigned to the algorithmic feed and 0 for those assigned to the chronological feed. Note that, by construction, Initial Chronoi = 1 − Initial Algoi and Treatment Chronoi = 1 − Treatment Algoi. Xi is a vector of control variables discussed below.

The coefficients of interest are β1 and β2, which estimate the ITT effects of switching the algorithm on and switching it off, respectively. Because we control for Initial Algo, those switching from the chronological to the algorithmic feed are compared with users who initially had the chronological feed and remained on it, whereas those switching from the algorithmic to the chronological feed are compared with users who initially had the algorithmic feed and remained on it.

We consistently report two sets of estimates. First, we report unconditional estimates, namely, controlling only for the initial feed setting and, when available, the pre-treatment value of the outcome variable. The unconditional model is justified by the balance in observables across treatment groups. Second, we also report estimates that flexibly control for pre-treatment covariates. Our baseline approach relies on GRFs—a machine-learning method that identifies the most relevant covariates and adjusts for them non-parametrically39. This approach limits the researcher’s discretion in selecting covariates and functional form, including mitigating concerns about model misspecification46. We use the full set of available pre-treatment covariates as input to the GRFs. The covariates include initial feed setting, gender, age, indicators of X use, all categories of educational attainment and race, frequency of X use and posting, political affiliation, the device used to complete the pre-treatment survey (such as a laptop or smartphone), life satisfaction, happiness and affective polarization. Technical details are provided in Supplementary Information section 2.2.1. Baseline ITT estimates are reported in Fig. 2.

Throughout the analysis (in the main text and Extended Data Figs. 4, 5 and 7 and Supplementary Information), we present ITT estimates in graphical form using coefficient plots. Estimates from the unconditional specification are shown in blue, whereas those from the specification controlling for pre-treatment covariates are shown in orange. For clarity, we refer to β1 estimates as ‘Chrono to Algorithm’ and β2 estimates as ‘Algorithm to Chrono’ in these coefficient plot figures.

In Supplementary Information section 2.4, we report results for all outcome variables using a range of linear specifications, progressively adding the following controls to equation (1): demographic characteristics (gender, age, indicators for being white and for high educational attainment), pre-treatment X use (frequency of use, frequency of posting and self-reported purpose of using X), and political affiliation (Republican, Independent, Democrat or other). We also present results from a specification that includes the predicted probability of having an algorithmic initial feed, obtained using LASSO on all pre-treatment characteristics. In addition, we report results from a specification that controls for post-treatment X use. Although post-treatment use is endogenous, robustness to including this control indicates that the estimated effects are not solely driven by changes in X usage.

In Supplementary Information section 2.4, we also present tests for the symmetry of the effects of switching the algorithm on and off. Specifically, we test whether β1 = − β2, and whether 7 weeks of algorithm exposure, induced by treatment assignment for users initially on the chronological feed, leads to outcome levels similar to those of users who were initially on the algorithmic feed and remained on it, that is, whether β1 = β3.

Compliance and LATE estimates

Due to minimal non-compliance with assigned treatments (Extended Data Fig. 2 and Supplementary Information section 1.8.1), the ITT estimates do not strictly represent average treatment effects. We estimate LATE for compliers, that is, for respondents who adhered to their assigned feed setting, using instrumental variables regressions, with the assigned feed setting used as an instrument for the feed setting actually used.

We define ‘actual’ feed use during the treatment phase using treatment assignments and respondents’ self-reported compliance from the post-treatment survey. Respondents are classified as compliant if they answered ‘always’ or ‘most of the time’ to the question, ‘Did you stick to the assigned feed?’ Based on this, we construct two variables: Actual Algoi and Actual Chronoi.

Actual Algoi equals 1 if respondent i was compliant and assigned to the algorithmic feed, or non-compliant and assigned to the chronological feed, implying they used the algorithmic feed. Conversely, Actual Chronoi equals 1 if the respondent was compliant and assigned to the chronological feed, or non-compliant and assigned to the algorithmic feed. By construction, Actual Chronoi = 1 − Actual Algoi.

We estimate two-stage least squares regressions, with the second stage specified as follows:

$$\begin{array}{c}{Y}_{i}=\alpha +{\gamma }_{1}{\rm{H}}{\rm{a}}{\rm{t}}({\rm{I}}{\rm{n}}{\rm{i}}{\rm{t}}{\rm{i}}{\rm{a}}{\rm{l}}\,{{\rm{C}}{\rm{h}}{\rm{r}}{\rm{o}}{\rm{n}}{\rm{o}}}_{i}\times {\rm{A}}{\rm{c}}{\rm{t}}{\rm{u}}{\rm{a}}{\rm{l}}\,{{\rm{A}}{\rm{l}}{\rm{g}}{\rm{o}}}_{i})\\ \,+\,{\gamma }_{2}{\rm{H}}{\rm{a}}{\rm{t}}({\rm{I}}{\rm{n}}{\rm{i}}{\rm{t}}{\rm{i}}{\rm{a}}{\rm{l}}\,{{\rm{A}}{\rm{l}}{\rm{g}}{\rm{o}}}_{i}\times {\rm{A}}{\rm{c}}{\rm{t}}{\rm{u}}{\rm{a}}{\rm{l}}\,{{\rm{C}}{\rm{h}}{\rm{r}}{\rm{o}}{\rm{n}}{\rm{o}}}_{i})\\ \,+\,{\gamma }_{3}{\rm{I}}{\rm{n}}{\rm{i}}{\rm{t}}{\rm{i}}{\rm{a}}{\rm{l}}\,{{\rm{A}}{\rm{l}}{\rm{g}}{\rm{o}}}_{i}+{{\bf{X}}}_{{i}}^{{\prime} }{\boldsymbol{\Gamma }}+{\varepsilon }_{i}\end{array}$$
(2)

Hat() denotes the fitted value from the corresponding first-stage regression. The two first-stage regressions predict the following:

Hat(Initial Chronoi × Actual Algoi) with Initial Chronoi × Treatment Algoi and Hat(Initial Algoi × Actual Chronoi) with Initial Algoi × Treatment Chronoi. With a compliance rate of 85.38%, the instruments are extremely strong: in all specifications, the F-statistics for the excluded instruments in the first stage exceed 1,000.

In equation (2), γ1 represents the LATE estimate of the effect of switching the algorithm on, and γ2 the corresponding effect of switching it off. The main LATE estimates are reported in Extended Data Fig. 3. Given the high compliance rate, the ITT and LATE estimates are very similar.

In Supplementary Information Fig. 2.13, we also report LATE estimates using a more conservative definition of self-reported compliance, in which a respondent is considered compliant only if they answered ‘always’ to the question ‘Did you stick to the assigned feed?’ For those who answered ‘almost always’, we assume that they complied with the assigned setting at any moment only with probability 0.5. In this case, the instruments remain very strong and the results are very similar.

The results are also robust to restricting the analysis to respondents whose compliance was confirmed through observation, that is, those who installed the Google Chrome extension and remained on their assigned feed setting (Supplementary Fig. 2.12).

Attrition and Lee bounds estimates

Significant attrition was observed between the pre-treatment and post-treatment surveys, common to studies like ours. Attrition rates did not differ across treatment groups. We formally assess whether selective attrition could bias our results using Lee bounds and conclude that attrition does not meaningfully affect our findings. Details are provided in Supplementary Information section 1.8.2, and results are reported in Extended Data Fig. 6.

Ethics statement

Ethical approval for this research was granted by the Ethics Committee of the University of St. Gallen, Switzerland. All procedures involving human participants were performed in accordance with the institutional guidelines and regulations. Informed consent was obtained from all participants at the time of their enrolment. Participants received compensation for their time: they were compensated in points—a currency used within our implementation partner’s (YouGov) platform, with an exchange rate of 1,000 points equalling US$1.00. For the pre-treatment survey, they received 500 points (US$0.50), with clear information that they would earn 2,500 points (US$2.50) for using the assigned feed setting between the two surveys and returning for the post-treatment survey. Furthermore, participants could earn an extra 2,000 points (US$2.00) for using the Chrome extension during the pre-treatment survey. After completing the post-treatment survey, they were paid the 2,500 points (US$2.50) that had been announced during the pre-treatment survey. Again, they could earn additional compensation for running the Chrome extension, this time 2,500 points (US$2.50). Therefore, participants could earn up to 10,000 points (US$10.00) if they completed both surveys and used the Chrome extension both pre- and post-treatment. Sharing the X handle was not incentivized.

The experimental interventions were limited to the choice of feed setting, which is freely available to every user on the X platform. No artificial manipulation of users’ content was introduced. Additional information on ethical safeguards and considerations is provided in Supplementary Information section 1.1.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.