In early 2025, the United Nations Development Programme (UNDP) launched the AI Trust & Safety Re-imagination Programme, where we heard from practitioners in 60 countries who are addressing emerging AI-related harms. The programme is designed as a networked ecosystem of Trust and Safety leaders, innovators, researchers, think tanks and civil society organizations in countries, drawing on the lived experiences of stakeholders and highlighting emerging insights to solve for AI-driven harms.

These insights from countries are critical in shaping the new paradigm for Trust and Safety models  – shifting from universal, centralized approaches towards solutions tailored to local contexts. 

Submissions by region  

Needs and opportunities by region 

• Africa: Submissions highlighted a wide spectrum of challenges and critical gaps – from inadequate policy enforcement to the unregulated deployment of AI models. Notably, the marginalization of low-resource languages emerged as a key barrier to advancing AI safety across the continent. There was also a strong emphasis on community-led solutions to address infrastructure gaps, with a focus on cybercrime, misinformation, child safety and youth reskilling.
• Asia-Pacific: Priorities included addressing harms affecting migrant groups, financial fraud and linguistic diversity. Submissions called for a balance between centralized governance and culturally specific adaptations. Common themes were misinformation, gender-based violence, child safety and identity fraud. These reflect growing concerns around AI and digital risks that disproportionately affect vulnerable populations.
• Latin America and the Caribbean: Governance gaps, financial fraud and deepfakes used for disinformation were identified as paramount. Submissions leaned towards combining centralized compliance tools with locally grounded approaches that empower civic leaders and civil society groups.
• Europe and Central Asia: Concerns centred on low-resource languages and algorithmic fairness in diverse contexts. While some favoured EU-inspired centralized frameworks, others stressed the need for localized approach for marginalized groups.
• Arab States: Governance gaps and disinformation are key concerns, and localized approaches are seen as essential.
  North America: Misinformation, data provenance, deepfakes and bias were identified as top issues. 

Top emerging themes

Across regions, submissions reimagine AI Trust and Safety as ‘decentralized with shared foundations,’ aiming to strike a balance between global and local approaches. Shared foundations, such as global datasets and AI Safety infrastructure, led by community-driven implementation strategies were proposed to counter the pitfalls of one-size-fits-all models while avoiding the fragmentation of purely local solutions.

The Global South led in the number of submissions, pointing to early warning signs of emerging AI harms and providing evidence of their widespread, cross-border amplification of AI harms. Common themes include misinformation, deepfakes, bias, fraud, and child safety. These challenges are particularly significant in low-resource contexts, where limited digital literacy, weak regulations, and cultural or linguistic diversity amplify risks because existing solutions are scaled and optimized primarily for presumed mono-cultural and mono-lingual populations.

• Disinformation / deepfakes: Solutions such as AI detection tools (e.g., WITNESS's TRIED benchmark, Karna’s audio deepfake detection) or monitoring platforms (e.g., Phoenix in Kenya).
• Bias and discrimination / gender-based violence: Frameworks for culturally adaptive AI or chatbots for migrant workers (e.g., PoBot by Migrasia in Hong Kong).
• Financial fraud / identity theft / authentication and cybersecurity: Tools such as fraud detection apps (e.g., Silverguard in Brazil) or personas for inclusive fintech (e.g., Assessed Intelligence).
• Child Safety/Exploitation: Reskilling programmes (e.g., Trust & Safety Africa Academy in Ghana).
• Safety impact of low-resource languages: Datasets and benchmarks addressing low- resource language inaccuracies that lead to misinformation. (e.g. Distant Voices work by Ushahidi, Trustweave, Duco and Tattle).