TLDR: A research paper titled “Between Myths and Metaphors: Rethinking LLMs for SRH in Conservative Contexts” explores the challenges Large Language Models (LLMs) face in understanding sexual and reproductive health (SRH) communication in conservative, low-resource contexts like Pakistan. The study, conducted by Ameemah Humayun, Bushra Zubair, and Maryam Mustafa, identifies a unique framework of indirect communication strategies, including euphemisms, colloquialisms, figurative language, myths, and gestures. It reveals that current LLMs struggle with semantic drift, polysemy, and providing culturally appropriate, easy-to-understand explanations. The paper proposes design recommendations for culturally-situated SRH communication, advocating for synchronous terminology management, designing for miscommunication as a default, and multimodal privacy-conscious approaches to create AI that adapts to communities rather than the other way around.
A recent study delves into the complex world of sexual and reproductive health (SRH) communication in conservative, low-resource countries, specifically focusing on Pakistan. The research highlights a critical challenge: while large language models (LLMs) hold immense promise for improving healthcare, they often falter when faced with the indirect, culturally nuanced language used in such sensitive contexts.
Maternal deaths remain a significant issue in low-resource countries, with Pakistan being one of the top contributors. Many of these deaths are preventable, and LLMs could play a vital role in automating health communication and risk assessment. However, the way SRH is discussed in conservative societies often relies on euphemisms, metaphors, and indirect phrasing, which can obscure meaning and complicate AI-based interventions.
Understanding the Communication Landscape
The study, titled Between Myths and Metaphors: Rethinking LLMs for SRH in Conservative Contexts, was conducted by Ameemah Humayun, Bushra Zubair, and Maryam Mustafa from Lahore University of Management Sciences, Pakistan. It involved a two-stage approach: first, analyzing real-world communication data from clinical observations, interviews, and focus groups with clinicians and patients; second, evaluating how five popular LLMs interpreted this data.
The researchers identified a unique framework for SRH communication, categorizing it along two axes: the ‘referential domain’ (what is being discussed, e.g., symptoms, procedures, conditions, processes, anatomical parts) and the ‘communicative approach’ (how it’s expressed). The communicative approaches include:
- Euphemisms: Polite ways to discuss sensitive topics, like referring to sexual intercourse as ‘meeting the husband’.
- Colloquialisms: Informal, standardized medical shorthand, such as ‘big operation’ for a C-section or ‘small operation’ for an episiotomy. These often exhibit ‘semantic drift,’ where a general term takes on a very specific medical meaning within the context.
- Figurative Language: Descriptive metaphors used by patients with limited medical vocabulary, for example, ‘inside is falling out’ for vaginal prolapse or ‘stringy pieces of fat’ for white vaginal discharge.
- Myths and Misconceptions: Folk beliefs or medically inaccurate explanations, like ‘athra’ (a curse of miscarriages) or believing white vaginal discharge means ‘bones dissolving’.
- Gestures: Non-verbal cues and vague expressions that are unintelligible without physical context, such as pointing to an area and saying ‘hurts there’.
LLMs Struggle with Nuance
The evaluation of LLMs (LLaMA 3.2, Gemma 3, GPT-OSS, GPT-4o, and Claude Sonnet 4) revealed significant challenges. While proprietary models like GPT-4o and Claude performed better than open-source ones, all struggled with the intricacies of indirect communication. Key issues included:
- Polysemy: LLMs often fixed on only one meaning of a term, failing to recognize other clinically relevant interpretations. For instance, ‘rasoli’ (tumor/fibroid) is often used by patients to refer to cysts or ectopic pregnancies, but models typically only recognized fibroids.
- Semantic Drift: Models failed to understand terms like ‘is the child normal?’ when patients meant to ask about the mode of delivery (vaginal vs. C-section), or ‘small operation’ referring specifically to an episiotomy.
- Overly Technical Explanations: Even when correct, LLM responses were often too complex and filled with medical jargon, making them incomprehensible to low-literacy patients.
- Persona Inconsistencies: Models sometimes failed to maintain the instructed female gynecologist persona, occasionally using masculine forms of address.
- Linguistic Instability: Lower-performing models often produced incoherent text, mixed languages, or even switched to entirely different scripts, rendering the information unusable.
Also Read:
- Indian Riddles Uncover LLM Overconfidence and Multilingual Reasoning Gaps
- Guiding AI’s Moral Compass: A Five-Step Framework for Diverse Value Alignment
Designing for Real-World Communication
The study concludes with crucial design recommendations for future LLM-based SRH interventions:
- Synchronous Terminology Management: AI systems should recognize and manage multiple coexisting terminology systems—formal medical terms, formal euphemisms, and informal clinical practice terms—and adaptively map them based on the communicative context and patient’s health literacy.
- Default Miscommunication Architecture: Instead of assuming successful communication, LLMs should be designed with miscommunication as the baseline, incorporating active verification protocols to clarify ambiguous terms without questioning patient credibility.
- Multimodal Privacy-Conscious Design: Given the prevalence of gestural communication and privacy concerns around video, interfaces like body diagrams where patients can mark problem areas could be a practical solution to accommodate non-verbal cues.
This research underscores that effective healthcare AI in culturally complex settings must adapt to community communication practices rather than forcing communities to conform to standardized medical discourse. It calls for a fundamental rethinking of LLM design, prioritizing equity and real-world linguistic realities in healthcare technology.
