Can you fact check in other languages ?

1. Performance is possible but uneven: models understand many languages, yet accuracy varies

Large models such as GPT-3.5 and GPT-4 demonstrate impressive multilingual understanding and can classify claims in multiple languages, but studies show systematic drops in reliability outside English and across different tasks, with English consistently outperforming other languages on many factuality metrics ^{[6] [1] [2]}. Benchmarks like X-FACT and later multilingual datasets reveal that automated systems struggle: top models achieve modest F-scores on multilingual verification tasks, indicating realistic limits to out-of-the-box multilingual fact-checking ^{[7] [8]}.

2. Translation and cross-lingual transfer are pragmatic workarounds — and sometimes boosters

A common pipeline translates non-English claims into English and runs English-trained fact-checking or retrieval systems; several studies find that translating into English often improves retrieval and matching for low-resource languages and can leverage high-resource tooling, though translation can also introduce meaning shifts that corrupt veracity checks ^{[3] [4] [5]}. Research on cross-lingual transfer and fine-tuning shows that tailoring models on multilingual examples or using English-heavy supervision yields measurable gains, but that these gains depend on language coverage in training data ^{[4] [9]}.

3. Retrieval of previously fact-checked items across languages is a high-value strategy

Rather than predicting veracity from scratch, systems that retrieve previously fact-checked claims (PFCD) across languages help human fact-checkers by surfacing relevant counters and evidence; recent work documents that multilingual LLMs can enhance cross-lingual retrieval and that translations plus cross-lingual embeddings are effective in practice ^{[3] [10]}. This approach reduces the burden of raw veracity prediction and plays to LLM strengths in matching and synthesis while leaving judgment to human reviewers ^[3].

4. Low-resource languages sometimes paradoxically show different error patterns

Surprisingly, some evaluations find a negative correlation between model accuracy and the amount of internet content for a language, suggesting LLMs may overfit noisy high-resource signals or that benchmarks capture different claim types across languages; nonetheless, broad reviews underscore persistent weaknesses for underrepresented tongues and topical diversity beyond political claims ^{[5] [11]}. Large-scale multilingual datasets such as MultiSynFact and X-FACT are attempts to close this gap by expanding claim-source coverage, but they also reveal how far automated systems remain from human-level, cross-topic reliability ^{[12] [7]}.

5. Hallucinations, stale knowledge, and evaluation limits are the clinchers

Multilingual fact-checking inherits core LLM failure modes: hallucination (fabricated citations or facts), outdated knowledge, and brittle sensitivity to phrasing; papers explicitly call out that metrics and hallucination detectors are weaker in multilingual settings, and that veracity prediction is especially hard for novel factual claims ^{[2] [1] [11]}. Consequently, the literature recommends combining model outputs with retrieval, human verification, provenance checks, and transparent uncertainty signals rather than treating LLM outputs as definitive ^{[4] [3]}.

6. The landscape: what this implies for real-world use and hidden agendas

Practical deployments use layered systems—translate when useful, retrieve prior fact-checks, use multilingual fine-tuning, and route high-uncertainty cases to humans—because research shows no single LLM reliably replaces multilingual human fact-checkers yet ^{[4] [3]}. Academic incentives and vendor messaging sometimes overemphasize raw multilingual competence; scrutiny of datasets and tasks reveals that evaluation often focuses on narrow topics or languages, which can inflate perceived readiness and favor commercial narratives ^{[11] [5]}.