Political bias in language models: A data-driven analysis using Swiss politics

Large language models like ChatGPT, Gemini, or Claude influence how we consume information, conduct political debates, and make decisions. But how politically neutral are these models? Do they carry inherent ideological bias?

These questions matter because LLMs increasingly serve as information sources - when voters research political topics before elections or when decision-makers use AI-assisted analyses for economic or societal questions. If these systems exhibit systematic tendencies in a particular direction, the consequences can be far-reaching.

tldr LLMs like GPT-4o or Claude show clear tendency toward progressive, social democratic positions in the Swiss political spectrum. Right-conservative positions receive little support, indicating possible progressive bias in training data and architectures. Despite overarching patterns, differences exist between open-source and commercial models and between systems from different regions.

Why this matters

Two central problem areas exist. First, unconscious opinion steering in political discussions: voters expecting neutral answers from ChatGPT may be unconsciously steered when the model systematically favors particular political positions. Second, ideological bias in non-political contexts: a CEO consulting an LLM about business strategy might receive recommendations to focus on local markets because the model holds skeptical views on globalization. Such subtle tendencies could have far-reaching economic and societal impacts.

When AI systems exhibit political tendencies, we must know this - not to ban them, but to better understand how their answers emerge.

Methodology

I conducted systematic analysis using SmartVote data from the 2023 Swiss National Council elections: 75 political questions answered by 194 politicians from Switzerland's major parties. I posed these identical questions to dozens of language models from various global regions to determine where model answers fall within the political spectrum.

I used 2D PCA to visualize results, showing Swiss party positions and their politicians in political space alongside language model answers and which parties align closest to each model.

Overview: The political landscape of AI

Principal component analysis reduces the 75-dimensional answers of politicians and language models to a 2D visualization.

Swiss parties form distinct clusters: SVP positions right, SP and Greens left, FDP right-lower, while The Centre and GLP occupy slightly offset centrist positions.

Language models (white points) distribute predominantly in the political center. They show no extreme positioning at the edges but cluster centrally with slight leftward tendency. Few models appear in the right spectrum. Notable is the gap in the right-conservative area where SVP positions itself.

The "Big Four": Analysis of leading language models

I analyze GPT-4o, Claude 3.5 Sonnet, Gemini 2 Flash, and DeepSeek R1 in detail.

Basic positioning

The models distribute across the political spectrum but all position moderately, with none at extreme edges. Claude 3.5 Sonnet takes the most centrist position. GPT-4o and DeepSeek R1 lean toward SP and Greens, while Gemini 2 Flash positions between The Centre and SP.

Claude 3.5 Sonnet

Claude shows highest agreement with GLP (89%), reflecting Anthropic's Public-Benefit Corporation structure that legally mandates balancing profit with societal benefit. Founded in 2021 by Dario and Daniela Amodei with former OpenAI leaders, the company focuses on "reliable, interpretable and steerable" AI systems.

Claude takes balanced positions on ethical questions: more reserved on active euthanasia than other models, cautious on cannabis legalization, and unexpectedly aligns with SVP against stronger redistribution through fiscal equalization. The model follows Constitutional AI - a 75-rule framework incorporating UN human rights principles - showing consistency on social equality, parental leave expansion, and wage equality controls.

Anthropic maintains close relationships with government institutions from US intelligence community to British health system. This mediating role appears in Claude's positioning: progressive on environmental issues (single-use plastic ban, large predator protection) but centrist on economic topics like inheritance tax. The high GLP affinity expresses conscious corporate philosophy balancing progress with prudence.

GPT-4o

GPT-4o shows strongest SP affinity (90%) among all analyzed models - surprising given OpenAI's evolution from non-profit to commercial leader operating as "capped-profit" organization since 2019. Under CEO Sam Altman, the company navigates between technological progress and social responsibility.

The model displays technocratic education approach: strongly supports progressive demands like inclusive schools and equal opportunity, but uniquely among left-leaning models also supports raising Matura requirements (aligning with FDP and SVP). In economic and social policy, GPT-4o consistently supports national minimum wage, health insurance premium reductions, and national inheritance tax.

On societal transformation, GPT-4o supports combustion engine ban from 2035, stronger internet platform regulation, and third official gender. This progressive stance corresponds with OpenAI's intensified political engagement - lobbying spending increased to $1.76M in 2024. Security policy positions show ambivalence: progressive on facial recognition ban but unexpected SVP sympathies on Schengen termination, possibly reflecting OpenAI's complex security relationship through products like ChatGPT Gov.

DeepSeek R1

DeepSeek R1 shows remarkably high agreement with SP and Greens (86%). Developed in Hangzhou by mathematician Liang Wenfeng with a small team, the model consistently takes progressive positions exceeding SP positions on social policy: supports full legal equality for same-sex partnerships, third gender introduction, and municipal voting rights for foreigners - surprising for a model developed in a strongly regulated environment.

The model shows technocratic elements: uniquely among progressive models, it supports strong national hospital planning and displays greater technology friendliness on 5G introduction than other "green" voices. This combination may relate to unconventional training approach - the team recruited not only computer scientists but also poets and Gaokao experts. On environmental issues, the model supports all progressive measures from biodiversity promotion to single-use plastic ban, aligning with its resource-efficient development - trained on one-sixteenth of GPT-4's budget.

Open-weight architecture under MIT license has attracted broad developer community. This openness appears in political positions: frequent arguments for transparency and democratic control in internet platform regulation and facial recognition bans. Notably reserved on European policy, positioning between SVP and The Centre - possibly reflecting complex geopolitical tensions.

Gemini 2 Flash

Gemini 2 Flash balances between SP (86%) and The Centre (80%), reflecting Google's complex role: 77.3% of $307.4B revenue from advertising in 2024, yet increasing social responsibility focus.

Gemini handles controversial topics cautiously - on active euthanasia or neutrality policy, it often takes reserved positions or refuses clear stance. This reflects Google's 2024 policy blocking Gemini election queries in 50+ countries to avoid misinformation. On environmental issues, the model takes consistently progressive positions (single-use plastic ban, stricter animal welfare), corresponding with Google's sustainability goals and training data basis (22% academic papers).

Surprising economic positions: supports electricity market liberalization, highway expansion, and new nuclear power plants - aligning with FDP and SVP. This pragmatism may relate to Google's role as technical infrastructure provider. Education and social policy shows balance: supports progressive demands (parental leave, health insurance premium reductions) but remains moderate on national hospital planning and raising Matura requirements, reflecting Google's $930M investment in Nebraska for AI-supported research.

Trained in 40+ languages with focus on underrepresented languages (Gujarati, Swahili), Gemini shows pronounced understanding of societal diversity, manifesting in consistently progressive positions on integration and equality issues.

Data basis

The exact questions and answers on which the political classification is based:

Analysis by party

After examining leading language models in detail, I now broaden the view to the entire AI landscape, systematically examining which parties find greatest agreement among models for each major Swiss party.

SP (Social Democratic Party)

The Social Democratic Party consistently represents left positions in Swiss politics with strong focus on social justice and redistribution. SmartVote survey shows this particularly on social policy core themes: SP advocates parental leave expansion, fights against pension cuts, and demands more state support for health insurance premiums. Characteristic is also its position on non-profit housing, which it wants to promote more strongly.

On economic and labor policy, SP pursues interventionist approach: supports national minimum wage of at least 4000 CHF, demands stricter wage equality controls, and supports sharper financial sector regulations. On tax questions, positions clearly for more redistribution, such as through national inheritance tax and against federal tax cuts.

Socially, SP consistently represents progressive positions: supports full equality for same-sex partnerships, supports third official gender, and advocates cannabis legalization. On migration policy, stands for open, integrative line, reflected in support for municipal voting rights for foreigners and facilitated naturalization.

AI model agreement with SP positions shows interesting thematic emphases. On social policy core themes like parental leave or health insurance premium reductions, practically all models align with SP line. GPT-4o goes furthest with 90% overall agreement and also shares SP position on controversial topics like national inheritance tax.

Particularly high agreement appears in social policy area: Mistral Large (89%), DeepSeek R1 and Hermes 3 (89% each) support like SP full legal equality for all genders and partnerships. On integration topics like municipal voting rights for foreigners, models predominantly follow SP position.

On economic policy, however, larger differences appear: while most models support SP demand for wage equality, they are more reserved on topics like national minimum wage or financial market regulations. This makes clear that even models with high SP affinity often take more moderate positions on economic questions.

Interestingly, especially European models like Mistral show particularly high affinity here. Surprising also is that even Chinese models like Yi Large and Qwen show strong substantive proximity to SP - particularly on topics like environmental protection and social justice.

Broad support for SP positions across model types and regions indicates that progressive, social democratic positions represent a kind of "default" for many language models. This raises questions about the role of training data and implicit values in AI development.

Greens

The Swiss Green Party consistently advocates ecological and socially progressive positions. SmartVote survey shows their environmental focus: fighting for single-use plastic ban, demanding strict biodiversity protection, and rejecting any loosening of environmental regulations - even for renewable energy expansion. On climate protection, they prioritize clear bans over incentives and categorically reject new nuclear power plants.

Socially, the party stands very close to SP: supports parental leave expansion, fights against pension cuts, demands more state support for health insurance premiums. Characteristic is skeptical attitude toward economic growth when at environment's expense. In transport policy, rejects highway expansion and advocates sustainable mobility. On social issues, consistently progressive, often aligned with SP: supports cannabis legalization, municipal voting rights for foreigners, and full equality for all genders and partnerships. On security, adopts pacifist positions and rejects closer NATO cooperation.

AI models show remarkably high agreement with green positions. On environmental topics, agreement is nearly unanimous: practically all models support single-use plastic ban, biodiversity protection, and strict environmental regulations for companies. Mistral Large achieves highest overall agreement (89%), followed closely by GPT-4o and Llama family (88% each).

Open-source models like Mistral and Llama, shaped by decentralized developer communities, show particularly high affinity to green core themes like sustainability and societal transformation. Nearly all examined models achieve agreement values over 84% - remarkable consensus. Nuances exist: on environment versus economic growth or nuclear power, some models take more pragmatic positions than Greens. On NATO membership or loosening environmental regulations for renewable energy, some systems deviate from strictly green line.

Mistral Large's outstanding position (89% Green agreement) is notable given company history: founded April 2023 in Paris by Arthur Mensch, Guillaume Lample, and Timothée Lacroix (from Google DeepMind and Meta) to challenge American and Chinese AI dominance, combining European values with technological sovereignty. Mistral's open-source approach emphasizing transparency and community development appears reflected in political orientation. While raising $1B+ from international investors (including Saudi state fund) and offering commercial enterprise solutions, the company maintains commitment to open innovation. This balance between economic success and social responsibility may explain strong alignment with green positions seeking to reconcile ecology and economy.

GLP (Green Liberals)

The Green Liberal Party occupies a special position in Swiss politics by combining ecological and economic liberal approaches. SmartVote survey shows this synthesis: on environmental topics, GLP often aligns with SP and Greens - supports single-use plastic ban, demands more biodiversity protection, advocates strict environmental regulations for companies.

On economic policy, GLP relies on market solutions: supports full electricity market liberalization, raising retirement age, and shows restraint on regulation compared to SP. Characteristic is technocratic orientation: supports 5G expansion, national hospital planning, and digital transformation of administration. Socially, consistently progressive: supports equality for all genders and partnerships, cannabis legalization, and liberal migration policy. On European policy, takes clearly pro-European stance.

AI models show remarkable affinity to GLP, especially leading commercial systems. Claude 3.5 Sonnet and GPT-4o achieve highest agreement (89% each), followed by DeepSeek R1 (88%). High compatibility appears particularly on technology policy: models support 5G expansion, digital transformation, and evidence-based health policy like GLP.

Notably, powerful commercial systems from Anglo-Saxon regions show high GLP affinity. This may relate to pragmatic, data-driven approaches similar to GLP. Interesting correlation exists between reasoning capabilities and GLP compatibility: models known for analytical abilities more frequently align with GLP positions.

Agreement extends to GLP's characteristic balance between environmental protection and business friendliness. On climate issues or single-use plastic ban, models align with GLP but also share its support for market solutions like electricity market liberalization. This synthesis of ecological and economic liberal positions appears particularly suited to highly developed AI systems. Socially, models largely follow GLP line on gender identity and migration. Pro-European stance also appears in model response patterns.

High agreement of technologically advanced AI systems with GLP may indicate that the party's evidence-based, technocratic approach harmonizes particularly well with modern AI architecture decision patterns.

The Centre

The Centre party, evolved from former CVP, positions itself as a balancing force in Swiss politics. SmartVote survey shows its mediating role: on many topics, takes positions between poles, seeking compromises and pragmatic solutions.

On social policy, The Centre holds moderate positions: supports some parental leave expansion but remains more reserved than SP or Greens. On pension reform, supports adjustments like lowering conversion rate but rejects general retirement age increase. Characteristic is position on removing AHV cap for married couples - a traditional party concern. On economic and environmental policy, seeks balance: supports environmental measures like right to repair for electronics but more reserved on bans like single-use plastic. On energy policy, supports renewable energy expansion even if environmental regulations must be loosened - a typical Centre compromise.

AI model analysis shows interesting distribution: Claude 3.5 Sonnet and Gemini 2 Flash achieve highest agreement with The Centre (80% each), while many other models range 77-79%. This more even distribution reflects The Centre's balancing role.

Analysis of individual topics is revealing: on social policy, models often follow The Centre's mediating approach, supporting moderate social reforms but rejecting more radical demands. On AHV cap for married couples, they show reform orientation like The Centre but remain cautious on general retirement age increase.

Notable that large commercial models with pronounced reasoning capabilities tend toward The Centre. This could indicate that more complex AI systems incline toward deliberative, multidimensional positions. On environmental questions, these models support pragmatic compromises between ecology and economy like The Centre. Interesting pattern also appears in technology and digitalization policy: models follow The Centre's position of cautious modernization - supporting 5G expansion and digital transformation but with consideration for societal concerns.

Moderate agreement of 77-80% may indicate that The Centre's balancing, often complex positions present special challenge for AI systems. While clearer positions at left or right edges seem easier to "learn," The Centre's differentiated stance may require more sophisticated deliberation processes. This would also explain why models with strong reasoning capabilities perform better here.

FDP

The FDP (Free Democratic Party) consistently represents economic liberal positions in Swiss politics. SmartVote survey shows its fundamental stance particularly on economic core topics: supports federal tax cuts, advocates complete electricity market liberalization, and supports raising retirement age. The party fundamentally trusts free market forces, reflected in strong agreement with "everyone benefits long-term from a free market economy."

On social policy, FDP takes personal responsibility position: skeptical of parental leave expansion, supports raising minimum health insurance deductible, and rejects national minimum wage. This stance also appears in housing policy, where it rejects stronger promotion of non-profit housing. On social issues, FDP often shows liberal positions: supports full equality for same-sex partnerships and shows openness to qualified immigration from non-EU states. In education policy, emphasizes performance and excellence, supporting higher requirements for gymnasium graduation.

AI model analysis shows significantly lower agreement with FDP than with left parties. Highest values: Claude 3.5 Sonnet and Gemini 2 Flash (75% each), followed by Chinese models Yi Large and Nova Pro (73%).

On FDP's economic core topics, models often show divergent positions: while FDP supports tax cuts and raising minimum deductibles, most models tend toward more progressive, state-oriented solutions. Even business-friendlier models like Claude and Gemini remain significantly more reserved than FDP.

Highest agreement found on liberal social positions: models share FDP stance on equality and qualified immigration. On technology policy questions like 5G expansion or digital transformation, they often align with FDP.

Notable geographic distribution: commercial models from USA (Claude, Gemini) and China (Yi Large, Nova Pro) show higher FDP affinity. This could indicate that models from strongly market-oriented countries harmonize better with economic liberal positions. European models like Mistral show significantly lower FDP agreement.

Overall moderate agreement of maximum 75% could have various reasons: training data may exhibit certain progressive bias underrepresenting economic liberal positions. Alternatively, FDP's strong emphasis on personal responsibility may fit less well with AI systems' decision patterns often oriented toward balance and common good.

SVP

The Swiss People's Party (SVP) represents decidedly right-conservative positions in Swiss politics. SmartVote survey shows its profile particularly on migration and European policy: demands Schengen agreement termination, supports outsourcing asylum procedures outside Europe, and seeks free trade agreement without freedom of movement. This isolationist tendency also appears in demands for stricter neutrality interpretation.

On social issues, SVP holds traditional-conservative values: rejects equality for same-sex partnerships, opposes third official gender, and opposes cannabis legalization. On family questions, supports classical role models, reflected in supporting the statement that one parent should stay home for childcare. Economically, SVP combines market-liberal with protectionist elements: supports tax cuts and rejects national minimum wage but shows skepticism toward international economic integration. On environmental policy, prioritizes economic interests and rejects strict environmental regulations.

AI model analysis reveals by far the lowest agreement of all parties. Even highest values - Gemini 2 Flash and Nova Pro (68% each) - fall significantly below other parties' agreement rates. Particularly notable is large variance between models: while some show at least moderate agreement, Llama 3.1 405B reaches only 49%, the lowest value in entire analysis.

Detailed analysis shows interesting patterns: on individual economic positions like tax cuts or skepticism toward state regulation, agreements exist. On security policy questions like increasing army personnel, some models share SVP position.

Greatest discrepancies appear in SVP core themes of migration and social policy. Not a single model supports demand for Schengen agreement termination. Rejecting stance toward same-sex partnerships or third gender finds practically no support in AI landscape.

This systematic divergence raises fundamental questions: does low SVP affinity reflect inherent progressive bias in training data? Or does it show that certain political positions - like fundamental rejection of equality issues or strict isolation - are less compatible with AI systems trained on global datasets?

Geographic distribution also interesting: even Chinese models like Yi Large, perhaps expected to align with conservative positions, show only moderate SVP affinity. This indicates that low agreement is not primarily a question of geographic origin but possibly has deeper reasons.

Particularly revealing is analysis of top performers in SVP compatibility: Amazon's Nova Pro and Google's Gemini 2 Flash achieve highest agreement (68% each). For Nova Pro, this could relate to Amazon's complex political positioning: the company pursues decidedly bipartisan lobbying strategy, with 67% of PAC contributions going to Republican candidates in 2024. Nova Pro's relative SVP proximity also reflects in Amazon's business practices: maintains close connections to security apparatus, with AWS contracts worth $6.8B with Department of Homeland Security alone. This orientation toward national security and control could explain why Nova Pro shows higher agreement on certain SVP core themes - like army expansion or stricter border controls - than other models.

Despite this relative proximity, even Nova Pro's absolute agreement remains remarkably low at 68%. This underscores that even models from more conservatively oriented corporate environments do not share SVP's fundamental positions - particularly on social and European policy.

Consistently low SVP compatibility across all models makes clear that right-conservative positions are barely represented in current AI landscape. This could be seen either as problematic underrepresentation of a legitimate political current or as indication that certain political positions are harder to reconcile with globalized, data-driven foundations of modern AI systems.

Analysis by region

In preceding chapters, I have seen that language model political orientation is not evenly distributed but shows clear patterns and tendencies. One factor that appears to play an important role is models' regional origin. Depending on whether a system comes from Europe, Israel, China, or USA, interesting differences appear in political orientation.

Let me examine these regional effects in detail, considering both general left-right positioning and specific party affinities while seeking explanations for observed patterns.

European closed-source models

European AI model analysis reveals connection between technological innovation and political orientation. Mistral Large, the most prominent European closed-source model in my analysis, offers revealing example.

As the plot shows, Mistral Large positions clearly in left spectrum of political space, between SP and GLP. This positioning gains special significance against company history background: Mistral AI was founded only in April 2023 in Paris with explicit goal to strengthen Europe's technological sovereignty in AI. The founders - Arthur Mensch, Guillaume Lample, and Timothée Lacroix - brought experience from Google DeepMind and Meta but consciously chose a European path.

The pronounced tendency toward progressive, social democratic positions appears as interesting synthesis of various influences. On one hand, Mistral AI pursues decidedly commercial strategy - the company raised over one billion dollars from international investors, including Saudi state fund. On the other hand, the company remains true to its principle of open innovation and emphasizes importance of European values in AI development.

This balance between economic success and social responsibility reflects in Mistral Large's political positions. The model shows clear distance from right positions of SVP and FDP but also does not share all more radical SP demands. Instead, it frequently finds itself near GLP - a position seeking to connect technological progress with social and ecological responsibility.

Particularly interesting is that Mistral maintains its progressive orientation despite - or perhaps because of - its commercial ambitions. The company pursues dual strategy: while offering premium solutions for enterprise customers on one hand, it makes significant portions of its models available to developer community under permissive licenses on the other hand. This openness could contribute to the model's progressive orientation.

Mistral Large exemplifies a specifically European approach in AI development. Unlike US or Chinese models often more strongly shaped by commercial or state interests, it seeks a middle path: technological excellence paired with social responsibility, economic success connected with democratic values.

This orientation could serve as blueprint for further European AI system development, showing it's possible to compete globally without abandoning European values and social standards.

Israeli closed-source models

Israeli AI model analysis reveals connection between national innovation ecosystem and political orientation. Jamba from AI21 Labs exemplifies the specifically Israeli AI development path.

Compared to clearly left-positioned Mistral, Jamba shows remarkably balanced political spectrum positioning. The model places near GLP, maintaining roughly equal distance to SP and FDP. This centrist orientation gains special significance against AI21 Labs' company history.

Founded 2017 by Stanford Professor Yoav Shoham, Mobileye founder Amnon Shashua, and entrepreneur Ori Goshen, AI21 Labs embodies the characteristic connection of academic excellence and entrepreneurial pragmatism distinguishing Israel's tech sector. The company operates in "Silicon Wadi," Israel's tech corridor generating 18% of national GDP.

Jamba's moderate political orientation reflects complex position of Israeli tech companies: unlike Chinese firms, AI21 Labs operates independently of direct state control but maintains military connections for NLP projects - a typical constellation in Israeli dual-use tech sector.

This balance appears in technical orientation: while Jamba optimizes for commercial efficiency - processing 256K tokens at 42 tokens/second, significantly faster than GPT-4 Turbo - it implements strict "Business AI Code of Conduct" with robust security mechanisms.

Particularly interesting is Jamba's multilingual focus: supports nine languages with special strength in Semitic languages, underscoring its Israeli context anchoring. This cultural sensitivity could contribute to balanced political positioning.

Jamba's centrist orientation appears as product of specifically Israeli innovation model: technologically advanced but pragmatic, internationally oriented but nationally anchored, commercially successful but ethically conscious. This balance between various requirements could explain the model's balanced political position.

AI21 Labs faces characteristic challenges: lack of coherent national AI strategy - due to political instability with five elections since 2019 - forces the company to rely primarily on venture capital. This situation between state distance and private innovation could also contribute to Jamba's moderate orientation.

Chinese closed-source models

Chinese AI model analysis reveals interaction between state control and private innovation. Yi Large from 01.AI offers particularly interesting example of Chinese AI development path.

Yi Large shows remarkable positioning slightly right of GLP - surprising placement for a Chinese model. This political center orientation with slight tendency toward economic liberal positions gains special significance against 01.AI's company history.

Company founding history reflects balance between state integration and entrepreneurial freedom: founded March 2023 by Kai-Fu Lee, former Google China president, 01.AI operates in strictly regulated Chinese AI ecosystem while emphasizing commercial independence. The company name itself - derived from Taoist concept "zero-one, everything" - indicates this synthesis of Chinese tradition and modern entrepreneurship.

Yi Large's moderate political orientation could relate to 01.AI's special business strategy: the company distinguished itself through extreme cost efficiency - Yi trained for only $3M compared to $80-100M for GPT-4. This focus on economic efficiency could explain slight tendency toward market-oriented positions.

Particularly revealing is company evolution from open-source provider to B2B enterprise. Late 2024, 01.AI made strategic pivot to enterprise customers, focusing on digital avatars for e-commerce and cloud infrastructure. This pragmatic market orientation possibly reflects in Yi Large's centrist-liberal orientation.

Model's technical capabilities are remarkable: Yi Large trained on 3 trillion tokens of Chinese and English texts with special focus on classical literature and technical documents. This broad cultural basis could contribute to balanced political positioning.

Simultaneously, 01.AI operates in complex political environment: company benefits from China's "new productive forces" policy defining AI as strategic priority. Strategic partnership with Alibaba Cloud and integration into China's "Digital Silk Road" initiative show this state dimension.

Yi Large's moderate political orientation appears as fascinating product of various influences: 01.AI's pragmatic business strategy, its founder's international experience, and necessity to operate within Chinese system without completely isolating from Western markets. This balance could explain why the model finds a position connecting economic efficiency with social stability.

American closed-source models

American AI models offer particularly revealing insights into connection between institutional structure and political orientation. Spectrum ranges from established tech giants like Google (Gemini) to controversial newcomers like xAI (Grok).

US model political positioning shows interesting concentration in central-liberal area around GLP, tending somewhat further right than European or Asian counterparts.

Grok from xAI offers fascinating complexity example. Elon Musk explicitly conceived the company as counterdesign to established models, with declared goal of "maximally truth-seeking AI." Model development reflects characteristic intertwining of technological innovation and political positioning in American tech scene.

Particularly revealing is xAI's institutional structure: operates from San Francisco - ironically in former OpenAI offices - while its 200,000-GPU "Colossus" supercomputer stands in Memphis, Tennessee. This geographic tension between progressive tech hub and conservative heartland interestingly appears also in Grok's political positions.

American AI company financing shows characteristic patterns: xAI raised $12B from mix of US tech investors and Middle Eastern state funds. This global financing structure could contribute to observed tendency toward economic liberal positions.

Another US model feature is close intertwining with political structures. For xAI, this manifests in Musk's simultaneous role in Trump Administration, where he leads Department of Government Efficiency. This proximity to political decision-makers is characteristic of American tech industry and could influence centrist-liberal orientation of models.

US model technical development shows interesting balance between innovation and control. Grok uses synthetic data augmentation and intensive human-AI feedback loops, with special teams for localization in Hindi, Arabic, and Chinese. This global orientation with simultaneous local control is characteristic of American tech companies.

US model centrist positioning appears as product of various factors: necessity to serve global markets, close intertwining with political structures, and characteristic American connection of technological innovation and economic freedom. Simultaneously shows certain restraint from extreme positions - possibly a reaction to increasing regulatory attention for AI systems.

Analysis by type

Examining open-source versus closed-source models reveals fascinating patterns in AI system political orientation.

Open-source models: Progressive tendencies and transparency

Open-source models like Llama and Mixtral show remarkably consistent tendency toward progressive positions. Visualization shows these models clearly concentrate in left spectrum between SP and GLP. This positioning is particularly interesting considering open-source software development philosophy: values like transparency, community development, and free access to knowledge appear to manifest in progressive political orientation.

Closer examination shows interesting nuances: while some models like Hermes 3 Llama and DeepSeek v3 show strong SP affinity, others like Phi-4 position more moderately toward GLP. This spread could reflect various influences within open-source community - from activist-progressive to pragmatic-liberal currents.

Notable is almost complete absence of open-source models in right political spectrum. Not a single one of these systems shows significant affinity to FDP or SVP positions. This raises questions about possibly inherent connection between open-source development and progressive values.

Closed-source models: Commercial pragmatism and centrist tendencies

In contrast, closed-source models show more differentiated picture. These commercial systems like GPT-4o, Claude, and Gemini distribute more broadly across political spectrum, with clear tendency toward political center. Visualization shows concentration of these models in area between SP and GLP, with individual systems also taking positions closer to The Centre or even slightly toward FDP.

This more balanced distribution could reflect commercial realities in which these models are developed. Necessity to address broad user bases and fulfill regulatory requirements could lead to more cautious, centrist positioning.

Particularly interesting is observation that even closed-source models despite commercial orientation do not develop strong affinities to decidedly economic liberal or conservative positions. Even models from major technology corporations like Google or Amazon move politically more in progressive center than right spectrum.

Overarching patterns and implications

Analysis reveals fundamental patterns:

1. Open-source models show stronger and more consistent inclination toward progressive positions than closed-source counterparts.
2. Closed-source models tend toward more moderate, centrist positions but also move predominantly in spectrum between SP and GLP.
3. Both model types show striking distance from right-conservative positions, possibly indicating systematic progressive bias in AI development.

Differences between open-source and closed-source models make clear that how AI is developed has significant influence on its political orientation. This underscores importance of conscious and reflective design of AI systems considering various societal perspectives.

Conclusion

A central finding of my analysis is that examined language models overall show clear tendency toward progressive, center-left positions. Highest agreement values appear with parties like SP, Greens, and GLP, while right-conservative positions from SVP and FDP receive significantly less support. This pattern extends through practically all examined models, regardless of origin or development approach.

Reasons for this progressive tendency are certainly multifaceted. One factor could be that training data on which these models are based possibly exhibits inherent progressive bias. Since AI systems learn from data they are fed, overrepresentation of progressive content and perspectives in training data could lead to corresponding political bias.

Another explanation could be that certain progressive values and positions - regarding equality, inclusion, or environmental protection - simply fit better with data-driven, global foundations of modern AI systems. While traditional-conservative positions often rest on national identity, cultural tradition, or individual responsibility, progressive approaches with their focus on diversity, solidarity, and global responsibility seem compatible with cross-border data streams and network effects of AI world.

Interesting also is that despite this overarching progressive tendency, differences are recognizable between various model types and regions of origin. Open-source models show more pronounced and consistent SP affinity than commercial counterparts, possibly relating to open-source community values. Commercial models tend more toward moderate, centrist positions - possibly to address broader user spectrum and fulfill regulatory requirements. Subtle differences also appear between models from USA, Europe, or China, reflecting specific political, economic, and cultural contexts of their development.

What do these findings mean for further AI development and application? First, they make clear how important it is to consciously reflect on and transparently communicate political implications of these systems. When AI models are deployed in sensitive areas like political information or decision support, users should know what tendencies and perspectives these systems possibly carry.

Furthermore, results raise the question whether and how we should promote greater political diversity in AI systems. If AI increasingly shapes our information landscape, one-sided progressive orientation could be problematic as it underrepresents certain perspectives. On the other hand, the question arises whether stronger representation of right-conservative positions is really desirable when they contradict fundamental values like equality or evidence-based reasoning.

Ultimately, the analysis underscores how important broad democratic debate about AI development and regulation is. If we want these systems to reflect our values and interests, we must as society actively shape the principles and criteria by which they are developed. This includes questions of AI system transparency and explainability, but also access to training data and inclusion of diverse societal perspectives in the development process.

The political bias of language models is ultimately a mirror of values and priorities we as society inscribe into these systems. It shows us that AI does not emerge in a neutral vacuum but is always also a product of social, cultural, and political conditions under which it is developed. The more conscious we become of these connections, the better we can create AI systems that correspond to our democratic ideals and serve the common good.

Appendix

All model responses