The X Algorithm’s Challenge in Türkiye: Politics, Engagement, and Disinformation

F

ollowing Elon Musk’s acquisition of Twitter in late 2022 and its subsequent rebranding as “X,” expectations emerged that the platform would undergo fundamental changes in the way it delivers content. The algorithmic “For You” feed, which largely replaced the chronological timeline, became the default main feed presented to users. That algorithm was introduced with the stated aim of ranking posts that users might find engaging regardless of whom they follow. However, over time, intense debates arose, particularly regarding political content in Türkiye, over whether the algorithm operates in a neutral and genuinely personalized manner.

In 2023, several journalists and researchers, together with Daily Sabah, argued that X’s recommendation algorithm was manipulating the content flow in Türkiye. According to those claims, X disproportionately amplified opposition-leaning content regardless of users’ interests or political preferences, to the extent that even pro-government users were consistently exposed to opposition posts in their feeds. In response to those criticisms, Elon Musk acknowledged that the algorithm had shortcomings and pledged to open the code for the sake of transparency and to implement improvements.

So what has changed since then? Does X’s content recommendation algorithm in Türkiye genuinely operate with political bias? And how does this affect user experience and public discourse? To address these questions, I conducted a series of experiments toward the end of 2025. From the final days of November 2025 through early December 2025, I closely examined X’s “For You” feed using a total of eight accounts that were created from scratch on entirely new devices. Those accounts were configured with different locations and distinct political orientations. The findings obtained indicate that X’s recommendation algorithm exhibits serious problems with both technical and societal dimensions.

The “for you” feed: Personalized or a uniform digital enforcer?

In our experiments, the situation Ix encountered when logging into X with zero-profile accounts (no followers, likes or prior activity) offered a striking illustration of this issue. Despite the complete absence of user data, the algorithm immediately presented political content and accounts belonging to political actors. Approximately 90 percent of the initially recommended accounts to follow were users known for opposition-oriented views. Of the remaining 10 percent, roughly half consisted of official accounts such as the Presidency and ministries, while the other half included one or two pro-government accounts.

Unless users explicitly switched to the “Following” tab, without having expressed any interests, new accounts were directly exposed to popular tweets featuring government criticism and accounts widely recognized for their oppositional stance in Türkiye. This pattern suggests that the algorithm prioritizes currently trending and sensational political debates on the platform rather than attempting to assess user preferences through gradual interaction signals.

Within the scope of the experiment, four of the accounts I created followed only pro-government politicians, journalists, and affiliated accounts. The remaining four accounts followed exclusively figures and accounts aligned with the opposition. Despite this deliberate differentiation, all eight profiles were predominantly presented with content featuring the rhetoric of opposition politicians, government criticism, and posts from popular opposition-oriented accounts, as well as misleading or false news that framed the political authorities in a negative manner. Pro-government profiles were rarely exposed in their feeds to posts from the accounts they explicitly followed. In some cases, even when a journalist’s original tweet from an account that the user followed was not shown in the feed, the user was nonetheless exposed to an opposition-oriented post because that same journalist had replied to it, thereby amplifying the visibility of the oppositional content.

Conversely, the profiles designed with an opposition-oriented stance were exposed to a similar feed structure. While posts from the opposition accounts they followed were consistently displayed, those were frequently supplemented by content from additional opposition-aligned accounts that the users did not recognize or follow. Notably, those opposition-oriented profiles were almost never recommended pro-government content. As a result, in both scenarios, the algorithm effectively enclosed users within a bubble dominated by opposition discourse.

That situation points to two possible explanations. The first is that the algorithm attempts to introduce viewpoint diversity by presenting opposing content but fails to maintain a meaningful balance. The second is the presence of a structural bias embedded within the system’s underlying architecture. Although X engineers argue that the recommendation system operates not on the basis of political labels but according to interaction networks and aggregate engagement metrics, the higher levels of engagement typically generated by opposition content in Türkiye inadvertently transform the algorithm into a one-directional echo machine.

The misinterpretation of comment-based engagement: the algorithm’s affective blindness

X’s algorithm interprets all forms of user interaction including likes, reposts, comments, and even views; as signals of interest. However, it does not differentiate between positive and negative forms of engagement. Following changes implemented after 2023, the tendency to treat all types of interaction as inherently positive has become even more pronounced. During the experiments, it was observed that when a pro-government profile commented on an opposition tweet with a remark such as “This information is false,” the algorithm did not reduce the visibility of similar content; on the contrary, it increased it. Even the act of commenting alone led to a higher volume of recommendations related to the same topic. Likewise, merely clicking on a controversial post or viewing that briefly was sufficient to cause similar content to proliferate within the feed.

At the core of those outcomes lies X’s failure to analyze the semantic content of posts and the qualitative nature of user engagement. This form of affective blindness causes undesirable content to cascade onto users’ feeds. In particular, posts that provoke anger or political polarization tend to generate large volumes of comments and quote reposts. The system interprets thas pattern as “high engagement equals high relevance” and consequently rewards such content.

Misinformation and algorithmic amplification

Perhaps the most alarming aspect of X’s recommendation algorithm is its capacity to enable the rapid and large-scale dissemination of posts containing misinformation. The claim examined in 2025 that “Türkiye donated 7,000 police vehicles to Syria” constitutes a striking example. Although this post lacked any official basis, it received millions of views, while subsequent corrections circulated only minimally. Whereas the false content exceeded 7 million views, the corrective post reached only 211,000 views. Moreover, when accounts that continued to amplify the false claim through quote reposts are taken into account, the total number of views was estimated to surpass 10 million.

A similar pattern emerged in a post shared by a politician affiliated with the İYİ (Good) Party. The user presented footage from a Christian religious ceremony as if it had been recorded in Türkiye’s İznik district, accompanying the video with a provocative message asking, “Are we christianized?” In reality, the video was neither recent nor filmed in Türkiye. It was older footage recorded in Spain, a fact that was uncovered shortly thereafter. Nevertheless, many users reacted with anger and outrage, initially accepting the fabricated “scandal” as genuine and responding emotionally before the correction became widely known.

These examples demonstrate how algorithmic amplification can pose serious risks to the information ecosystem. In psychology, the phenomenon known as the “truth effect” refers to the tendency for false information to be perceived as true through repeated exposure. X’s recommendation algorithm appears to intensify precisely this mechanism by repeatedly presenting the same false narratives to users, effectively imprinting them onto their cognitive frameworks. Users with low levels of media literacy or those unable to critically assess information to which they are repeatedly exposed are particularly susceptible to such waves of misinformation. In this way, “vibes” initially shape emotions and subsequently translate into behavior as an outcome of identity formation.