AI in DCE campaign development: from concept to content
Every polio misinformation campaign starts with the same choice: which story will land? Which frame will move a worried parent in Kano, a father in Karachi, a grandparent in Kabul — and which one will quietly miss them?
The DCE team make these decisions under significant constraints. Campaigns must respond quickly to emerging narratives, work across different cultural and linguistic contexts, meet technical and safeguarding requirements, and make the best possible use of limited resources.
AI is helping DCE compare possible creative directions, identify potential weaknesses earlier, support localisation and manage complex production workflows. But the principle remains constant: AI accelerates the process; people make the decisions.
“In complex environments shaped by misinformation, information alone is not enough. We deliberately pivoted toward emotionally resonant, community-focused storytelling powered by real-time intelligence. This shift delivered a +33% average increase in vaccination intent across regions,reinforcing that strategic digital engagement, when grounded in trust and cultural insight, can meaningfully strengthen immunization outcomes.”
— Adnan Shahzad, Digital Community Engagement Global Lead, UNICEF
Listen, decide, make, measure
DCE's campaign cycle has four stages. The parent note covers two of them: listen (the social-listening machinery) and measure (automated analysis of how content performs once it is out). This piece is about the middle: decide and make.
In the decide stage, AI helps the team test ideas before anything is produced. In the make stage, it helps adapt one idea into something that actually fits the families it is meant to reach. Both rest on the same principle the team has held throughout the AI work: AI gets us further faster, but the call is always made by people. AI does several jobs at once:
- Quality and compliance. Automated pre-checks against polio communication guidelines, technical specifications, and language and format requirements — applied across hundreds of campaign assets.
- Localisation. Adapting one idea — visuals, copy, colours, framing — into culturally targeted versions for different families and markets.
- Reporting and analysis. Pattern detection in campaign data and draft executive summaries, structured for human review and decision.
- Production efficiency. Project-specific workflows that automate parts of production, approval, and deployment — keeping hundreds of assets consistent and moving them to market faster.
Both stages rest on the same principle the team has held throughout the AI work: AI gets us further faster, but the call is always made by people.
Two definitions of “persona” in DCE's AI work
The term persona is used in two different ways within DCE’s AI-supported work.
The first refers to audience profiles used to guide content development. An AI engagement tool can be grounded in profiles representing groups such as mothers, fathers, grandparents, health workers, faith leaders or community leaders. These profiles help the system consider differences in information needs, trusted messengers, barriers and preferred formats when proposing draft content.
The second application is synthetic persona testing.
In this approach, a language model is prompted to simulate a specific audience perspective—for example, a mother of three living in a particular context—and to react to an early campaign concept. The generated response may resemble qualitative feedback, but it is not an interview and should not be treated as testimony from a real person or community.
Synthetic personas do not possess lived experience. They generate plausible responses from statistical patterns in their training data and from the information provided in the prompt. That data may be incomplete, unevenly representative or affected by linguistic and cultural bias.
Their value therefore lies in hypothesis-generation, not validation.
They can help teams:
- compare multiple concepts using a consistent framework;
- identify possible cultural or emotional concerns;
- surface questions that researchers may want to explore;
- challenge assumptions held by the creative team; and
- prioritise which ideas should proceed to real-world testing.
Any proposed direction must still be reviewed by cultural advisers, country teams and relevant technical specialists, and tested with actual audiences where appropriate.
In practice: polio misinformation, May 2026
In May 2026, the team had four angles on polio misinformation on the table, deliberately spanning the spectrum of what might work in priority polio markets. The point was to map where leverage sits, not to choose from a list of equals.
The four angles were intentionally different:
- A playful one. What if we made it fun and emotional to not spread misinformation? Positioned at the edge of the spectrum on purpose — to map where tone becomes counter-productive.
- A moral one. What if spreading misinformation became a question of right and wrong?
- A consequence one. What if one careless share could be shown to lead, traceably, to a child going unvaccinated?
- A disease one. What if misinformation itself was the virus — something families have to protect against, together?
The boundary was mapped. The playful register read as trivialising across every market and gender. Children's health and a parent's daily weight are not the place for that tone. The line is now documented.
Three frames held strongly across markets. The moral, consequence, and disease angles all carried, each with its own conditions. Consequence, for example, lands as a story (a parent who regrets a mistake, an elder warning of a hidden danger) and falls flat as a lecture from a government voice. Same frame, different delivery.
A synthesis came into focus. The strongest version pulls the three working frames together, told as culturally resonant storytelling, and pushed through where families actually share content — WhatsApp voice notes, not only social media. That direction matched what the team's strategists had been moving toward. The synthetic layer confirmed it across markets, fast.
The campaign now moves into in-country qualitative testing with cross-market evidence already in hand: strategic instincts confirmed where they held, sharpened where they needed nuance, and the edges of the spectrum mapped. Field testing focuses on what is most likely to land.
What changes when this works
Three concrete shifts.
Time. What used to take weeks of recruitment and moderation now takes days. The cycle starts earlier, on stronger evidence, with more markets evaluated in parallel.
Cost. Testing four directions across three markets in parallel — without recruitment, fieldwork logistics, or print runs — costs a fraction of equivalent qualitative work. Producing localised versions of content with AI keeps the per-market unit cost low too.
Reach. When content is targeted properly — adapted in tone, visuals, and framing to the family it is meant for — it works harder. A misinformation campaign last year reached three to four times as many people as the same content distributed without that step. That is the most concrete demonstration we can point to so far.
The part that stays unmistakably human is the work that turns evidence into a campaign worth shipping. Cultural advisors, in-country researchers, and field teams remain the authoritative voices on cultural fit and operational realism. AI does not make those calls. It widens the space the team can bring to them, and gets there earlier.
