After Geneva, AI Governance Must Confront the Trust Deficit

Last week, Geneva hosted the first session of the UN Global Dialogue on AI Governance, the most recent multilateral effort to coordinate international responses to the risks and opportunities of artificial intelligence. Opening the meeting, Secretary-General António Guterres warned:

The computing power, the data and the talent behind the most advanced systems are concentrated in a handful of companies, and in a handful of countries. Most nations — including many developing countries — have had no say in decisions that will shape their futures. When power imbalances are hard-wired into technology, inequality becomes part of the code.

The Dialogue, alongside the release of a preliminary report from the Independent International Scientific Panel on AI, signals that global governance is advancing from ethical principles toward questions of implementation and institutional capacity.

But a significant challenge remains: the Dialogue reaches organizations with the capacity to participate in multilateral processes. It does not naturally reach quilombola communities in Rio de Janeiro who associate AI with surveillance risk, nor low-income families in Albuquerque who learn technology through community libraries. This is not a design flaw unique to the Dialogue; it is a structural trust deficit. Without trust, participation becomes decoration.

Two contexts, one structural problem

Between 2024 and 2025, Public Knowledge, UnidosUS, and the National Digital Inclusion Alliance conducted focus groups in Denver, Atlanta, New Mexico, and Appalachia. The resulting report, “The Blueprint for Equitable Digital Participation,” documents something that rarely appears in AI debates: what it actually means to live in digital exclusion. Despite the barriers described, participants did not reject technology — they sought digital learning through libraries and trusted local networks, not through government portals or technology companies. Communities had already built the social infrastructure; what was missing was institutional recognition of it.

In Brazil, the Territórios Digitais research — conducted by GERATE Lab with Instituto Peregum and UNEafro Brasil in Quilombo Santa Rita do Bracuí and the Multiethnic Indigenous Village Filhos da Terra — reached a remarkably similar conclusion: trust is produced by territorial ecosystems of legitimacy — leaders, educators, health workers, and, in particular, women community leaders. In both cases, the lesson is the same: digital policies fail when they treat communities as end users and ignore the institutional ecosystems that already organize trust locally.

Recent initiatives reinforce this shift. The newly created Participatory AI Research Network, inspired by the work of researchers such as J. Nathan Matias and Megan Price, argues that communities, journalists, and researchers should collaborate to produce reliable evidence about AI's impacts. Even so, Latin American territorial methodologies remain largely absent from these emerging international conversations — highlighting the need to diversify not only participation, but the very empirical foundations of AI governance.

The community navigator and sociotechnical mediation

Both contexts point to a practical answer that AI governance has only begun to recognize: the role of trusted intermediaries embedded within communities. The “Blueprint for Equitable Digital Participation” describes digital navigators — trained community members who help neighbors access connectivity — as one of the most effective interventions for digital equity. In Denver, the housing authority conducted outreach through already-trusted spaces; in Atlanta, without that mediation, the same programs remained invisible. The difference was not the quality of the program — it was the presence, or absence, of trusted mediation.

Territórios Digitais names this layer sociotechnical mediation: the relationship between communities and technological systems is never direct or neutral, it is mediated by power, language, and a history of trust. Within quilombola and Indigenous territories, leaders, elders, and educators already mediate relationships between different knowledge systems. When this mediation operates without recognition or sustained investment, communities bear the weight of maintaining trust while external institutions capture the benefits.

This does not mean that every territorial leadership is automatically representative or immune to capture — communities also face internal disputes over legitimacy, and recognizing trust infrastructures does not replace accountability mechanisms within the territories themselves. The argument here is not that community mediation is infallible, but that it already exists, is functional, and is systematically ignored by governance designs that treat communities merely as recipients of policy, not as its co-authors.

Matias and Price argue that involving people with lived experience improves the scientific evaluation of AI itself — not only its democratic legitimacy. This perspective now extends beyond evaluation: the creation of the Participatory AI Research Network reflects the recognition that trustworthy AI depends on community-centered approaches to evidence production. Territorial mediation extends this argument further: communities are not simply participants in AI evaluation — they are the institutional infrastructures that make community-centered AI implementation possible.

The governance problem no one has named

The caution shown by the communities studied is not technological rejection — it is a rational response to a history of externally introduced technologies that served institutional interests before community needs. The research documented strong interest in using AI to strengthen local capacities and preserve cultural knowledge. What is being questioned are the conditions under which the technology is introduced, governed, and validated.

The experience of Cacique Alex, a participant in the Territórios Digitais research from the Filhos da Terra Village, illustrates this distinction: faced with a bill that threatened his community's rights, he used generative AI to understand the legal language — but validated the information with a trusted territorial partner before acting. The trust was not in the AI, but in the human who validated its result.

The risks identified by communities go beyond surveillance. The Territórios Digitais research documented how disinformation reshapes trust relationships and interferes with collective decision-making. The case of Dona Marilda, a leader of Quilombo Santa Rita do Bracuí, illustrates the concrete cost of this. After the Public Prosecutor's Office suspended a construction project for environmental violations, a false narrative attributed the decision to the quilombo. Repeated across messaging apps, the lie generated real threats and forced the family to temporarily leave their home under state protection. "We can no longer walk alone, not even at night," she reported. "We lost that freedom within our own territory." The disinformation was never publicly corrected, according to Marilda.

This episode reveals a dimension absent from the AI governance debate: who bears the costs when information systems fail. For marginalized communities, this means loss of political legitimacy, restricted movement, and erosion of community leadership — not just abstract disinformation.

Secretary-General Guterres raised a related concern in Geneva, warning that a machine-generated lie can now persuade as effectively as the truth. As he put it: "A society that cannot agree on what is real cannot defend itself."

The gap the Scientific Panel itself acknowledges

The Preliminary Report of the Independent International Scientific Panel on AI confirms, in its own words, why this agenda matters. The report argues that "the artificial intelligence divide is not just about access, but about capacity to influence artificial intelligence development." The evidence base itself reflects this concentration:

The evidence on the impacts of AI is concentrated in high-income, English-language contexts. Economic studies are biased towards advanced economies, large firms and formal work. The AI evaluation infrastructure remains linguistically and geographically concentrated.

This is not incidental: in 2025, 91% of notable AI models originated from the private sector, and institutions based in the United States produced 59 notable AI models, compared with 35 in China and just 13 in the rest of the world, according to research cited in the Preliminary Report.

AI governance capacity and public-sector capability compound this concentration. The report defines this dimension as the ability to understand, guide, regulate, and support AI development, and finds it strikingly absent where it is most needed. According to figures from the United Nations Conference on Trade and Development cited in the report, 118 countries, predominantly in the Global South, are not engaged in major AI governance discussions, and less than a third of developing countries have developed national AI strategies.

Notably, the Panel itself acknowledges the limits of its own evidence base on a point closely related to this text's central argument: how AI interactions translate into social erosion over time. The report admits that "the pathway from individual-level AI interactions to societal-level outcomes such as epistemic erosion, civic participation and social cohesion remains poorly understood," and that "existing evidence captures snapshots — engagement metrics, documented harms, case studies — but not the cumulative trajectory." Long-term territorial research, such as that conducted by Territórios Digitais, does not close this scientific gap on its own — but it is needed to broaden the lens and offer a concrete starting point for the kind of longitudinal evidence that is missing: the case of Dona Marilda, for instance, documents in detail exactly this progression — from an isolated false narrative to the rupture of a community leader's freedom of movement.

Three principles for implementation

The gap the Panel itself acknowledges — between individual-level AI interactions and their cumulative effects on social cohesion — points to three implementation principles, which respond directly to the challenges the report identifies.

First, recognize Territorial Trust Infrastructures as implementation infrastructure. In Denver, families learned about programs from staff who personally visited their homes; in quilombola and Indigenous territories, participants turned to leaders, not digital platforms. Trust is already institutionalized in these actors — they are the ones who make information credible and translate technologies into local realities. Governance frameworks that fail to map these networks before designing participatory processes systematically exclude those most affected by AI.

Second, invest permanently in sociotechnical mediation as a public good. The US Digital Equity Act funded digital navigators as public infrastructure, until its arguably unlawful cancellation by the Trump administration in May 2025. AI governance should fund territorial articulators, community navigators, and Indigenous knowledge keepers in an equivalent and permanent way — not as temporary intervention, but as a structural part of governance itself.

Third, treat long-term territorial research as a primary source of evidence, not as a supplementary case study. This is precisely the kind of research that can fill the gap the Panel identifies regarding cumulative trajectories of social erosion. Initiatives such as the Participatory AI Research Network, the Latin American AI Diaspora Network (Fundar and nadIA), and territorial research labs such as GERATE Lab already point in this direction, connecting Latin American AI professionals to territorial knowledge and local development agendas — not as symbolic participation, but as part of AI governance's own scientific foundation.

Beyond participation

The question is no longer whether communities should be included in AI governance, but whether that governance is prepared to recognize community-centered institutions as legitimate actors in its implementation — not merely as beneficiaries, nor simply as sources of data to be extracted. Territórios Digitais contributes to this emerging agenda by demonstrating that quilombola and Indigenous territories also produce knowledge about governance, including regarding gaps that the UN Scientific Panel itself acknowledges it cannot yet explain on its own. This knowledge can circulate on terms that preserve authorship and territorial autonomy, as evidenced by the gradual publication of the research in the Mozilla Data Collective since the fieldwork concluded. The Collective offers a path away from data extractivism, favoring community autonomy instead.

As Secretary-General Guterres stated in Geneva: "If AI is to be trusted, those who build it must be accountable." This accountability also hinges on recognizing that community-centered institutions of trust are not confined to any single region. They include territorial leaders, quilombola associations, and local communicators in the Global South who have long governed technology through their own systems of trust — and they include the housing authority staff in Denver and the digital navigators once funded by the US Digital Equity Act, before its cancellation. Both are part of the same missing infrastructure that AI governance has yet to recognize. As he put it: "Let this meeting be remembered as the moment governance began to catch up with the technology." The next phase must ensure that governance keeps pace with all the communities already doing this work — and does so on their terms.