By Alberto Cottica and Lazar Pop Ivanov

We are no strangers to warnings about how information pollution erodes public trust, deepens social and political divides, and limits access to accurate, reliable knowledge. UNDP’s 2023-24 Human Development Report highlights the growing challenge of digital communication in already polarized societies, where algorithms amplify division and further fuel misinformation. The UN has similarly cautioned that mis/disinformation fuel conflict, threaten democracy, and even hinder climate action. In this context, the erosion of the integrity of the information ecosystem risks compounding existing vulnerabilities and standing in the way of the Sustainable Development Goals (SDGs). Strengthening information integrity is therefore fundamental to development itself. Recognizing this, UNDP’s new Strategic Plan (2026 -2029) calls for scaling efforts to confront digital harms, including algorithm-driven polarization and online disinformation that erode social trust.

This brings us to Artificial Intelligence (AI), and the paradoxical role it plays in the landscape of information pollution. On one hand, generative AI technologies—such as large language models, or LLMs—enable the creation of highly convincing disinformation at scale, amplifying the reach and sophistication of false narratives. This capability is compounded by the tracking systems built into the modern web, that enables bad actors to target misinformation at vulnerable groups and “wrap” users into filter bubbles, making the problem significantly worse. Perhaps most problematic, preliminary research indicates that reliance on LLMs might itself weaken the user’s ability to think critically, lowering his or her “immune defenses” against misinformation. On the other hand, AI offers promising solutions to counter these harms. For example, advanced detection systems leveraging natural language processing and machine learning can identify misleading content, support fact-checking, and even generate counter-narratives to debunk false claims. 

Fact checking and debunking, while important, are by no means a silver bullet. The claim that the Earth is flat has been authoritatively debunked, yet it persists. The potential usefulness of AI, though, stands: it is, after all, a general-purpose technology. What else can we do with AI to neutralize disinformation and support information integrity? Which specific technologies carry the most potential? What precautions and guardrails should we build for them to be effective? Several Accelerator Labs have deployed different interventions aiming to use local solutions to tackle the problem of online mis- and disinformation in their specific countries, and some Labs are taking that a step further with information integrity-supporting AI deployments. This is highly practical work: they build prototypes, test them and learn from those tests, aware that failures are as rich in learnings as successes.

In this post, we focus on the learning journeys of the Bangladesh and the Mexico Accelerator Labs, which, though different from one another, are related. In partnership with the Japan Cabinet Office – and inspired by the work of innovative Japanese companies such as Spectee – the two Labs started from the same point of departure: they looked for ways to use AI to supercharge a practice called prebunking. Prebunking is the process of debunking lies, tactics or sources before they strike. Its proponents describe it as a sort of vaccine for the mind, that strengthens people’s ability to think critically and makes them immune to most misinformation or propaganda. Like with vaccines, it is not necessary to immunize everybody: when prebunking reaches a critical threshold of the susceptible population, a phenomenon similar to herd immunity takes place: immune individuals do not repeat or deny false claims outright, preventing their spread. 

Both prototypes have the same ultimate purpose: gain practical knowledge and insights on how to calibrate AI tech, and how to design interfaces and contexts, to support prebunking initiatives. 

Anticipating misinformation in Bangladesh with the help of young people

From this point on, the paths of the two Labs diverge and complement each other. The Bangladesh prototype starts by observing that the country has a large population of young people, digital natives who are very comfortable online. Many of them are already active online and trying to use social media responsibly by verifying the authenticity of claims before sharing them. However, this commendable work does not reach the broader population. 

The UNDP Bangladesh Accelerator Lab conjectures that these young people can take on the role of credible sources by leveraging established trust within their families and communities. In doing so, each of them would function as a prebunking service for their older relatives and enhance information integrity. To test this conjecture, the team is developing an AI-powered pre-bunking application that provides young users with timely, personalized, and credible information. The prototype will leverage agentic AI to predict potential rumors by analyzing historical data and real-time online discourse; profile its young users to tailor the tone and framing of responses; and provide them with inoculation messages to explain to others (especially their families and the older adults within their networks) common techniques of information distortion and why certain content may be misleading. 

A test deployment will allow the Lab to assess whether an anticipatory misinformation shield, mobilizing young people as digital first responders, can strengthen public trust and build a scalable civic model for crisis communication in Bangladesh.

Teaching cognitive resilience and media literacy in Mexico

The prototype in Mexico focuses on the large role that social media plays in shaping people’s perception of security – a high-stakes issue in a region that faces significant threats from criminal violence. In the state of Zacatecas, where the pilot is being tested, social media is the leading source of information about public safety issues. This makes people’s perceptions of security particularly vulnerable to disinformation and propaganda. The authorities have responded by offering facilitated media literacy workshops, delivered through the state’s schools in a classroom setting. While commendable, the initiative struggles to reach the scale and speed required by the populace’s sense of insecurity and the flood of inflammatory social media content. 

The UNDP Mexico Accelerator Lab is designing a version of those workshops that uses an LLM to enhance and conduct the learning experience. Its conjecture is that such an AI-supported workshop can be administered by a regular teacher and still help young participants to develop their skills to recognize and resist security-related misinformation. If that conjecture turns out to be correct, the schools in Zacatecas (and elsewhere) could deploy these workshops at a much higher scale and limited per-pupil costs.

The Lab is designing its AI-supported workshop, with input from Zacatecas public officials and participants to previous iterations of the workshop. Next, it executed A/B tests: one group of students took part in the human facilitator-only version of the workshop, while another engaged in the AI-supported version. By comparing the behavior of the two groups, the Lab aims to compare the two versions’ performance in terms of enhancing the capacity to recognize and resist misinformation.

The results of the prototypes deployed by the Bangladesh and Mexico Accelerator Labs will improve knowledge of how AI-assisted prebunking works in practice. In turn, that knowledge will inform UNDP’s ongoing work in support of information integrity.