How an AI research tool helped challenge support for police violence in Brazil

Art by Marcas, used with permission.

This article, written by Daniel Stanley, Marianna Araujo and Diogo Santos, is published here through an agreement between Instituto Fogo Cruzado and Future Narratives Lab with Global Voices. For more information, see the report “Challenging narratives on police violence in favelas.”

In 2025, as Instituto Fogo Cruzado, Brazil’s leading civil society monitor of armed violence, and UK-based research organisation Future Narratives Lab were finalising a study on messaging around police violence in favelas, we were shaken by a grim turn of events. In a single favela raid in Rio de Janeiro on October 28, 122 people were killed by Brazilian police, the deadliest policing operation in the country’s history. Polls conducted in the days after the raid showed continued broad public support for such action, despite longstanding evidence that the raids have no impact on reducing crime rates.

In 2023, Brazil recorded 45,747 intentional violent deaths — among the highest rates in the world. Like many democracies, the country has long grappled with the issue of large-scale public support for state violence, which has contributed to a broader societal sense of fear, fatalism, and desensitisation. Changing this requires understanding what works to challenge certain misconceptions. The broad support for the October 28 raids, despite the casualties, highlighted the critical need for grassroots organisations to discover messaging that resonates with the public, especially in a polarised political environment like Brazil.

Understanding what drives support for violence

Civil society organisations working on contentious issues face a problem: many audiences are supportive of policies and actions that run counter to their own interests, yet are difficult and expensive to reach for research, especially in politically uncertain or highly polarised contexts. Such organisations, particularly in the Global South, often lack the funding and resources to carry out the types of in-depth research that Global North-based civil society organisations use. Focus groups and qualitative interviews are the standard route to understanding why people hold the views they do, but they are slow and costly, often putting them out of reach for the smaller organisations doing the most urgent work.

What if an artificial intelligence (AI) tool could lower these barriers and allow small, grassroots organisations to understand how their target audiences think? The large-scale adoption of AI tools for everyday use has created greater opportunities to learn research skills and process information. But, as with any technological innovation, there are ethical considerations around the use of such technology, including the implications of using tools created by Big Tech corporations and the environmental consequences of data centres.

Clearly, it is not a technology that can be used lightly. As we developed the tool, we had many discussions about how to use AI in a way that was consistent with our overall beliefs and aims. Ultimately, we decided to maintain an approach we saw as capable of contributing to advancing a human rights movement, while taking specific steps to protect user privacy, prioritise open-source software-based solutions, and plan future actions to address further concerns.

Using AI to uncover underlying beliefs

Before this project, Future Narratives Lab had developed an AI Interviewer tool specifically for civil society organisations looking to understand their audiences. After a successful trial of the tool in a project testing messages on the criminal justice system in the UK, Instituto Fogo Cruzado approached Future Narratives Lab for support using it in a new project on state-sponsored violence in Brazil, funded by the International Resource for Impact and Storytelling (IRIS), as part of their “New Tech, New Rules” programme.

The main use of the tool in this project was for primary research, laying the foundations for the development of new message and counter-narrative approaches. Twenty-eight in-depth qualitative interviews were conducted via the chatbot-based AI interviewer, with each participant able to engage in the conversation using their own devices via a supplied URL and participant code.

Participants were selected and recruited based on a demographic profile identified in prior desk research as being more likely to support an increase in police violence as a means to reduce crime: those who identified as Christian (Evangelical or Roman Catholic), had not completed higher education, and had a household income of less than 500 BRL (approximately USD 98).

The AI interviewer carried out a customised, structured but adaptive conversation with each participant for up to 30 minutes, asking follow-up questions and probing for inconsistencies and reasoning. Each interview was based on an overall brief setting out what the research aimed to achieve and what happened, concurrently, but unique to each participant. To give a sense of how a typical interview went:

It consistently started with the question: “What comes to mind when you hear ‘violence’ in Brazil?” Participants responded with a range of associations and, sometimes, specific experiences.
The interviewer would then ask responsive questions to clarify or request further explanation as needed. Other subjects were then explored, such as who the main victims of such violence are and who holds responsibility for reducing it.
One key point was the presentation of official data showing a decrease in homicides, to which participants were asked to react. Many expressed surprise or disbelief; these responses were then probed for further explanation.
Questions then explored different, narrower aspects of the topic, including opinions on police violence specifically, the root causes of crime, and recommended actions to address it.

The process of message development that followed uncovered a range of new insights and potential opportunities for persuasion.

What the AI Interviews allow

The ability to carry out multiple interviews, concurrently and remotely, enables organisations to compress what would normally take weeks of fieldwork into days at a fraction of the cost. For under-resourced civil society organisations across the world, this tool creates a significant shift: it makes in-depth audience research about what would move their audiences viable.

In this case, Fogo Cruzado used the tool to investigate what drives underlying beliefs and opinions about police violence. The AI interviews, combined with their years of experience and expertise, suggested three potential areas where messaging could sway reluctant audiences:

police unpreparedness undermines the effectiveness of raids;
raids leading to innocent killings via stray bullets;
the failure of raids to address root causes.

This foundational research directly led to the next step in the study: surveying over 500 Brazilians from a “hard to persuade” demographic in an RCT-style test, to understand which of these messaging options might contribute to a decrease in support for police raids.

What we found

The message that performed best was not centred on victims, empathy, or structural injustice — the framing narrative that advocates in the sector often reach for first. It was a pragmatic critique of police competence: “The police aren’t sufficiently trained or prepared, so carrying out raids in crowded favelas only leads to more shootouts and more innocent victims.” This framing emerged directly from the AI-conducted interviews, where participants repeatedly described police as underprepared rather than malicious.

Equally striking, Evangelical participants, who in Brazilian political discourse are often assumed to be aligned with conservatism, especially in a context where churches frequently collaborate with security forces, were less supportive of police raids than other groups. Their open-text responses revealed moral frameworks rooted in redemption, structural critique, and a distrust of elites, which sits in tension with the punitive politics often attributed to Evangelicals.

The big picture

Favela residents are routinely misrepresented in both domestic and international political and media narratives. Our study raises their voices in ways that aren’t often done. It makes us rethink which audiences we consider “persuadable.”

The affordable, fast, and transferable AI methodology was a key factor in how we were able to make this happen, and our goal is for other grassroots organisations to do the same. To start, Fogo Cruzado has already shared information on this approach with 20 Brazilian organisations through a “communication camp,” ahead of a crucial general election in October 2026, in which public security will be a central issue.

As the impact of AI on society becomes an ever more heated topic of discussion, there are increased expectations and responsibilities to ensure that its use takes into account the many ethical considerations at play.

Between our organisations, we feel in a strong position to do this — Future Narratives Lab has published about the negative impacts of mainstream AI use, and Fogo Cruzado has a close connection with grassroots communities to ensure their views are taken into account.

Steps have already been taken to move the AI interviewer away from more extractive “frontier” models, like Claude or ChatGPT, toward models that respect privacy and data protection and uphold higher overall standards. In a time when we’re seeing populism rise globally, the possibility of changing minds seems far-fetched, but with technology allowing us to meet audiences where they are, we get closer to building collective power.