A paper dropped this week on arXiv that should be required reading for anyone building on top of commercial language models. Maxim Chupilkin's "Geopolitical alignment: Endorsement effects in large language models" finds that LLMs systematically shift their evaluations of policy-relevant information based on who is perceived to endorse it. Name a position as coming from a high-credibility geopolitical actor, and the model rates it more favorably. This is not a bug in the sense of broken code. It is a structural feature of training on human-generated text, which is itself saturated with geopolitical bias.
The Endorsement Effect and the Trained World
Chupilkin's finding connects to a broader cluster of AI safety concerns that the academic community has been circling for two years. The same week, a separate arXiv paper on interval certifications for multilayered perceptrons by Papamichail et al. addresses the foundational problem of making formal guarantees about what neural networks will and won't output. The gap between these two papers is the gap between what we can formally verify and what actually happens when you deploy a model at scale into politically contested domains. Anthropic is localizing Claude pricing for India. The EU is considering age restrictions on social media. Both moves assume that AI systems are relatively neutral delivery mechanisms for content. Chupilkin's paper suggests that assumption is wrong in ways that compound across geographies.
AI Neutrality as Marketing, Not Architecture
The geopolitical endorsement effect has a direct downstream consequence for journalism, policy analysis, and education, exactly the domains where LLMs are being most aggressively deployed. A 2026 benchmark study, L2-Bench by Edgell et al., notes that rigorous evaluation of AI-powered educational systems remains thin despite rapid adoption. If the model summarizing foreign policy documents for students is systematically favoring certain geopolitical framings based on source attribution, the educational neutrality argument collapses. The arc from Chupilkin's endorsement paper to Claude pricing to the New Yorker's chatbot family story is shorter than it looks. The AI entering homes as emotional infrastructure and the AI summarizing geopolitics for policy analysts are the same system, just at different zoom levels. The bias is not optional. It is the model.