How DCE Uses AI — Part 2: From the Machines to the Field

Vaccine decisions used to be shaped across public and private information spaces: on social media, in community meetings, through mass media, and in conversations between families and health workers. That’s still true. However, new actors have recently entered these spaces, gaining every day more traction and trust: a mother asking ChatGPT whether the polio vaccine is safe, a father copy-pasting a WhatsApp rumour into Gemini for a second opinion, or a health worker checking Grok before a campaign briefing.

These conversations are invisible to conventional monitoring. Private conversations with AI assistants cannot be directly observed. And yet the answers they provide may influence perceptions at the moment a vaccination decision is being made.

This is the second of two parts. In the companion piece, “How DCE Uses AI — Part 1: Reading the Conversation,” we showed how DCE listens to and makes sense of the human conversation about vaccines. Here we turn to the machines — and to what DCE does next.

Listening to the machines

Generative AI is becoming part of the information environment in which vaccine decisions are made. Monitoring how AI systems communicate about immunisation is therefore an emerging component of DCE’s work.

DCE does not access private conversations between users and AI assistants. Instead, it uses two complementary approaches.

First, structured test prompts are submitted regularly to systems such as ChatGPT, Gemini, Claude and Grok. This makes it possible to compare how different models and model versions answer selected questions about childhood immunisation.

Second, DCE monitors publicly available examples of AI interactions, including shared screenshots, copied responses and online discussions about answers produced by AI systems.

The analysis examines:

how benefits and risks are framed;
which evidence, institutions and authorities are referenced;
whether answers remain consistent across languages and models;
how systems respond to misleading premises or conspiracy narratives; and
where factual gaps, excessive simplification or potentially harmful ambiguity may appear.

This does not reveal the full universe of private AI use. It provides a structured way to evaluate an increasingly influential source of health information that remains largely outside traditional social and media monitoring.

From detection to action

AI does not only support the identification of risks. It can also help teams design, test and adapt engagement responses.

The system supports fast design and deployment of tailored engagement campaigns, drawing on pre-collected audience profiles, messenger credibility mapping, platform-specific engagement rules, and a dynamic set of personas (parents, grandparents, faith leaders, community leaders) across contexts (urban, rural, displaced, nomadic).

Synthetic personas can support early concept testing by helping teams examine how different audiences might interpret a message, what questions it may provoke and where language may be confusing or counterproductive. This is an early validation step, not a replacement for testing with real communities or for specialist review.

During active campaigns, AI connects to real-time data streams from digital interventions to identify where and when adjustments are needed. If content is bouncing on TikTok but performing on YouTube, the system flags it. If one district is over-reached while another is under-served, it recommends redistribution. The loop from insight to action tightens from days to minutes.

Turning data into accessible guidance

DCE’s AI-supported tools can generate analytical briefs on demand or according to a predefined schedule — covering any time period, any geography, any combination of risk, misinformation, influencer, or campaign data.

Reports can include custom charts, referenced content, geographic breakdowns, and demographic analysis. For recurring reports, teams can set templates, frequency, and distribution lists. Recurring products can be configured using standard templates, reporting frequencies and distribution arrangements. This allows teams to spend less time compiling information and more time interpreting findings and deciding what to do next.

Built for speed, built for independence

The entire ecosystem runs on fully automated, open-source data pipelines with end-to-end latency of 1–2 minutes. During an outbreak, that speed is the difference between reacting and leading.

DCE’s AI agents are trained on operational guidelines, the most recurring risk narratives from Polio Pulse, ten years of polio SBC literature from Poliokit, UNICEF style guides, two years of social listening data and digital campaign content, and real-time monitoring feeds including Poliokit analytics, campaign data, and uInfluence contributions.

Looking ahead, DCE is developing a fully local, standalone AI engine — owned and governed internally, protected from external interference and model drift. Development is supported through agile, pro-bono technical arrangements. Design, governance, and day-to-day operations remain fully inside DCE. The goal: an AI capability that belongs to the programme, not to a vendor.

How we work

No cost to countries. Full deployment is available at no additional cost, with high local customization and ownership, to all polio outbreak, endemic, and surveillance countries.

From dashboards to conversations. DCE is replacing complex dashboards with AI-supported interfaces that translate data into plain-language insights, usable regardless of digital or data literacy level. The Infodemic Assistant is the first of these tools. Read more: Meet the Infodemic Assistant →

From dashboards to control rooms. DCE is moving toward real-time control rooms that give emergency coordinators and leadership continued monitoring and evidence-driven decision support during epidemic management.

Data protection. Data is accessed and processed in accordance with relevant licences, agreements and UNICEF data-protection requirements.

Sustainable by design. DCE is building internal capacity and reusable infrastructure to reduce dependency on short-term contracts and isolated technical solutions.