In my last post, I pondered whether tech companies are deliberately steering society to the right. My conclusion was that this has less to do with direct influence and more to do with how the technologies—specifically the large LLM language models—are structured. I’d like to elaborate on this a bit further here.   In general, this is not a direct ideological strategy pursued by the major tech companies. Rather, it is a complex process that intertwines strategic content design, sociocultural appropriation, and the technological functioning of algorithmic recommendation systems.

YouTube Songs: Right-Wing Music

But even far-right music today is no longer simply characterized by a form of hard punk; rather, it is evolving into a blend of conventional pop and a right-wing sense of national identity. The producers of such music operate in a sphere that initially tends to conceal their political aims. It doesn’t necessarily come across as far-right hate propaganda, but rather as songs that focus on traditional values such as home and one’s own homeland.

Many songs appear apolitical or “patriotic,” but then draw on codes and themes from the far-right scene. This is also evident in the hashtags suggested on YouTube. These include terms like “homeland,” “Schlager,” “Schlager music,” “Germany,” “satire,” “party Schlager,” and “folk music.”   Among them are songs that might initially be interpreted as general social commentary. For example, in a Schlager song about freedom of speech, the lyrics go: “People, freedom of speech is in the Basic Law, but just so you know, if you say something wrong, the police will come.” And then it continues: “Shut up—everything is meticulously monitored, while the country is falling apart at the seams.” It’s only when you listen closely that you realize how everything is geared toward agitation for the AfD: “Billions for the world have to go,” and when it comes to banning the AfD, there’s applause. Almost any current political issue can be turned into a rhyme with the help of chatbots—gas, pensions, sugar tax—to release a new pop song every day. For example, there’s a song about the “health insurance deficit.” It meticulously lists everything that’s supposed to be cut. But you don’t hear a single word about how the AfD would pay for it.   On the one hand, topics like love of one’s homeland, nature conservation, or discontent with the current government are exploited to convey nationalist ideologies to the audience through simple pop rhythms. At the same time, however, there are also more and more songs that label Scholz or Merz as “liars,” essentially launching a direct attack.

The Function of Algorithms

However, the hashtags that creators use to tag their videos on YouTube now play a crucial role, as they are used to generate recommendation lists. This is because YouTube isn’t simply an archive for all the songs uploaded there; rather, it acts as an amplifier for certain music that is supported by high user engagement within its recommendation algorithm. So whoever gets more exposure on YouTube also receives more recommendations from other users. For example, the system detects overlaps in user behavior. If, for instance, many users recommend traditional pop hits or watch right-wing political commentary, the algorithm will link these categories as recommendations from similar users. But this also means that users with right-wing leanings are increasingly likely to receive more and more extreme content over time. In this way, “bubbles” form in which such like-minded songs circulate.   This does not mean that YouTube’s operators are sitting idly by. They are addressing this problem with their “4Rs Framework” (Remove, Raise, Reward, Reduce). In many borderline cases, however, the algorithm reacts uncertainly and very often allows these recommendations to go through.

Large Language Models (LLMs) and Their Focus

But what about the background of large language models? Can we rule out the possibility that they are driven by their own political agendas—ones that shape the data they are trained on or the algorithms that define them? Researchers took an interest in this issue early on. The main finding is that we cannot assume that today’s large LLM models actively promote any particular ideology. First and foremost, we must assume that these language models have processed millions of texts in their training, and that the underlying political attitudes were incorporated into the results. Nearly all models therefore tend toward a center-left position or a more liberal-progressive worldview. LLMs are trained on vast amounts of human text; as a result, they also absorb the ideological structures contained within it. This does not necessarily have to pertain to political ideologies. Language itself can also lead to differences. In one of the studies, answering questions in French resulted in a clear shift toward the right-wing authoritarian spectrum.   But of course, Chinese models can only be partially compared with those from the U.S. and Europe: Chinese models, for example, express much less concern about climate change, and they fall silent when the issue of Taiwan or criticism of the political leadership is raised. According to recent studies by the University of Washington and the Technical University of Munich, a comparison of the largest models yields the following results:

• ChatGPT is often described as the most left-leaning model. It places great emphasis on social justice, environmental policy, and liberal social models.
• Claude from Anthropic, on the other hand, is often considered the most centrist model within the liberal group. It takes a moderate liberal approach with a focus on economic openness.
• Google’s Gemini positions itself as moderately liberal with a focus on social freedom. However, Google can sometimes overreact by not answering a question because it considers it too political.   Elon Musk’s Grok, however, represents a special case, as it was deliberately designed to be “anti-woke.” Grok employs a bipolar algorithmic model. While it takes left-leaning positions on economic issues, it tends to emphasize right-leaning viewpoints on social topics such as immigration or gender issues. Grok also tends to give a lot of prominence to extreme positions while downplaying centrist views. The example of Grok in particular suggests that tech companies still have room to align themselves more closely with certain ideological positions in the future.

The example of Grok shows that the political positioning of AI could be a competitive advantage in the future. For us as users, this means we must not mistake the AI’s answers for “objective knowledge.” They are the product of algorithms and the values of the programmers who feed and weight them. In the age of AI, knowledge is always also a matter of architecture and those responsible for it.