Responsible AI in Global Health: Solutions from the Global South

This prerequisite encompasses the development of robust regulatory and policy frameworks that ensure the safe, ethical, and effective use of AI in health systems, balancing the need for innovation with regulatory compliance to protect patient rights and public health.

Many countries lack dedicated legislation for the use of AI in healthcare. Instead, they rely on broader medical device regulations that fail to address AI-specific challenges, which increases risks related to data privacy, accountability, and equitable access. 

Best practices:

• Enforce data protection laws
• Promote transparency in AI decision-making
• Generate clinical evidence for AI tools
• Establish clear accountability lines
• Develop ethical guidelines with human oversight
• Research pricing regulations to improve accessibility
• Foster collaborative interoperability standards from a Global South perspective

Using high-quality, representative, and disaggregated data when developing responsible AI systems is essential for effectively addressing diverse health needs.

Low-quality data can result in various forms of bias that compromise the fairness and accuracy of AI-driven decisions. These issues are exacerbated by structural weaknesses in data availability, governance, and the underrepresentation of linguistic diversity, particularly in low-resource settings.

Best practices:

• Address poor data quality stemming from under-representation, misrepresentation and over-representation
• Strengthen robust data triangulation during model training
• Ensure secure interoperability of data systems
• Assess the availability and reliability of local datasets
• Promote cross-border data sharing with strong governance practices

Because health outcomes for women, men, and gender-diverse individuals are shaped by biological, social, cultural, and political factors, AI-driven health solutions must actively address gender disparities and systemic inequities.

AI technologies risk perpetuating and amplifying existing inequities if their design and implementation do not consider systemic oppression, exclusion, and gender norms. Without deliberate action, AI solutions may reflect the dominant perspectives of those in power, resulting in both visible and invisible forms of discrimination. 

Best practices: 

• Ensure equitable access to AI-enabled tools and connectivity
• Understand how social and gender norms influence AI solutions and regulations, alongside health-seeking and health provider behaviors
• Address impacts of data biases and underrepresentation of certain groups
• Design solutions to support individuals with disabilities
• Mitigate bias and exclusion in diagnosis and treatment
• Diversify pool of locally trained talent across disciplines

Aligning AI solutions in healthcare with ethical principles that protect human rights is paramount, along with promoting transparency and ensuring equitable outcomes.

Current AI systems have been relying on granular personal data more and more. For this reason, safeguarding privacy and minimizing harm from data misuse are crucial measures for preventing the exacerbation of health inequities. 

Best practices:

• Align AI solutions with human rights principles
• Minimize bias to ensure fair treatment across user groups
• Improve transparency and clarity of explanation
• Ensure appropriate human oversight in decision-making processes
• Enhance data security to protect privacy
• Reduce the environmental impact of AI technologies

Prioritizing locally-led research and decision-making ensures that experts and institutions from the Global South drive the translation of evidence into relevant policies and practices.

Addressing historical power imbalances between the Global North and Global South is essential to give those most affected by AI and global health policies a meaningful role in shaping them.

Best practices:

• Strengthen research ecosystems in the Global South
• Support lead authorship from local institutions
• Prioritize strengths & needs of Global South-based organizations in partnerships
• Embed equity principles in multi-partner collaborations to ensure fair resource distribution and inclusive decision-making

The AutoAI-Pandemics hub in Brazil developed Dominique, an AI-powered conversational assistant that combats disinformation—such as that on vaccine use—by analyzing statement truthfulness, promoting fact-checking, and providing accessible verification tools. Built using a GRU-model trained on a Portuguese news corpus, Dominique achieved a 95.7% accuracy rate and is available in Portuguese and English, earning international recognition for enhancing public information consumption.

Responsible AI Elements: The project prioritizes transparency, fairness, and ethical information sharing, with a user-friendly design that promotes critical thinking and reduces vulnerability to misinformation.

The NatallA project in Guatemala is developing an AI-enabled tool to help midwives perform fetal ultrasounds without the need for specialized personnel, helping address maternal health disparities among Indigenous women. Leveraging convolutional neural networks (CNNs), the tool extracts diagnostic features from ultrasound videos, enabling decentralized primary health care in local languages and cultural contexts.

Responsible AI Elements: The project ensures transparency and inclusivity by involving Indigenous communities, protecting data privacy, and using representative datasets to minimize bias.

A multidisciplinary team at Eminence Associates for Social Development in Bangladesh developed an AI-enabled tool to assist non-specialized healthcare providers in detecting prenatal and perinatal depression. By using a graph neural network model, the tool assists these providers in analyzing visual, facial, and acoustic cues from pregnant and postpartum women. This intervention complements traditional PHQ-9 assessments and reduces bias in the data collected.

Responsible AI Elements: The solution incorporates human-in-the-loop validation, prioritizes patient privacy, and uses culturally-relevant datasets to ensure fairness and empathy in mental health diagnostics.

Researchers at the University of Ghana developed an AI-powered solution to improve sexual and reproductive health (SRH) information for adolescents with hearing, speech, and visual disabilities. Using fine-tuned large language models (LLMs) like ChatGPT and Gemini and based on surveys with 400 students, the AI system generates culturally-sensitive, age-appropriate content.

Responsible AI Elements: The project addresses inclusivity through offline options, engages stakeholders for content validation, and ensures bias mitigation through culturally-informed prompt engineering of the AI model.

The Medical Rescue Association of Turkey developed an AI-powered chatbot to improve SRH knowledge among refugee women, overcoming barriers such as language, xenophobia, and digital literacy. The tool provides personalized, culturally-sensitive advice, accessible via WhatsApp, and engaged 110 users with 248 conversations in its beta phase.

Responsible AI Elements: The chatbot emphasizes ethical data use, user privacy, and adaptability to meet users' cultural and situational needs.

Jacaranda Health’s PROMPTS platform provides real-time, personalized maternal health information in English as well as Swahili, Sheng, and five other less common African languages. By classifying users into vulnerability segments, the tool provides health guidance while collecting socioeconomic data to address maternal health disparities.

Responsible AI Elements: PROMPTS employs open-source models, ensures human-in-the-loop validation, and incorporates user feedback to improve fairness and usability.

To aid malaria control efforts, the Kwame Nkrumah University of Science and Technology (KNUST) developed a deep learning AI model that classifies mosquito species based on wingbeat sounds. Trained on over 25,000 recordings, the model achieved 92% accuracy identifying species, offering a non-invasive alternative to traditional identification methods.

Responsible AI Elements: Community engagement and inclusive consultations ensure the tool addresses local needs and is accessible to vulnerable populations.

The Health Research Institute of the Universidad Autónoma de Santo Domingo is developing an AI-enabled early warning system to predict mosquito-borne diseases such as dengue and Zika viruses. By integrating epidemiological and environmental data, the tool provides timely alerts to improve public preparedness and response.

Responsible AI Elements: The project incorporates community input, ensures transparency in data use, and focuses on equitable access to outbreak information.

A team at the Universitas Gadjah Mada is developing an AI-enabled prediction model to enhance dengue surveillance by integrating environmental and syndromic surveillance data. Working with Indonesia’s Ministry of Health, the project aims to predict infection rates and improve data quality in healthcare facilities.

Responsible AI Elements: Stakeholder engagement ensures transparency, alignment with local health needs, and ethical use of medical records data.

Wits University and iThemba LABS developed a cost-effective, AI-powered air quality monitoring network to combat pollution in Gauteng province, South Africa. Affordable sensors provide real-time data visualized through dashboards to guide interventions against respiratory and cardiovascular diseases.

Responsible AI Elements: The project makes data publicly accessible, incorporates community feedback, and ensures transparency in AI predictions.

The Institut Pasteur de Tunis, in partnership with the Ministry of Health, developed an AI-enabled dashboard to monitor wastewater for high-priority pathogens and antimicrobial resistance. Covering 21 treatment plants, the tool supports outbreak prediction and resource allocation.

Responsible AI Elements: The model promotes transparency, ethical data use, and equitable resource distribution, particularly benefiting marginalized communities.

Jimma University developed PolioAntenna, a mobile app leveraging deep learning to enhance acute flaccid paralysis (AFP) surveillance. Trained on 428 images from six Ethiopian regions, the AI system automates detection and supports early polio response efforts.

Responsible AI Elements: The project involves local health workers, prioritizes gender-inclusive data collection, and ensures culturally-sensitive AI model deployment.