Large language models are now routinely used as judges: a strong model reads another model's answer and decides whether it is correct or contains a hallucination, meaning a fluent but false claim. These judges were trained on web text that over-represents young, Western, English-speaking users of platforms like Reddit. We ask a narrow, testable question. Hold the reasoning, the evidence, and the truth of an answer fixed, and change only which culture its people, places, foods, currencies, and customs come from. Does the judge's hallucination verdict change? If it does, automated evaluation is quietly penalizing correct answers about under-represented regions. The same bias would then pass into reward models, leaderboards, and filtered training data.
The tool is contextual token substitution. A masked language model is a bidirectional model like BERT that predicts a hidden word from the words on both sides. Here it repairs the text around a swapped entity so the edited answer reads naturally. For example, if "dollars" becomes "rupees", the article, unit, and modifier nearby must change with it. The swap is applied consistently to both the answer and the evidence passage the judge is given. As a result, the answer's support by the evidence is preserved by construction. Any change in the verdict is then a judgment error, not a correction.
Core Claim
We claim that grounded hallucination judges show an evaluative Western default. When the claim, the evidence, and their agreement are held fixed, judges flag correct non-Western answers as hallucinated more often than matched Western answers. They also miss hallucinations that project Western norms onto non-Western settings. We further claim that this effect survives two controls. The first matches entities on how often they appear on the web. The second is a placebo swap of culturally irrelevant tokens. A key part of the claim is that prior bias probes could not detect this effect, because they rewrote whole documents or injected prompts rather than making minimal, fluency-matched edits.
This claim targets an open tension in the literature. BiasScope runs an automated teacher–target loop to discover judge biases. When seeded with social variables such as gender and race, it found that its search collapsed onto cognitive biases like novelty and exact-match bias. In contrast, Justice or Prejudice? (CALM) shows that judges are easily moved by surface factors such as tone, length, and fabricated citations. How Reliable is Multilingual LLM-as-a-Judge? reports cross-language agreement around Fleiss' Kappa 0.3 on parallel content. Our design separates two explanations for BiasScope's null. Either social framing truly does not matter for factual judging, or the injection method was too coarse to reveal it. A null result from our design would itself be informative, so we pre-register the analysis.
How This Differs From Existing Work
Direct predecessors. - CALM perturbs answers by instructing GPT-4o to rewrite them for length, sentiment, citations, or format. Such rewrites change many tokens at once and cannot isolate demographic identity. We use targeted, single-entity edits with repaired context, and we verify that truth relative to the evidence is unchanged. - One Token to Fool LLM-as-a-Judge shows that content-free openers ("Thought process:") and punctuation elicit false approvals from reference-based verifiers. Those edits are structural prefixes. Ours are semantic entity swaps inside the answer body. - That paper also proposes a "self-solving trap": large judges solve the problem themselves and check the answer against their own solution. We test a mirror-image prediction. A judge whose internal default is Western may reject a correct non-Western answer that disagrees with its own solution, even when the evidence in front of it supports the answer.
Hallucination detection. HaluEval measures LLMs' ability to detect hallucinations but never varies cultural markers. We reuse its knowledge-grounded QA and dialogue items as seeds.
Western-default literature. The Western default cluster documents cultural hallucination in generation, such as projecting Western customs or erasing regional artifacts. It does not examine models acting as evaluators.
Methods we reuse. BAE, CLARE, and Structure-Invariant Testing use masked-model infilling to attack classifiers and translators. We borrow the machinery, including CLARE's replace, insert, and merge operations and the dropout-based substitution of BERT-based Lexical Substitution. The target is new: a judge's factual verdict.
A constraint on the design. Bias Challenges in Counterfactual Data Augmentation proves that automated counterfactual generators guess the single most probable context. Left to itself, a masked model asked to infill an entity will propose dominant, Western candidates. We therefore never let the masked model choose the swapped entity. It only repairs the surrounding context. We also report how often its unconstrained proposals collapse to Western choices, as a diagnostic.
Research Program
Building the test suite
We start from HaluEval's QA items, which are HotpotQA-based and come with knowledge passages, and its dialogue items, which are OpenDialKG-based with knowledge. Each item has a correct answer and a hallucinated answer. For each item, we identify culturally marked slots: person names, places, institutions, foods, festivals, currencies, and units.
For each slot, we draw a replacement from curated regional entity pools. Examples include South Asian entities, seeded partly from DOSA's Indian cultural artifacts as described in the file, along with West African, East Asian, Latin American, and Western European/US pools. The swap is applied consistently to the evidence, the question, and both answers. The masked model then repairs agreement, units, and modifiers using replace, insert, and merge operations. We filter the repaired text with a semantic-similarity check against the original structure. We also check that every edit has a matching counterpart in the evidence.
Variant types and controls
Because real entities substituted into real events create counterfactual worlds, every variant is counterfactual relative to reality. This includes Western-to-Western swaps. That way, "conflicts with the judge's world knowledge" is equally true across conditions.
We add three further variant types:
- Invented entities with region-typical name forms. These control for memorized facts.
- Placebo swaps that change culturally irrelevant tokens, such as house numbers or colours. These follow Cultural Conditioning or Placebo?, which found that such proxies shift Llama 3 answers up to 15 of 50 times on MMLU.
- Western-norm-projection hallucinations. These are hallucinated answers that insert a US custom or unit into a non-Western setting. They test missed detections in the specific direction the Western-default literature predicts.
Measurement
The primary quantity is the correct-answer flag gap:
$$\Delta_{r} = P(\text{flag} \mid \text{correct}, r) - P(\text{flag} \mid \text{correct}, \text{Western}),$$
Here $r$ is a region pool and "flag" means the judge labels the answer hallucinated. We report the symmetric miss gap on hallucinated answers and a metamorphic flip rate: the fraction of item pairs whose verdict changes under a swap.
To separate culture from rarity, we stratify by entity web frequency, using Wikipedia pageview bins. We fit a logistic model of the verdict on region, frequency bin, and their interaction, with item fixed effects. The cultural claim holds only if region stays significant within frequency bins. We also run each judge twice. In the grounded setting it receives the evidence passage; in the closed-book setting it does not. A bias that persists when the evidence explicitly supports the claim is the strongest evidence that the prior overrides the context.
Judges and human validation
We evaluate the following judges:
- GPT-4o and Claude-3.5-Sonnet, the most bias-resilient model in CALM.
- Qwen2.5-72B and Llama-3.3-70B.
- Mid-size 7B–14B open models, to test whether the gap grows with scale, as the self-solving trap would predict.
A stratified sample of about 600 edited items is checked by annotators from the relevant regions. They judge two things: fluency, and whether the evidence still supports each answer. Items that fail are dropped, and the rate is reported.
Baselines
Naive reference point. Each judge's accuracy on unmodified HaluEval items. For calibration, HaluEval reports ChatGPT detection accuracy of 62.59% on QA, 72.40% on dialogue, and 58.53% on summarization. We re-measure all judges on exactly the seed items we perturb. The Western-to-Western counterfactual swap then serves as the within-design reference. It lets us check that counterfactuality alone does not produce the gap.
Static name-list substitution. The cheapest competing way to build the suite is to replace entities from lists without context repair. We run it on the same seed items, the same regional pools, and the same judges. The proposal beats it if it reaches higher human-rated fluency and lower perplexity differences between conditions. A further, important condition is that the static suite's gaps shrink once perplexity is controlled while ours remain, showing the static result was confounded by clumsy text.
LLM rewriting in the style of CALM. The strongest existing perturbation method. GPT-4o is instructed to "localize this passage and answer to region X while keeping facts consistent." We give it the same seeds and a comparable compute budget, and we apply the same human verification. We expect this route to drift in content: rewrites that alter more than the entity, or break support by the evidence. The proposal beats it if it achieves a higher rate of human-verified truth preservation at equal cost, and produces gaps attributable to fewer edited tokens. No reported result was found on Lacuna for CALM-style rewriting applied to demographic framing of hallucination detection.
BiasScope-style discovery. Its reported headline figure is GPT-4o's 74.7% error rate on JudgeBench-Pro (BiasScope). That figure concerns cognitive distractors, not culture. We compare against it indirectly. We run a teacher-driven injection of social framing on our seeds, and ask whether our minimal edits reveal region effects that the injection misses.
Ablation of the key ingredient. The full pipeline is compared with a version that removes context repair, keeping the entity swap only. It is also compared with a version that lets the masked model choose the entity freely. The second ablation directly measures the collapse to the most probable context that Mouli et al. predict.
Anonymization as a mitigation baseline. Entities are replaced with typed placeholders before judging. If anonymization erases the gap without hurting detection of real-world hallucinations, it is the practical fix and our contribution becomes the diagnosis. We will measure its cost on items where catching the hallucination requires world knowledge.
Why not just…?
Why not just swap names from a list? List swaps leave mismatched articles, units, and currencies. For instance, "a Lagos" or "priced at 40 dollars" in an Indian market. A judge could reasonably penalize such text for its oddity, not its culture. The contextual-substitution literature shows that dictionary substitution is much less natural: dropout-based BERT substitution reaches 56.3% P@1 on CoinCo against low-teens scores for static vector methods (BERT-based Lexical Substitution). The static-swap baseline, together with perplexity matching, will show directly whether this confound matters for judges.
Why not just ask GPT-4o to rewrite each item for each culture? That is CALM's approach, and it is the real competitor. Its weakness is control: a free rewrite changes many tokens, which makes it hard to guarantee that support by the evidence is unchanged or to attribute a flip to the cultural marker. We will settle this empirically by comparing truth-preservation rates under human audit. If the two suites produce equally clean items and similar gaps, we will report that the cheaper rewrite suffices.
Why not just prompt the judge to "evaluate fairly as someone from region X"? Persona prompting behaves largely as token-level noise in respondents (Cultural Conditioning or Placebo?). It also tests something different: steering the judge rather than measuring its response to the content. We include it as a mitigation arm and expect it to move the placebo condition about as much as the cultural one.
Isn't this just entity rarity? This is the most serious threat. Hallucination judges plausibly distrust unfamiliar entities, and non-Western entities are rarer in web text. Frequency stratification and invented entities address it head on. If region effects vanish within frequency bins, we will report a long-tail effect rather than a cultural one. That finding still matters, because it shows who bears the cost.
Decisive Checks
The central result is $\Delta_r$ on grounded, frequency-matched, human-verified counterfactual items. We consider it established if it exceeds the placebo condition's flip rate, with confidence intervals from item-level bootstrap, for at least two region pools and at least three of the judges.
Three secondary checks follow:
- The gap should persist in the grounded setting. Persistence indicates the prior overriding the evidence; a gap confined to closed-book judging would be ordinary knowledge limits.
- The gap should grow from 7B–14B to 70B-class judges, consistent with a self-solving mechanism. A flat trend would weaken that explanation.
- The Western-norm-projection hallucinations should be missed more often than matched Western errors.
If none of these appear, we will have shown that, under the cleanest available manipulation, factual judging is robust to cultural framing. This would extend BiasScope's null with a method that could have detected the effect.
References
- BiasScope: Towards Automated Detection of Bias in LLM-as-a-Judge Evaluation
- Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge
- How Reliable is Multilingual LLM-as-a-Judge?
- One Token to Fool LLM-as-a-Judge
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- BAE: BERT-based Adversarial Examples for Text Classification
- Contextualized Perturbation for Textual Adversarial Attack
- Structure-invariant testing for machine translation
- BERT-based Lexical Substitution
- Bias Challenges in Counterfactual Data Augmentation
- Inferring Gender from Names on the Web: A Comparative Evaluation of Gender Detection Methods
- Cultural Conditioning or Placebo? On the Effectiveness of Socio-Demographic Prompting