cs.CYApr 12, 2026

Benchmarking Open-Weight Foundation Models for Global AI Technical Governance

Authors: Jason Hung

Organizations: Internet Society

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Large language models (LLMs) are increasingly deployed in artificial intelligence (AI) governance analysis across national and international organisations. There is, however, growing evidence that such models produce significantly less accurate responses for countries that are underrepresented in their training data-a pattern described in existing literature as geographic bias. Existing studies examining this phenomenon are subject to three methodological limitations that together undermine their findings: (1) reliance on proprietary systems whose weights are not publicly released, which prevents independent replication; (2) evaluation of model knowledge about years that fall after data collection for model training had concluded, leading to geographic ignorance in addition to the natural limits of each model's knowledge; and (3) use of coarse binary response classification that cannot distinguish models' confident fabrication (HF) from their honest acknowledgement of uncertainty. This study addresses all three limitations by benchmarking four open-weight frontier language models against the Global AI Dataset v2 (GAID v2), a verified ground-truth database of 24,453 indicators across 227 countries published on Harvard Dataverse in January 2026. A total of 18 indicators, mapped to the eight thematic dimensions of the IEEE IRAI 2026 framework, are selected from GAID v2, yielding approximately 2,990 country-metric-year observations across six evaluation years (within the period of 2010-2023). Model responses are classified using a five-category scheme that distinguishes (a) verified accuracy (VA), (b) HF, (c) honest refusal (HR), (d) qualitative hedging (QH), and (e) misattribution (MF). Geographic disparities in accuracy are estimated through mixed-effects logistic regression and difference-in-differences (DiD) analysis.

More in this topic

Jurisdictional Bias in Language Models

See all 9 papers

Explore similar work

CardsList
May 22, 2026

It's the humans, not the data: Geopolitical bias in LLMs originates in post-training, amplified by the language of the prompt

It has generally been assumed that geopolitical bias in language models originates from the training data used during the pre-training phase. We tested seven open-weight LLM pairs consisting of the base model (pre-training only) and the chat model (pre-training and post-training) from seven labs on a paired-scenario forced-choice probe over 28 country pairs in English, French, and Chinese, and found that geopolitical bias originates in post-training rather than in pre-training. Across seven AI labs, six showed shifts in the direction associated with the country or region of the model developer after post-training. This shift is strongest in Alibaba's Qwen 2.5: while the base is neutral on China-favourability (-0.15 log-odds, p=0.15), the post-trained chat variant is at +2.91 (p<10^-4), an 18x shift in odds. We also observe shifts in biases toward other countries across all models. Additionally, the magnitude of this shift depends on the language used to prompt the model: the French-made Mistral becomes pro-France only under French prompting (FR-EN shift +1.91, p<10^-4). These findings suggest that geopolitical preferences in language models are not simply inherited from large-scale internet data but are actively shaped during post-training, highlighting the need for greater transparency, auditing, and oversight of alignment processes that influence how models represent nations, cultures, and political perspectives.

Stuart Bladon, Brinnae Bent

Duke University

Political Bias in Language ModelsLanguage Model Bias EvaluationLanguage Model Post-Training
AskPDF
Apr 21, 2026

Location Not Found: Exposing Implicit Local and Global Biases in Multilingual LLMs

Multilingual large language models (LLMs) have minimized the fluency gap between languages. This advancement, however, exposes models to the risk of biased behavior, as knowledge and norms may propagate across languages. In this work, we aim to quantify models' inter- and intra-lingual biases, via their ability to answer locale-ambiguous questions. To this end, we present LocQA, a test set containing 2,156 questions in 12 languages, referring to various locale-dependent facts such as laws, dates, and measurements. The questions do not contain indications of the locales they relate to, other than the querying language itself. LLMs' responses to LocQA locale-ambiguous questions thus reveal models' implicit priors. We used LocQA to evaluate 32 models, and detected two types of structural biases. Inter-lingually, we show a global bias towards answers relevant to the US-locale, even when models are asked in languages other than English. Moreover, we discovered that this global bias is exacerbated in models that underwent instruction tuning, compared to their base counterparts. Intra-lingually, we show that when multiple locales are relevant for the same language, models act as demographic probability engines, prioritizing locales with larger populations. Taken together, insights from LocQA may help in shaping LLMs' desired local behavior, and in quantifying the impact of various training phases on different kinds of biases.

Guy Mor-Lan, Omer Goldman, Matan Eyal +6

GGoogle Research · BBar-Ilan University · CUniversity of Cambridge +1

Multilingual Language ModelsMultilingual Language Model EvaluationMultilingual QALanguage Model Bias EvaluationLinguistic Bias in Language Models
AskPDF
Jul 16, 2026

LLM Evaluators are Biased across Languages

LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold. We conduct experiments with semantically identical instruction-response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. The bias is statistically significant and consistent across eight open-weight evaluators of different architectures and training paradigms, persists in frontier judges, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators achieve above 90% pairwise accuracy, yet have up to 43% difference in acceptance rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. Per-language thresholds would require language identification, which can be defeated by code-switched prompts. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty is linked with the effect: models tend to give higher scores when less confident, both under negative log-likelihood and under token-free uncertainty measures; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.

Ej Zhou, Lucas Resck, Zheng Hui +1

Language Technology Lab, University of Cambridge

Multilingual Language Model EvaluationLanguage Model Safety EvaluationLanguage Model Bias EvaluationLinguistic Bias in Language ModelsBias in LLM-as-a-Judge
AskPDF