Mar 10, 2025

AI Hallucinations in Healthcare: Cross-Cultural and Linguistic Risks of LLMs in Low-Resource Languages

Peace Silly, Zineb Ibnou Cheikh, Gracia Kaglan

This project explores AI hallucinations in healthcare across cross-cultural and linguistic contexts, focusing on English, French, Arabic, and a low-resource language, Ewe. We analyse how large language models like GPT-4, Claude, and Gemini generate and disseminate inaccurate health information, emphasising the challenges faced by low-resource languages.

Download

Review Project

See Code

Reviewer's Comments

This project seeks to assess risks from hallucination in healthcare settings as it varies by languages (those of "low" vs "high" resource). This is an important topic with clear safety implications for detecting hallucinations in impactful settings. The cross lingual checker is a really neat tool with practical utility.

I have several suggestions for how to improve the reported work.

- the introduction refers to psychology and cognition but it is unclear how those connect to what is studied here

- a definition of a 'high resource' language should be provided (is it about of training data, or something else?)

- the question set is relatively small; it could also be more related to naturalistic queries used in healthcare settings

- while hallucination is well defined for factual questions, reasoning and cultural specificity are much more subjective and therefore seems like a very different construct. Perhaps, alignment rather than hallucination. But those items should not be interpreted as reflecting hallucination.

- it should be described how the test items were selected and validated

- for LLM response ratings, the number of raters, the instructions given to them, and their inter-rated agreement should be reported

- analyses (e.g, comparisons of models) should be performed with statistical tests.

- Plots should show aggregated data over specific prompts to better reveal trends in the data. Bars for different languages should be presented side by side for better comparison. Right now the differences are difficult to see.

- Data for English and French not shown.

Thank you for submitting your work on AI Hallucinations in Healthcare: Cross-Cultural and Linguistic Risks of LLMs in Low-Resource Languages - it was interesting to read. Please see below for some feedback on your project:

1. Strengths of your proposal/ project:

- Your paper quickly outlines how this project contributes to advancing AI safety.

- The paper is interesting, timely and topical, and focused on an under-researched area

2. Areas for improvement:

- It would be helpful if you defined/ expanded upon some of the key terms that you are using (e.g., global catastrophic risks; AI hallucinations), as well as contexts (e.g., ‘AI-driven healthcare decisions’).

- A clearer threat model and examples would greatly strengthen the paper. E.g., you mention that ‘AI hallucinations in healthcare, which poses significant risks to global health security’ - how/ in what ways? Which areas of healthcare? Could you provide citations for the results/ implications you mention under Threat Model and Safety Implications (p.2)

- Some sentences need to be expanded - e.g., ‘Existing research on AI hallucinations often focuses on technical aspects’ - what does this mean? Technical aspects of what?

- The methods section could be a bit clearer, which would help anyone looking to replicate your methods.

Strengths:

-Transparency and Reproducibility: The methodology is clearly documented, and the study can be easily replicated, which is a significant strength.

-Relevance: The paper addresses a critical issue in AI safety—hallucinations in low-resource languages—with clear implications for healthcare.

-Systematic Approach: The use of a multi-step evaluation framework (factual accuracy, safety recommendations, uncertainty transparency) is well-structured and effective.

-Policy Recommendations: The inclusion of policy recommendations adds practical value to the study.

Areas for Improvement:

-Limitations and Negative Consequences: The paper does not explicitly discuss the limitations of the study or potential negative consequences of its findings. Adding this would strengthen the analysis.

-Mitigation Strategies: While the paper investigates risks, it does not propose concrete solutions or mitigation strategies. Including this would enhance its impact.

-Discussion and Conclusion: The discussion section is brief and does not explore future implications, next steps, or shortcomings. Expanding this would provide a more comprehensive conclusion.

-Comparative Analysis: Including a more thorough comparison between high-resource and low-resource languages would add depth to the analysis.

Suggestions for Future Work:

-Conduct larger-scale experiments with more languages and questions to validate the findings and improve generalizability.

-Explore mitigation strategies for hallucinations and biases in low-resource languages.

-Investigate the framework’s performance in diverse cultural and regulatory contexts.

-Address practical constraints and limitations in future iterations of the study.

Cite this work

@misc {

title={

AI Hallucinations in Healthcare: Cross-Cultural and Linguistic Risks of LLMs in Low-Resource Languages

author={

Peace Silly, Zineb Ibnou Cheikh, Gracia Kaglan

date={

3/10/25

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Recent Projects

View All

Apr 27, 2026

OliGraph: graph-based screening of large oligopools

Existing synthesis screening tools cannot evaluate short oligonucleotide pools, whose overlapping fragments can be reassembled into regulated sequences via polymerase cycling assembly (PCA) yet fall below gene-length detection thresholds. We present OliGraph, an open-source tool that constructs a bi-directed overlap graph from an oligonucleotide pool and extracts contigs for downstream gene-length screening. An optional PCA mode retains only cross-strand overlaps consistent with PCA chemistry. We validated OliGraph in a blinded study across ten simulated pools (70–9,184 oligonucleotides, 30–300 bp) spanning four risk categories. BLAST screening of individual oligonucleotides failed to identify sequences of concern in most pools: three returned zero hits, and vector noise obscured true positives in the remainder. After OliGraph assembly, contig-level BLAST matched the longest assembled sequences (up to 1,905 bp) to sequences of concern at 97–100% identity. In one pool, assembly collapsed 1,634 individual BLAST results into 10 hits from a single contig, all assigned to the same source organism. PCA mode correctly distinguished assemblable from non-assemblable fragments within the same pool. Two pools with no assemblable structure yielded no contigs. OliGraph processed all pools in under 0.2 seconds, fast enough for real-time order screening and consistent with proposals to bring oligonucleotide orders within the scope of synthesis screening regulation.

Apr 27, 2026

BioRT-Bench: A Multi-Attack Red-Teaming Benchmark for Bio-Misuse Safeguards in Frontier LLMs

Frontier AI laboratories are expected to maintain safeguards against biological misuse, but whether deployed models actually refuse bio-misuse queries under adversarial pressure is largely unmeasured in the public literature. We introduce BioRT-Bench, a benchmark that runs four attack methods (direct request, PAIR, Crescendo, and base64 encoding) against four frontier models (Claude Sonnet 4.6, GPT-5.4, DeepSeek V4-flash, Kimi K2.5) across 40 prompts spanning five biosecurity-relevant categories. Responses are scored by a calibrated judge extending StrongREJECT with two bio-specific dimensions: specificity and actionability. We measure Attack Success Rate (ASR), where 0 means the model fully refused and 1 means it provided specific, actionable bio-misuse content. Our results reveal a sharp robustness divide: Chinese frontier models (DeepSeek, Kimi) have under 5% refusal rates even under direct request (ASR 0.88 and 0.79), while Western models (Claude, GPT) maintain substantially stronger safeguards (ASR 0.15 and 0.16). Crescendo is the most effective attack across all models, both in bypassing refusal and in eliciting actionable content. Claude Sonnet 4.6 is the most robust model tested, achieving 100% refusal against base64-encoded prompts.

Apr 27, 2026

PROTEUS (PROTein Evaluation for Unusual Sequences): Structure-Informed Safety Screening for de novo and Evasion-Prone Protein-Coding Sequences

AI protein design tools like RFdiffusion, ProteinMPNN, and Bindcraft make it trivial to produce low-homology sequences that fold into active, potentially hazardous architectures. However, sequence homology-based biosafety screening tools cannot detect proteins that pose functional risk through structurally novel mechanisms with no sequence precedent. We present a tiered computational pipeline that addresses this gap by combining MMseqs2 sequence alignment with structure-based comparison via FoldSeek and DALI against curated toxin databases totaling ~34,000 entries. AlphaFold2-predicted structures are screened for both global fold similarity (FoldSeek) and local active/allosteric site geometry (DALI), capturing convergent functional hazards that sequence screening misses. The pipeline was validated against a panel of toxins, benign proteins, structural mimics, and de novo-designed Munc13 binders, as well as modified ricin variants with residue substitutions. We additionally tested robustness to partial-synthesis evasion, where a bad actor submits multiple shorter coding sequences intended for downstream reassembly into a full toxin-coding gene. We found that while sequence-based screening did not identify any de novo ricin analogues with high certainty, the combined pipeline with FoldSeek and DALI identified all 24 tested de novo ricins as toxic.

Apr 27, 2026