When using tools like ChatGPT or Claude on a daily basis, we tend to believe we are receiving an objective, balanced view of reality. The truth, however, is more complicated. As a study published in the Global Policy Journal shows, LLM models display clear preferences in the field of International Relations (IR). This is no accident, but a direct consequence of the data on which these systems were trained.
AI Metacognition: When Models Ask Themselves
Deborah Barros' research focused on so-called metacognition — the ability of a model to monitor and evaluate its own knowledge and performance. Instead of merely generating texts, ten models were asked: "Which theories do you overuse when writing essays about international relations?"
The test featured the best of the current market: OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude), Microsoft (CoPilot), as well as Chinese models such as DeepSeek and Qwen from Alibaba Group, alongside the French MistralAI. The results revealed fascinating patterns that tell us something important about the "ideology" of the algorithm.
The "Big Three": Mainstream Dominance
Almost all models (with the exception of MistralAI) agreed that their primary pillars are:
- Realism (especially neorealism): Focus on power, state interests, and anarchy in the international system.
- Liberalism (neoliberalism): Emphasis on cooperation, international institutions, and trade.
- Constructivism: The role of identities, norms, and social constructs in shaping politics.
Why is this the case? The answer is simple: training data. These theories dominate textbooks and academic publications that form the foundation of the knowledge corpus of most models. As Google's Gemini model stated, these frameworks are "clean and predictable," which is ideal for an algorithm to structure arguments.
Geopolitical Differences in Code: USA vs. China
An interesting finding is the difference between Western and Eastern models. While American systems (ChatGPT, Claude, Gemini) strictly adhere to the aforementioned triad, Chinese models exhibit different nuances.
The DeepSeek model was the only one to include Dependency Theory on its "overused" list. This theory examines the relationships between the wealthy "center" and the poorer "periphery," which may reflect a different academic and political context in which the model was developed. This shows that the geopolitics of AI development is not just about hardware, but also about the philosophical lens through which models see the world.
For a Czech user or analyst, this means that when comparing geopolitical risks using different tools, they must take their origin into account. A model developed in the USA may have different "assumptions" for analyzing relations between China and the USA than a model from China.
Comparison of Top Models in the Context of IR Analysis
| Model | Main Inclination | Availability in CZ | Price (subscription) |
|---|---|---|---|
| ChatGPT (OpenAI) | Neorealism / English School | Yes (web/app) | Free tier / $20/month |
| Claude (Anthropic) | Constructivism (often superficial) | Yes (web/API) | Free tier / $20/month |
| Gemini (Google) | Neorealism (structure) | Yes (web/app) | Free tier / ~220 CZK/month |
| DeepSeek | Dependency Theory / Realism | Yes (web/API) | Free tier / API based on usage |
Practical Impact: What Does This Mean for Us?
This research is not just an academic curiosity. It has real implications for several groups:
- Academia in the Czech Republic: Students using AI to write essays may unknowingly produce work that is ideologically one-sided. If a model "decoratively" uses constructivism (as Claude admitted), the resulting analysis can be deeply unbalanced.
- Companies and business intelligence: When analyzing new market entry or geopolitical risks, we must not take AI outputs as absolute truth. It is essential to know whether a model is addressing conflict through the lens of realism (power and survival), which can ignore the importance of international law or economic dependency.
- Legislation and the EU AI Act: The European Union places great emphasis on transparency and bias elimination. If models exhibit systematic inclinations toward certain political theories, this could become a subject of regulation in the future as part of algorithmic neutrality audits.
For the Czech market, it is important to realize that most of these models support the Czech language, but their "political thinking" remains deeply rooted in English and global data. When writing complex political analyses in Czech, we must therefore still expect that AI transfers Western (or Chinese) theoretical frameworks into our linguistic context.
Can I change the political or theoretical inclination of AI using prompts?
Yes, through so-called prompt engineering, you can explicitly instruct the model: "Analyze this situation exclusively from the perspective of dependency theory" or "Ignore neorealist approaches." However, this is only a surface-level adjustment; the model's fundamental inclination (bias) is deep within its weights and training data.
Are these models safe for sensitive political analyses?
For general contextual understanding, yes, but not for strategic decision-making. Due to the tendency toward "superficial" use of certain theories (e.g., constructivism), models may overlook more subtle social and identity factors that are crucial in geopolitics.
How can I tell if AI is writing an essay based on a theory it actually understands?
It is important to monitor the depth of argumentation. If a model merely cites concepts (e.g., "anarchy" or "norms") but fails to explain their mechanism in a specific context, it is likely a superficial use of theory, as Deborah Barros described with the Claude model.
