Digital and AI-enabled Social and Behavior Change: Snippets from the 2026 SBCC Summit

The Social and Behavior Change (SBCC) Summit took place in Panama earlier this summer under the theme “The Power of Connection: Reimagining Knowledge, Action and Equity in a Changing SBC Landscape.” 

There was a strong focus on how the landscape in which our organizations are working has rapidly changed over the past few years. Key sub-themes included challenging power structures and building relationships that prioritize the voices and leadership of those most directly impacted by the issues that Social and Behavior Change (SBC) work aims to address. Sessions related to digital SBC and AI were interlaced throughout the program.

Early in the week, I led a workshop on AI and SBC with Ana Mirzoyants (representing iMedia). I spent the rest of my time attending sessions to learn more about how organizations are using digital and AI in their work, and ended up with 90 pages of notes!

Below are some observations and snippets from the sessions that provoked my thinking*. I have not linked to all the organizations or individuals, but you can find more information on the Summit program or do an online search to find out more about their work. We are hosting an online event about these topics in early September; you’re welcome to join us!

Some key uses of AI for SBC

• Processing large volumes of text that human teams could not get through at speed, including comments, chatbot logs, subtitles, forum posts, rumor logs, field notes. For example, Big Cabal Media used AI to track sentiment and rumors across 11 Nigerian states in near real time and fed findings directly back to communications teams for message and campaign adaptation.
• Use of AI for tasks that standard research cannot accomplish. The World Bank’s MapGen initiative is using AI to analyze thousands of films and millions of online videos across multiple countries, measuring how gender is represented in media over time. The “norms in the wild” methodology — extracting and clustering comments from online forums like Nairaland — aims to surface what people actually think, not what they say when they know someone is listening. The idea is that this could be a way to address survey unreliability on sensitive topics due to social desirability, and that unmoderated digital spaces might offer something closer to real attitudes.
• Voice-based data collection in low-literacy, low-connectivity settings. HumanTruths tested the uptake and performance of an AI voice interviewer across multiple recruitment channels, namely random digit dialing, a local research panel, SMS, WhatsApp, and digital advertisements. The AI voice interviewer engaged caregivers of children under five in 10-minute AI-led interviews about beliefs and decisions around childhood vaccination. An AI agent can receive a WhatsApp photo of a child’s vaccine card, analyze the image, verify the number of doses, and follow up with the caregiver in real time. This enables record verification without requiring a human interviewer. The team evaluated response completeness, adherence to protocol, and participant experience. Participants’ most important criterion was feeling understood. The team was able to conduct 15,000 interviews in 5 days, with 88% of the completed calls reaching the “high quality” threshold. Comparative AI vs. human results are in the works.  

What AI cannot do

• Handle qualitative nuance. With qualitative data, nuance is the point. Off-the-shelf models do not have the language specificity or cultural grounding needed. One speaker gave the example that “Are you OK?” in Kenyan English is a greeting, not a distress signal, but AI trained on global English corpora gets this wrong. Using AI to code qualitative data without local language expertise built in produces something that looks like analysis but is not accurate or nuanced.
• Detect or appropriately respond to mental health crises. Multiple sessions raised this. How far do you let the AI go, and when does it hand off? What does a good handoff look like? Early warning detection for GBV, suicidal intent, and other safeguarding concerns has to be built in deliberately; it does not emerge from general language models. AI can help organizations manage volume, but humans should handle judgment.
• Reach the offline majority. 3.4 billion people are offline. It will take an estimated 39 years to close the connectivity gap (GSMA 2025). In Sub-Saharan Africa, fewer than 25% of the poorest 40% own a smartphone, with rural and female-led households further behind. WhatsApp bots and mobile apps are designed for the connected few. Viamo’s voice-first GenAI, which is accessible on any basic phone, toll-free, no internet or literacy required, tries to address this gap. There is quite a lot of interest in Voice AI and alternative ways to use AI to reach offline people.
• Drive community change. Chatbots for health behavior change are designed to work at the individual level. Multiple organizations showed how multi-pronged approaches are needed rather than a single channel; e.g., just a chatbot vs a chatbot plus community outreach plus health worker outreach plus social media plus traditional media. Several sessions shared evidence that community is the most important determinant of health behavior. This points to it being important to integrate chatbots into wider-reaching programs, not roll them out in an isolated way. 

Amplified bias: WEIRD and CURVEY 

• Gender representation and bias are a core issue with AI that better prompting cannot fix because the problem is enmeshed deep in the model and its underlying data. This is problematic for a field that is increasingly producing AI-assisted content (e.g., scripts, social media posts, training materials) yet that aims to address topics like gender equity and gender based violence. While above I noted the huge interest in using voice AI, there is evidence that it underperforms on women’s voices, low-resource languages, and non-standard dialects. HumanTruths, for example, tested eight different Swahili voices before finding models that worked well for their use case. Most commercial voice AI companies do not do this type of testing. Each new technology layer, as one participant put it, inherits and compounds the biases of the last.
• Digital data and AI are skewed towards the CURVEY. The concept “WEIRD” was coined to describe how psychological research is dominated by Western, Educated, Industrialized, Rich and Democratic worldviews. WEIRD has also been used to describe some of the inherent biases in AI. Our team at MTI has coined a complementary term (“CURVEY”) to reflect that in LMICs, digital data and AI bias skews towards the Connected, Urban, Resourced, Vocal, Extreme, and Young. It’s important to note that if your target population is primarily offline, digital analytics are not reliable ways to learn about your target users. Several sessions made the point about the mismatch between online and offline populations and the importance of designing with this in mind or treating these groups as separate.
• Social media bias and misinformation have increased. Upswell reported that Meta platforms no longer yield the deep signal they once did because the data that can be pulled is mostly from influencers who have a vested interest in driving content (again, “CURVEY” people dominate!). They also recommended giving up on the head-on fight against misinformation directly on platforms and to focus on building the peer trust infrastructure that makes communities more resistant to it..
• AI-generated summaries are a real and unresolved problem. Young people are increasingly getting health information from AI summaries that may contain misinformation. Additionally, organizations can no longer rely on search traffic or website visits as reach indicators because of AI summaries. I asked one group of panelists about this, as I had not heard anyone at the Summit talking about it. A speaker from Ipas Kenya said it is a big issue for them and could be a whole session on its own! For many, the way AI is changing search and SEO has not been identified as a problem, yet it’s something that our sector is going to need to figure out soon. (See this link about how to get Chatbots to pick up your content here).

Chatbots

• The privacy and non-judgment factor is real. Coach Mpilo, a project developed by PSI and Audere in South Africa, found that users were more willing to discuss sensitive HIV and TB issues with their AI coach than with a human provider. Nivi found the same across five countries on reproductive health. People often came in not knowing exactly what their question was, and the process of chatting helped them articulate the actual problem they were trying to solve. This has implications for how chatbots are designed. Privacy and space for on-device conversation and advice that avoids fear of stigma or showing vulnerability (including for men and boys) were important elements. Chattiness and lack of clear focus on a specific SBC goal can raise tensions with donors (who may want a more targeted focus) and also cost more tokens, but might yield better results.
• User data is routinely ignored by organizations. Nivi analyzed 9,714 real user messages and found their app’s menu structure was misaligned with what users actually needed and were looking for. Redesigning based on that analysis produced a 38% improvement in retention almost immediately. The behavioral data was there, but they had not been using it to inform their programming. This is a widespread issue I’ve been seeing since the “ICT and Mobile for Development days” with a fairly easy fix! (See this post from 2025 on how designers and MERL practitioners can work together)
• Tiered design outperforms any single modality. Viamo’s recommended architecture is menu-driven pre-recorded content first (co-designed with communities), then AI for questions that fall outside the menu, then human escalation for complex or sensitive issues. Every AI interaction has a financial cost; infinite AI responses are not viable because of cost. The goal is to get most questions answered by the cheaper layers, then move to AI which is more costly than IVR of menu based voice bots,  and reserve human time for what actually requires it. This brings MEL challenges as it’s very difficult to tease out effects of different integrated approaches in a hybridized product. RAES did try to unpack the question of which channel is doing what, finding that the two channels – offline and online – served different but complementary roles.
• Government endorsement matters more than most implementers acknowledge. Viamo spoke about Pakistan’s Khadija HPV helpline, a GenAI voice agent trained exclusively on Ministry-approved content. It reached nearly 157,000 people in three weeks, with 86% medical accuracy confirmed by independent testing. Government endorsement was central to adoption. (I remain curious whether a 14% error rate is acceptable for a health-focused initiative and how it compares to human-led medical information provision).

Unresolved tensions

The tensions can be organized into a few different buckets:

Over-investment in AI?

• Donors may be over-investing in AI tools at the expense of on-the-ground research. Several discussions raised this. AI seems fast, innovative, and cheap. People-centered qualitative research is slow and expensive. If funding shifts away from the latter, the evidence base for this field will get shallower. While the cost savings and efficiencies that AI affords appear to be real and improving; the risk of what gets defunded to pay for AI approaches is a big concern. Will the habitually excluded groups be left out again as face-to-face fall out of favor or are considered too expensive and slow?
• Where is the funding for high-quality services? Related to the bullet above, the importance of making it easy to take an offline action was mentioned. Social media awareness and SBC outreach need to point to a specific service so that people can easily take the next step. This means that investment in social media and AI SBC needs parallel investment in services and infrastructure. There is a fear that reduced funding for development overall will lead to more digital SBC without any accompanying quality service investment. This is not a new issue, it’s a carryover from the m4D days, and there is even less investment now. If all the funding goes to AI demand generation, where exactly are all the AI bots sending people who need a physical service? What happens when there is no funding for healthcare, vaccines, reproductive health, etc.? 
• The proliferation of tools is not obviously good. Multiple AI health chatbots, voice agents, and social listening platforms are being built in parallel, often solving the same problems with no shared standards or metrics, no interoperability, and no shared learning. The question of whether the field needs lots of specialized local tools or a smaller number of well-tested common platforms was raised but not resolved.

MEL challenges

• The measurement problem is unresolved. There seems to still be a lot of vanity metrics (reach, engagement) and need for better metrics and creative ways to understand impact when social media and AI are used to promote access to information and behavior change. Engagement — defined as reaching users and getting them to interact — was sometimes confused with impact when chatbots and social media interventions were presented. Many also noted that engagement is not enough to show impact. “Signals” were used to point to impact, but it seemed difficult and expensive for organizations to measure beyond the engagement stage. Definitions of “meaningful” engagement were not standard. Self-reported data on behavior change (did you go to the clinic?) is the primary outcome measure for most of these tools, and it carries its own biases. The field is still figuring out how to measure whether social media and chatbots actually change anything beyond the conversation itself. While some are finding ways to track connection with a health professional or physical action, this is still elusive for most organizations. How do we go from reach, engagement, clicks, clues, and insights related to behavior change to concrete metrics that someone took an action or that a social norm or attitude changed?

Critical thinking and awareness about potential harms

• There is low awareness of the downsides of casual and programmatic use of AI. SBC practitioners, faced with huge budget cuts and pressure to do more with less are turning to AI as a support tool. Practitioners are using AI extensively at the personal level, and AI adoption at organizations is often more ad hoc than strategic. There is insufficient understanding of how AI works and the underlying issues with AI, and this could lead to SBC programs embedding bias, not addressing hallucinations, and exposing sensitive personal or organizational data. Several of our workshop participants expressed gratitude that we discussed these concerns. Overall, I would love to see the sector be less cavalier about its use of AI and to develop deeper critical thinking about the complexities of using it.
• AI governance is a power question, not a technical one. Who controls the data centers, who sets the standards, who owns the training data, who benefits from the insights — these questions were raised in the AI plenary and in table discussions, and there are no easy answers. The framing of “responsible AI” in most global health contexts still centers technical guardrails more than structural power.
• AI causes health and environmental problems. There was no mention in the sessions I attended of the health impacts of AI (aside from aspects of mental health). Data centers have known health impacts and are being built in low-income areas that already have poor health indicators. They are also being built in areas that already suffer from drought and heat, and higher energy costs are often pushed onto local populations. Climate and other environmental effects of AI also went largely unmentioned. This tension needs to be addressed, considering that greater use of AI will lead to greater impacts on the very challenges the SBC community is aiming to address. We raised the potential of Small Language Models in our workshop, and this option should be explored and invested in more fully.
• There is hunger for discussions around safe and responsible use of AI. Judging by our workshop and comments heard throughout the conference, there is a real appetite for learning more about how to apply AI safely and responsibly. Some participants expressed a kind of relief that we named and raised ethical issues head-on in our workshop, saying they did not feel alone now in their concerns. At the same time, it seems like the majority of heavy AI users have no awareness of the practical and higher-level ethical issues with AI, or an understanding of alternatives to Big Tech. There can be a tendency in the SBC sector to want to be at the cutting edge of media and technology, in this case, with use of AI, with less attention to safety and privacy.
• Interest in Voice AI for research and for SBC programming is exploding, potentially with insufficient guardrails. Voice AI is a tool that could reach wider populations in a more engaging way. Several voice AI examples were shared at the conference, often combined with other approaches. Voice AI can allow for data collection rapidly at wide scale. At the same time, privacy and data re-use concerns were not adequately addressed in most cases and need to be raised, considering the vulnerability of the populations and lack of consent for re-use of voice for training AI models and systems. Issues with accuracy for certain demographic groups also need to be addressed. The potential and excitement about Voice AI should be supported and encouraged with sufficient attention to the possible harms.
• MTI’s SBC Learning Group should create space to discuss the above! The SBC Learning Group will work on developing out sessions. training, and potentially some support tools or guidance to help the community better integrate AI safely and responsibly. A topic for more exploration, for example, might be about how AI is changing search and how to improve/move on from SEO and what other tactics can address that. Voice AI is another topic that people are very interested in!

We’ll be running an online session on September 2nd, at 11am ET to share and discuss more on these topics. If you attended the SBCC Summit, we’d love to hear about your experiences! And if you did not, join us to get the download. RSVP here.

*AI Use Disclosure: I attended all the sessions I have written about and took my own notes. I created a “long” version where I copy-pasted the links to the sessions and the abstracts from the conference program into a Word document. I used Claude (Sonnet 4.6, medium) to clean up the document (fix font sizes, smooth out my notes, etc). I reviewed and edited the notes to be sure they captured what I heard and my own analysis of the “so what”. To create a shorter version, I asked Claude to pull out a synthesis of key points. I reviewed and edited these to improve them and added additional points that were missed and shared with my team at The MERL Tech Initiative. Kevin Hong and Isabelle Amazon-Brown added lots of interesting comments and drew out additional insights. I edited the report into a blog post without the use of AI.