Digital Tools, Dark Patterns, and Unintended Consequences: Tech-Mediated SBCC Failures

En 2025, le Ministère de la Santé tunisien, la Tunisie a introduit le vaccin contre le VPH dans son programme national, visant les filles de 6e année, avec l'appui de l'UNICEF et de l'OMS. Malgré une certaine confiance des parents dans la vaccination scolaire, la campagne a échoué face à une vague massive de désinformation numérique. Une campagne coordonnée d'astroturfing a inondé les réseaux sociaux de contenus alarmistes et répétés, créant l'illusion d'un rejet populaire généralisé. Cette stratégie a faussé la perception sociale, affaibli les normes favorables et renforcé la défiance, particulièrement dans un contexte post-COVID marqué par la méfiance envers les institutions sanitaires.
L'absence d'une riposte institutionnelle claire et rapide a aggravé la situation. Le Ministère de la Santé n'a pas occupé l'espace numérique avec suffisamment de visibilité et de cohérence. Le calendrier de mise en œuvre, en pleine période post-Ramadan et vacances scolaires, a également nui à l'efficacité. Résultat : une couverture vaccinale de seulement 3 %, très loin de l'objectif de 90 %.
L'expérience tunisienne met en évidence la nécessité pour d'autres pays de considérer la lutte contre la désinformation comme un pilier central des campagnes de communication. Cela implique des SOPs clairs pour les situations de crise, une stratégie digitale proactive avec ambassadeurs et influenceurs, la réduction de la visibilité des auteurs de désinformation, mais aussi une communication interpersonnelle renforcée. Enfin, la mobilisation des médecins, leaders religieux, médias et acteurs communautaires est indispensable pour bâtir un front collectif contre la désinformation et rétablir la confiance.

Keluarga Berimun launched WhatsApp Groups and a 24-hour hotline in suburban Depok and Bekasi in November 2024 to help parents who were hesitant about immunization. Six months in, the contrast was stark: WhatsApp Groups generated 1,200–3,500 monthly messages, while the hotline stayed stuck at only 150–200 chats. Community health workers kept steering families toward WhatsApp Groups simply because both they and caregivers were already familiar with them. A "hotline" felt less intuitive.
We simplified the chatflow, removed long bot sequences, and allowed users to reach a live agent immediately. At the same time we promoted the hotline in multiple parenting communities on Facebook and provided direct links through community influencers. We also refreshed the copies used across platforms to make the service clearer and more inviting. But despite these efforts, hotline traffic barely moved.
The breakthrough came from adjusting our delivery. Beginning June 2025, we shifted our ad buying approach: Meta's Click-to-WhatsApp (CTWA) and Google's Demand Gen were added on top of impression and click objectives. CTWA allowed users to open a chat instantly. Meanwhile Demand Gen provided precise targeting across platforms, placing the hotline in front of caregivers already seeking health information.
Hotline conversations jumped from 205 in March to 950 in September. What this revealed was that the issue had never been caregiver interest, but the way the promotion was delivered. The experience shows every element must be tested in parallel as content, user flow, targeting, and ad delivery can each create silent bottlenecks.

Keluarga Berimun is a digital campaign designed to increase childhood immunization coverage in Depok-Bekasi through a social behavior change communication approach. One of its key strategies is leveraging multichannel decision support systems WhatsApp support group to transform hesitant parents into confidently vaccinate their children. 


However, measuring behavioral change through WhatsApp is challenging as it is part of the "dark social" ecosystem. Manual analysis is time consuming and resource intensive, highlighting the need for scalable approaches to extract behavioral insights.


To bridge this gap, the team piloted data mining using Large Language Models (LLMs) to analyze WhatsApp conversations and identify behavioral evidence. Early findings revealed both potential and limitations: while LLMs could categorize tone of voice and detect positive intent toward immunization, they often misclassified multi-topic chats and required human quality control.


The failure was not in the choice of technology, but in underestimating the preparatory work needed for the artificial intelligence (AI) to be an effective partner. It demonstrated that prompt design and data structure are as important as the AI model itself. LLMs perform better when given complete conversation context rather than isolated chat bubbles, supported by well-structured and indexed data. Refining prompts with clear analytical objectives and "golden examples" for classification improved consistency across iterations, with the margin of error predicted to decrease from the initial range of 20~40% to around 10~20%.


LLMs, therefore, is not a replacement for human judgment but a partner in generating credible, timely evidence in "dark social" ecosystems.

South Africa faces an ongoing challenge in retaining clients on HIV treatment. To address this, we developed an AI-powered, WhatsApp-based health coach to provide users with private, empathetic, personalized support. We aimed to integrate the AI coach into routine clinical workflows by asking nurses to introduce it to clients and share informational materials during consultations.


Although we observed interest and enthusiasm among healthcare providers following AI coach orientation sessions, subsequent user uptake was low, prompting us to look further into potential barriers to referral.


In-person discussions revealed several key barriers: some providers missed the initial training due to other responsibilities, once-off training was insufficient to build confidence in explaining the tool, providers had not engaged with the tool themselves, the group trained did not include client-facing non-clinical staff (counsellors, clerks, security guards), and the referral process was seen as too time-consuming in relation to other demands and priorities.


Solutions included engaging non-clinical staff, providing refresher trainings, offering multiple training slots on different days and times, leveraging waiting room health talks to reach many clients at once, conducting on-the-job shadowing and feedback, streamlining the referral process, sharing data on indicators that providers care about (e.g., retention of users in care), and creating targets and incentives.


The experience underscored the gap between provider intention and action and the importance of designing for providers' realities-simplifying workflows, empowering teams, aligning with their priorities, and ensuring they see tangible value. Otherwise, even well-conceived interventions risk limited adoption.