Motivation

In late March and early April 2026, X, formerly Twitter, reportedly began expanding Grok-powered automatic translation in users’ timelines and For You feeds. The public reporting is not fully uniform: one report described an April 7, 2026 worldwide rollout, while another described a staged English-Japanese rollout beginning in late March, with later expansion beyond the United States still pending (The Tech Outlook 2026; BigGo Finance 2026). Product Head Nikita Bier described the feature as “the largest cultural exchange in history” (The Tech Outlook 2026; BigGo Finance 2026). Whether that framing is accurate or hyperbolic, the structural change is potentially important: a major social platform is reducing language as a filter on political information flows, not by training users to read foreign languages, but by rendering foreign-language content directly inside ordinary feeds.

This matters for political communication research because the existing literature on foreign media exposure often rests on a selection assumption. Studies linking exposure to foreign media or foreign cultural content with attitudes toward foreign countries have typically examined users who already consumed foreign-language or foreign-origin media, and therefore possessed some combination of access, motivation, cultural interest, or linguistic capacity (Chen 2025; Park 2005). Automatic translation weakens this selection mechanism. A Korean user who would never actively click on an English-language or Japanese-language political post may encounter a translated version of that post inside an ordinary feed, without a deliberate search act.

Compounding this structural change is empirical evidence on X’s algorithmic architecture. A six-week audit of X’s For You timelines during the 2024 U.S. presidential election found that approximately 50% of tweets in user timelines were personalized recommendations from accounts users did not follow, and that new accounts experienced a right-leaning exposure bias in default timelines (Ye, Luceri, and Ferrara 2025). Earlier work on Twitter’s ranking algorithm found that the political right enjoyed higher algorithmic amplification than the political left in six of seven countries studied, and that the U.S. media landscape showed amplification favoring right-leaning news sources (Huszár et al. 2022). More directly, a 2023 field experiment with 4,965 U.S.-based X users found that switching from a chronological feed to the algorithmic feed increased engagement and shifted some political attitudes in a more conservative direction, particularly on policy priorities, criminal investigations into Donald Trump, and views on the war in Ukraine (Gauthier et al. 2026).

If the algorithm that selects which foreign posts are surfaced to non-English speakers follows the same engagement-optimization logic, the automatic translation layer may not produce culturally neutral cross-lingual exchange. It may instead become a conduit through which algorithmically privileged political content, including U.S. right-leaning content, is broadcast in translated form to users who were previously insulated by language barriers.

The research question this project addresses is: Does X’s automatic translation feature increase non-English-speaking users’ incidental exposure to U.S. conservative political content, and does this exposure shift attitudes toward Donald Trump and the Republican Party’s America First agenda?

Core Argument

X’s automatic translation rollout constitutes a potential shock to the information environment of non-English-speaking users. However, because public reporting on the rollout sequence is inconsistent, the design should not assume a clean global rollout unless availability timestamps can be reconstructed. The stronger design is therefore to treat rollout timing as an empirical object: first map when automatic translation became visible to users in different language pairs and regions, then use that variation for comparison.

The core argument is twofold.

First, translated political content delivered through X’s algorithmic feed will be ideologically non-neutral. Because a substantial portion of For You timeline content comes from accounts users do not follow (Ye, Luceri, and Ferrara 2025), and because prior studies find that X/Twitter’s algorithmic ranking can amplify right-leaning political content (Huszár et al. 2022; Gauthier et al. 2026), non-English users may receive a translated information diet that overrepresents America First messaging, Republican politicians, and right-leaning commentators relative to the baseline distribution of English-language political opinion.

Second, this algorithmically skewed translated exposure may shift non-English users’ attitudes toward U.S. conservative politics, moderated by their prior attitudes toward the United States and their awareness that the content they are reading has been algorithmically selected and machine-translated. Users with neutral or weakly held priors should be more susceptible to attitude change than users with crystallized prior beliefs, consistent with the elaboration likelihood model of persuasion (Petty and Cacioppo 1986). Users who notice the translation label or understand the algorithmic origin of the content may also discount the message more than users who process the content as ordinary native-language feed material. This expectation is consistent with experimental evidence that perceptions of foreign news credibility are shaped by whether the content fits the dominant narrative in the user’s home information environment (Bryanov et al. 2023).

Theoretical Grounding

Incidental exposure and attitude formation. The political communication literature distinguishes between deliberate and incidental news exposure. Deliberate exposure occurs when a user actively seeks out political information. Incidental exposure occurs when political information is encountered while the user is doing something else, such as browsing the web or scrolling a social feed (Tewksbury, Weaver, and Maddex 2001; Fletcher and Nielsen 2018; Schäfer 2023). Automatic translation represents a structurally novel form of incidental exposure: the user does not merely encounter foreign political content accidentally; the user encounters it in a domesticated linguistic form, with some of the ordinary cues of foreignness removed.

Parasocial contact and outgroup attitude change. The parasocial contact hypothesis holds that exposure to outgroup members through media can produce attitude change toward the outgroup through mechanisms analogous to direct intergroup contact (Schiappa, Gregg, and Hewes 2005). Applied to the political context, exposure to U.S. conservative politicians and commentators through translated content could produce either attitude warmth, if the content is consumed without strong counter-messaging, or attitude hardening, if it activates negative priors. The direction of the effect should depend on the emotional valence of the encountered content and the strength of the user’s prior attitude (Petty and Cacioppo 1986).

Algorithmic amplification and ideological bias. A six-week algorithmic audit of X’s For You timeline during the 2024 U.S. presidential election, using 120 sock-puppet monitoring accounts, found that X’s algorithm skewed exposure toward a small set of high-popularity accounts, that both left- and right-leaning users received amplified exposure to ideologically congruent accounts, and that new accounts experienced a right-leaning bias in default timeline exposure (Ye, Luceri, and Ferrara 2025). Earlier Twitter research found that the political right received higher algorithmic amplification than the political left in six of seven countries, and that U.S. right-leaning news sources were algorithmically favored (Huszár et al. 2022). A later field experiment found that switching from X’s chronological feed to its algorithmic feed shifted several political attitudes in a conservative direction among active U.S. users (Gauthier et al. 2026). These findings establish the premise: if the algorithm amplifies conservative content for English-speaking users, automatic translation may extend the reach of that amplification to users who were previously linguistically insulated from it.

Foreign news credibility and prior attitude moderation. Prior research on credibility perceptions of foreign news finds that perceived credibility depends not only on factual accuracy but also on concordance with the dominant media narrative in the user’s home country (Bryanov et al. 2023). This implies that attitude effects from translated X content should vary across countries. In countries where the prevailing domestic media frame is antagonistic toward Trump, algorithmically amplified pro-Trump translated content may be discounted. Pew’s 2025 24-country survey found that a median of 34% had confidence in Trump to do the right thing regarding world affairs, with high confidence in Nigeria, Israel, Kenya, Hungary, and India, and especially low confidence in Mexico, Sweden, Turkey, Germany, and Spain (Pew Research Center 2025). The implication is not that attitude effects will be absent, but that they should be strongest among users whose prior attitudes are neutral, weak, or cross-pressured.

Methodology

The project combines two studies with distinct identification strategies: an observational audit of translated political content and a survey experiment examining the mechanism through which translated political content affects attitudes.

Study 1: Algorithmic Audit of Translated Political Content

Research question: Does X’s automatic translation layer deliver politically non-neutral content to non-English users, and in which ideological direction?

Design: Sock-puppet account audit. A set of monolingual accounts, such as Korean, Japanese, Portuguese, Arabic, and German accounts, are created with neutral following patterns, for example 3-5 non-political accounts per language. These accounts passively collect For You content before and after translation availability is observed for each language pair. All translated political content appearing in the feed is collected, including source account, post text, displayed language, original language, translation label visibility, and whether the account is followed or out-of-network.

Rollout measurement: Because reporting on the rollout sequence is inconsistent, the first empirical task is to reconstruct rollout availability. The study records when automatic translation first appears for each language pair and account type. The design then compares language pairs and account regions that receive automatic translation earlier versus later, rather than assuming a clean universal rollout date (The Tech Outlook 2026; BigGo Finance 2026).

Ideological classification: Source accounts are classified by political orientation using a combination of: (1) pre-existing English-language media-bias indices, such as AllSides or Ad Fontes, for institutional media accounts; (2) known party affiliation for politicians and campaign accounts; and (3) LLM-assisted stance classification applied to post text. Classification is validated against a hand-coded gold standard subset, for example N ≈ 500 posts per language pair.

Primary outcome: The ratio of right-leaning to left-leaning political content in translated posts surfaced to neutral non-English accounts, compared against the baseline distribution of political content in English. A ratio significantly above 1.0 would suggest that automatic translation does not merely expose non-English users to the full distribution of English political content, but systematically privileges one ideological register.

Feasibility note: X’s API remains a constraint. The old Academic Research access route should not be assumed. Current X API documentation describes pay-per-use pricing, with resources such as post reads charged per resource fetched (X API Docs 2026). Data collection should therefore combine API-based collection, browser-extension-based passive collection by consenting volunteer accounts, and archived or third-party datasets where available.

Study 2: Survey Experiment, Effect of Translated Conservative Content on Attitudes Toward Trump and America First

Research question: Does incidental exposure to algorithmically typical translated political content from X shift non-English speakers’ attitudes toward Donald Trump and the America First agenda, and is this effect moderated by prior attitude strength and awareness of translation origin?

Participants: N ≈ 800-1,000 Korean adults recruited via an online panel vendor, such as Kantar, Macromill Embrain, or another Korean survey panel, stratified by age, gender, and prior attitude toward the United States. Korean participants are appropriate because South Korea is included in Pew’s 2025 cross-national survey on confidence in Trump, and because Korean public opinion displays meaningful variation by ideology and party support (Pew Research Center 2025).

Experimental design: 2 × 2 between-subjects factorial.

Factor Levels
Content type Pro-Trump/America First content vs. neutral international news
Translation label salience Translation label prominently displayed vs. minimized

Participants in all conditions view a simulated X feed interface containing eight posts. Treatment conditions receive four posts drawn from the audit-identified universe of algorithmically typical translated conservative political content. Control conditions receive four posts of matched neutral international news content. In the high-salience condition, each post is marked with a visible “Translated from English” label and a note that content is algorithmically selected. In the low-salience condition, the translation label is present but visually minimized, approximating ordinary feed exposure.

Dependent variables:

  • Primary: Thermometer rating of Donald Trump on a 0-100 scale; support for key America First policy positions, including trade tariffs, NATO burden-sharing, immigration restriction, and skepticism toward multilateral commitments, measured before and after feed exposure.
  • Secondary: Perceived credibility of X as an information source; intention to follow translated foreign political accounts; self-reported awareness of algorithmic curation and machine translation.

Moderator: Pre-treatment prior attitude toward the United States and toward Trump, measured at baseline using items adapted from Pew’s Global Attitudes Survey (Pew Research Center 2025). This operationalizes the theorized conditional effect: attitude updating should be largest among neutrals and weakest among strong priors (Petty and Cacioppo 1986).

Estimation: Mixed ANOVA with pre-treatment attitude as a continuous moderator; moderated mediation analysis in which perceived algorithmic awareness is tested as a mediator of the label-salience effect on attitude change. Because PROCESS is software-specific rather than a design requirement, the analysis can also be implemented as an interaction model or structural equation model.

Key Hypotheses

Hypothesis Prediction
H1, content audit Translated political content surfaced to neutral non-English accounts on X will overrepresent U.S. right-leaning sources relative to the baseline distribution of English political content
H2, attitude change Exposure to algorithmically typical translated conservative content will produce a positive shift in thermometer ratings of Trump and support for America First policies, relative to the neutral content control
H3, label moderation The attitude shift in H2 will be attenuated when the translation label and algorithmic-selection cue are prominently displayed
H4, prior attitude moderation The attitude shift in H2 will be largest among participants with neutral or uncertain prior attitudes toward the United States and weakest among participants with strong negative priors

Expected Contributions

Theoretical contribution. The project introduces the concept of algorithmically mediated incidental cross-lingual exposure as a distinct theoretical category. It is separable from deliberate foreign media consumption, which operates on a self-selected population with access and motivation (Chen 2025; Park 2005), and from domestic incidental exposure, which does not necessarily cross linguistic or national boundaries (Tewksbury, Weaver, and Maddex 2001; Fletcher and Nielsen 2018). The key theoretical claim is that automatic translation does not simply lower a friction cost; it changes the population reachable by foreign political content and changes the psychological conditions under which that content is processed.

Empirical contribution. Existing research on X’s algorithmic amplification of conservative or right-leaning content is concentrated on English-language or U.S.-centered exposure environments (Huszár et al. 2022; Ye, Luceri, and Ferrara 2025; Gauthier et al. 2026). This project provides a direct empirical test of whether such amplification extends to non-English populations through the translation layer, and a causal estimate, via survey experiment, of its attitude effects outside the United States.

Policy contribution. The findings bear directly on the governance of AI-powered translation on social media platforms. If automated translation functions as an ideological amplifier rather than a neutral conduit, not necessarily by design but as a byproduct of engagement-optimization logic applied to content selection, platform operators and regulators face a novel accountability question: to whom is the algorithm accountable when its effects on political opinion cross national borders?

Key Limitations

Content selection and algorithm opacity. X does not publicly disclose the criteria by which content is selected for automatic translation or surfaced to non-native-language users. The audit design infers these criteria from behavioral observation rather than technical access, and may fail to capture time-varying algorithmic updates. This limitation is shared by prior platform audits (Ye, Luceri, and Ferrara 2025).

Rollout ambiguity. Public reporting on X’s automatic translation rollout is inconsistent. Some reports describe worldwide rollout, while others describe a phased English-Japanese launch with later expansion (The Tech Outlook 2026; BigGo Finance 2026). The study therefore requires independent reconstruction of feature availability by account, language pair, and region.

Sample constraints. Restricting Study 2 to Korean participants improves internal coherence but limits generalizability. The direction and magnitude of effects may differ substantially in countries with higher baseline confidence in Trump, such as Nigeria or India, compared with countries with strong negative priors, such as Sweden or Germany (Pew Research Center 2025).

Short-term attitude measurement. The survey experiment measures attitude change immediately after exposure. Whether these short-term shifts translate into durable attitude change or behavioral consequences, such as following, reposting, or cross-lingual engagement, remains untested. Prior work on social media exposure suggests that effects can be context-dependent and may polarize rather than persuade in a simple directional sense (Bail et al. 2018).

Demand characteristics. Participants viewing a simulated X feed in a survey context may apply different epistemic standards than users casually scrolling their actual feed. Ecological validity is therefore a standing limitation of laboratory-analog designs in political communication research. This can be partially mitigated by using real audit-derived posts, realistic feed interfaces, and post-treatment manipulation checks.

API access. Study 1 depends on continued data access. X’s current API model is pay-per-use, and current pricing documentation emphasizes per-resource charges rather than stable academic access (X API Docs 2026). Project planning must include contingency protocols for reduced access, including browser-extension collection, participant data donations, and archived datasets.

Updated: