<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Tianrong Ding Terisa on Medium]]></title>
        <description><![CDATA[Stories by Tianrong Ding Terisa on Medium]]></description>
        <link>https://medium.com/@dtr0105?source=rss-13709fdd2256------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*CARVunphujYi694GfYU3kQ.jpeg</url>
            <title>Stories by Tianrong Ding Terisa on Medium</title>
            <link>https://medium.com/@dtr0105?source=rss-13709fdd2256------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 21:45:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dtr0105/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Toward Development: Data for AI as Digital Public Infrastructure]]></title>
            <link>https://medium.com/@dtr0105/toward-development-data-for-ai-as-digital-public-infrastructure-27ba9380bc17?source=rss-13709fdd2256------2</link>
            <guid isPermaLink="false">https://medium.com/p/27ba9380bc17</guid>
            <category><![CDATA[digital-infrastructure]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[development]]></category>
            <category><![CDATA[ai-data-collection]]></category>
            <dc:creator><![CDATA[Tianrong Ding Terisa]]></dc:creator>
            <pubDate>Thu, 26 Feb 2026 10:06:44 GMT</pubDate>
            <atom:updated>2026-02-26T10:06:44.773Z</atom:updated>
            <content:encoded><![CDATA[<h3>Toward Development: Data for Digital Public Infrastructure</h3><p>Artificial intelligence is poised to add $15.7 trillion to the global economy by 2030. Yet here’s the catch: only 3% of those economic benefits are expected to reach Latin America, 6% for developed Asia (excluding China), and a mere 8% for the combined populations of Africa, Oceania, and other Asian markets. The rest will flow to the handful of countries that have built the data infrastructure and governance frameworks necessary to harness AI’s potential.</p><p>This isn’t just a technology gap. It’s a data governance gap, an infrastructure gap, and ultimately, a development gap. For international development agencies, the question isn’t whether to engage with AI, but how to ensure the countries they support aren’t left behind in this transformation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jH0hwY9d-kcRXZqJ" /></figure><p>The answer lies at the intersection of two critical domains: data governance and <a href="https://www.undp.org/digital/digital-public-infrastructure">Digital Public Infrastructure</a> (DPI). Together, they form the foundation upon which AI-ready nations are built. Let’s break down what this means for development practitioners and the frameworks that can guide effective intervention.</p><h3>Understanding the foundations: Data as the cognitive layer of AI</h3><p>AI systems learn from data the way humans learn from experience. Large language models train on vast collections of text, computer vision systems analyze millions of images, and predictive models identify patterns in historical datasets. Without quality data, even the most sophisticated algorithms are essentially blind.</p><p>Governments hold unique advantages in the data economy. They maintain birth and death registries, land records, court rulings, legislation, tax records, and service delivery data. This public data represents a treasure trove for AI development, particularly for models trained to serve public functions.</p><p>But there’s a problem. Much of this data remains fragmented across agencies, locked in legacy systems, or exists only in unstructured formats that machines cannot easily process. A social protection agency might hold beneficiary records in one database, while health records sit in another, with no easy way to link them for comprehensive service delivery.</p><p>The breakthroughs that powered today’s AI advances came from open data initiatives. <a href="https://www.image-net.org/">ImageNet</a>, a human-annotated dataset of 3.2 million images created for scientific purposes, made the deep learning revolution possible. The <a href="https://www.rcsb.org/">Protein Data Bank</a>, established nearly half a century ago, enabled AlphaFold’s breakthrough in protein folding. These examples illustrate a fundamental principle: open, accessible data drives innovation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iqOp23Uy8VBZo3vS" /></figure><p>For development agencies, this means data governance isn’t just a compliance issue. It’s policy design for the AI era. Countries that get this right will have the raw material to train AI systems that serve their specific needs. Countries that don’t will find themselves dependent on models trained elsewhere, on data that doesn’t reflect their languages, cultures, or contexts.</p><h3>Dagital Public Infra: The Enabling Architecture</h3><p><a href="https://www.undp.org/digital/digital-public-infrastructure">DPI</a> is a set of foundational digital building blocks designed for public benefit. Think of it as the digital equivalent of roads and railways: the underlying infrastructure that enables everything else to function. Unlike physical infrastructure, though, DPI is built on open technology standards and governed for the public good.</p><p>The core components of DPI include:</p><ul><li><strong>Identity systems</strong> that let people prove who they are across services</li><li><strong>Data exchange mechanisms</strong> that enable secure information sharing between agencies</li><li><strong>Digital payments infrastructure</strong> for financial transactions</li><li><strong>Registries and credentials</strong> for everything from land titles to professional certifications</li></ul><p>What makes DPI different from traditional e-government is its architecture. The <a href="https://cdpi.dev/">Centre for Digital Public Infrastructure</a> identifies five key principles: interoperability (systems that work together), minimalism (reusable building blocks), ecosystem innovation (enabling diverse actors to build on top), federation (distributed rather than centralized control), and security by design.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tIYpq-P3UXHYDNtk" /></figure><p>Countries like <a href="https://www.indiastack.org/">India</a> (with its India Stack), <a href="https://e-estonia.com/">Estonia</a> (e-Residency and X-Road), <a href="https://www.singpass.gov.sg/">Singapore</a> (Singpass), and <a href="https://www.gov.br/governodigital/pt-br">Brazil</a> (Pix payments) demonstrate what’s possible when these principles are applied at scale. Citizens can open bank accounts, access social benefits, and prove their identity through interconnected systems that prioritize both efficiency and rights protection.</p><p>For AI development, DPI serves as the “rails” on which AI services can run. A well-designed digital identity system can authenticate users for AI-powered health consultations. A data exchange layer can enable AI systems to access the information they need while maintaining privacy safeguards. Payment infrastructure can support AI-enabled financial services.</p><h3>The DPI-AI Framework</h3><p>The <a href="https://digitalpublicinfrastructure.ai/wp-content/uploads/2026/01/Vision_Paper_-DPI-AI-Framework_2026_CDPI.pdf">CDPI DPI-AI Framework</a> provides a practical model for integrating AI with public infrastructure. It consists of three elements working together:</p><p><strong>AI Blocks</strong> are modular units of AI capability that can be invoked as needed. Foundational blocks handle universal functions like translation, speech-to-text, optical character recognition, and summarization. Sector-specific blocks address particular domains: eligibility verification for social protection, clinical decision support for healthcare, personalized tutoring for education. These blocks are designed to be callable functions, not monolithic systems, allowing them to be combined and recombined for different purposes.</p><p><strong>DPI Workflows</strong> form the orchestration layer that coordinates AI Blocks with existing systems. They define sequences of steps, ensure policy compliance, and enable auditability. When a citizen applies for a benefit, a workflow might verify their identity through the digital ID system, check eligibility against registry data using an AI Block, process payment through the payments infrastructure, and log everything for accountability.</p><p><strong>Public Agents</strong> are the interfaces through which AI capabilities reach end users. These can be human agents (benefits officers using AI tools), fully automated systems (chatbots handling routine queries), or hybrid approaches (AI-supported human decision-makers). The key is designing these agents to be accessible, accountable, and aligned with public values.</p><h3>Policy and Governance Model</h3><p>How countries govern AI development varies significantly, with implications for how they approach data and infrastructure. The <a href="https://documents1.worldbank.org/curated/en/099120224205026271/pdf/P1786161ad76ca0ae1ba3b1558ca4ff88ba.pdf">World Bank identifies 4 regulatory approaches</a>: industry self-governance, soft law (principles and guidelines), hard law (binding regulations), and regulatory sandboxes (controlled experimentation environments).</p><p><a href="https://www.sciencedirect.com/science/article/pii/S0963868724000672">Academic research</a> synthesizes seven principles for responsible AI governance: accountability, diversity and fairness, human agency and oversight, privacy and data governance, technical robustness and safety, transparency, and social and environmental well-being. These principles translate into three types of governance practices: structural (how organizations are set up), procedural (the processes followed), and relational (how stakeholders interact).</p><p>But principles alone don’t capture the full picture. The geopolitical competition between the United States and China is shaping how AI governance models spread globally, with developing countries caught in the middle.</p><h4>The US Approach</h4><p>The US model is characterized by private sector-led innovation backed by massive investment. In 2024, US private sector AI investment reached <a href="https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf">$109.1 billion</a>, far outpacing other nations. This investment concentrates power in a handful of tech companies, with <a href="https://unctad.org/press-material/ais-48-trillion-future-un-trade-and-development-alerts-divides-urges-action">100 firms accounting for roughly 40%</a> of corporate R&amp;D spending on AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9EdQbDHzp7C9psWB" /></figure><p>The policy approach, articulated in the <a href="https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf">America’s AI Action Plan</a> released in July 2025, rests on three pillars: accelerating domestic innovation, building infrastructure, and international diplomacy. The plan recognizes the geostrategic importance of maintaining US AI dominance and promotes “full-stack AI export packages” for allies and partners.</p><p>A key element of US strategy is the promotion of open AI models as the global standard. Models like <a href="https://www.llama.com/">Meta’s Llama</a> and platforms like <a href="https://huggingface.co/">Hugging Face</a> allow other countries to access powerful AI capabilities without building from scratch. However, the US also wields export controls to limit advanced semiconductor access to adversaries, creating tension for developing countries seeking to build their own capabilities.</p><h4>The China Approach</h4><p>China’s approach is state-directed, with substantial domestic investment in AI development. The country has demonstrated its competitive capability through models like <a href="https://www.deepseek.com/">DeepSeek R1</a>, which achieved performance on par with leading US systems. This success story has become a powerful tool in China’s diplomatic efforts to position itself as an alternative AI partner for the Global South.</p><p>The <a href="https://www.csis.org/analysis/open-door-ai-innovation-global-south-amid-geostrategic-competition">Digital Silk Road initiative</a> provides a comprehensive framework for exporting Chinese digital technologies, creating technological dependencies worldwide. China has also positioned itself as a champion of AI cooperation through initiatives like the <a href="https://www.mfa.gov.cn/eng/wjbzhd/202409/t20240927_11498465.html">AI Capacity-Building Action Plan for Good and for All</a>, co-sponsored at the UN.</p><p>However, China’s AI development faces significant constraints. Despite efforts to build domestic chip manufacturing, <a href="https://www.cnas.org/publications/reports/countering-ai-chip-smuggling-has-become-a-national-security-priority">only 2 of 22 models</a> developed exclusively by Chinese AI labs in 2025 were trained with Chinese chips. Hundreds of thousands of chips are smuggled into China annually, underscoring continued dependence on Western technology.</p><h4>Implications for Developing Countries</h4><p>For developing countries, the US-China competition creates both opportunities and risks. On one hand, competition drives investment and technology transfer. On the other, it forces difficult strategic choices that can have long-lasting geopolitical implications.</p><p>The risk of digital colonialism is real. Without the ability to adapt AI models to local contexts or build domestic capabilities, countries risk becoming mere consumers of technologies designed elsewhere. This is driving growing interest in “sovereign AI ecosystems”: efforts to develop, control, and govern national AI capabilities without complete dependence on foreign support.</p><h3>Data for DPI: Case Studies</h3><p>Africa illustrates both the challenges and opportunities of building AI-ready infrastructure. The continent has the potential to contribute <a href="https://acetforafrica.org/research-and-analysis/insights-ideas/digital-public-infrastructure-dpi-will-drive-ai-for-africas-economic-transformation/">$1.2 trillion to the AI economy by 2030</a>, yet currently holds only <a href="https://www.bloomberg.com/news/articles/2024-02-23/us-data-center-operator-equinix-plans-390-million-africa-build">1% of the world’s computing capacity</a>. <a href="https://dataprotection.africa/which-african-countries-have-a-data-protection-law/">40 African countries</a> have adopted data protection policies, and the <a href="https://au.int/sites/default/files/documents/44004-doc-EN-_Continental_AI_Strategy_July_2024.pdf">AU Continental AI Strategy</a> provides a unified regional framework.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*clM3krCyTNrlc6Lv" /></figure><h4>Tunisia</h4><p>Tunisia has emphasized financial inclusion through unified mobile platforms and digital government reforms. The government’s approach demonstrates how DPI can serve multiple objectives simultaneously: expanding access to financial services while building the data infrastructure needed for future AI applications.</p><h4>Morocco</h4><p>Morocco’s <a href="https://dig.watch/resource/morocco-digital-strategy-2030">Digital Strategy 2030</a> prioritizes robust infrastructure, digital literacy, and public-private partnerships. By harmonizing digital strategies across sectors, the country has established a foundation for DPI utilization nationally.</p><h4>Ghana</h4><p>Ghana’s initiatives include national ID systems, payment system integration, and mobile money interoperability. A particularly innovative element is the integration of reverse transition planning into DPI contracts, ensuring systems remain adaptable as technology evolves.</p><h4>Kenya and Nigeria</h4><p>Both countries have launched AI knowledge centers to build digital skills and foster innovation. The <a href="https://www.undp.org/kenya/blog/africa-center-competence-digital-and-artificial-intelligence-skilling">Nairobi-based Centre of Competence for Digital and Artificial Intelligence Skilling</a>, established with UNDP and Microsoft partnership, drives digital inclusion through public-private collaboration. Nigeria’s <a href="https://fmcide.gov.ng/nigeria-launches-ai-scaling-hub-in-collaboration-with-gates-foundation-to-accelerate-impact-in-health-education-and-agriculture/">AI Scaling Hub</a>, launched in June 2025 with Gates Foundation support, focuses on agriculture, healthcare, and education applications.</p><p>Major tech investments are following these initiatives. Microsoft and G42 are investing <a href="https://interface.media/blog/2024/05/24/microsoft-pivots-in-east-africa-with-1-billion-kenyan-data-centre-run-on-geothermal-power/">$1 billion in a geothermal-powered data center campus in Kenya</a>. <a href="https://aws.amazon.com/blogs/publicsector/aws-investment-south-africa-results-economic-ripple-effect/">AWS Africa’s Cape Town cloud region</a> has added approximately $673 million to South Africa’s GDP.</p><h3>Financing and international cooperation mechanisms</h3><p>Building AI-ready infrastructure requires significant investment. Frontier AI model costs have grown at <a href="https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models#updates">2–3 times per year</a> for the past nine years. For developing countries facing competing priorities like debt service and immediate social needs, this presents a genuine dilemma.</p><p>Several financing models are emerging:</p><p><strong>Public-private partnerships (PPPs)</strong> combine government oversight with private sector capital and expertise. These arrangements can reduce upfront costs while ensuring local regulatory control.</p><p><strong>Regional public cloud infrastructure pools</strong> allow countries to share computing resources, reducing individual investment requirements while maintaining data sovereignty within the region.</p><p><strong>The Africa AI Fund</strong> proposes <a href="https://allafrica.com/stories/202504140440.html">$60 billion in investment</a> to create 500,000 jobs annually and lift 11 million people out of poverty by 2030.</p><p><strong>The Global Fund for AI</strong>, recommended by the <a href="https://www.un.org/sites/un2.un.org/files/governing_ai_for_humanity_final_report_en.pdf">UN High-Level Advisory Body</a>, would provide dedicated financing for AI capacity building in developing countries.</p><p>Multi-stakeholder initiatives are also playing a role. The <a href="https://www.undp.org/digital/news/global-dpi-summit-outcome-statement">50-in-5 Campaign</a> aims to design, launch, and scale DPI in 50 countries within five years. The <a href="https://www.digitalpublicgoods.net/">Digital Public Goods Alliance</a> helps discover and deploy open source technology globally. <a href="https://www.govstack.global/">GovStack</a> works to ensure everyone can access government services using trusted digital technology.</p><p>Development agencies have a critical role as catalysts. They can help countries navigate technology choices, broker partnerships, provide technical assistance, and ensure that investments align with sustainable development goals.</p><h3>Building AI-Ready Nations through Data Infrastructure</h3><p>The transformation driven by artificial intelligence is just beginning. For international development agencies, the imperative is clear: help countries build the data governance frameworks and digital infrastructure needed to participate in this transformation on their own terms.</p><p>The framework presented here, integrating data policy, DPI architecture, and AI governance, provides a pathway forward. It recognizes that AI readiness is not just about having the latest algorithms or the most powerful computers. It’s about having the data, infrastructure, and governance in place to deploy AI in ways that are inclusive, sovereign, and beneficial to all.</p><p>The $15.7 trillion question is not whether AI will transform development. It’s whether the benefits of that transformation will be shared equitably, or concentrated in the hands of those who already have the data and infrastructure to capture them. The answer depends on the choices we make today.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=27ba9380bc17" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cryptos as Digital Public Infrastructure: Understanding the New Plumbing of Global Finance]]></title>
            <link>https://medium.com/@dtr0105/cryptos-as-digital-public-infrastructure-understanding-the-new-plumbing-of-global-finance-b3f6f449050c?source=rss-13709fdd2256------2</link>
            <guid isPermaLink="false">https://medium.com/p/b3f6f449050c</guid>
            <category><![CDATA[blockchain-technology]]></category>
            <category><![CDATA[digital-infrastructure]]></category>
            <category><![CDATA[governance-and-tech]]></category>
            <category><![CDATA[crypto]]></category>
            <category><![CDATA[digital-public-goods]]></category>
            <dc:creator><![CDATA[Tianrong Ding Terisa]]></dc:creator>
            <pubDate>Wed, 25 Feb 2026 03:18:23 GMT</pubDate>
            <atom:updated>2026-02-25T03:19:10.892Z</atom:updated>
            <content:encoded><![CDATA[<p>While headlines focus on cryptocurrency volatility and speculative trading, a more consequential shift is happening in the infrastructure layer of money itself. <a href="https://www.circle.com/usdc">Stablecoins</a>, once dismissed as merely a tool for crypto traders, have evolved into a $230 billion-plus financial infrastructure processing over $27 trillion in annual transaction volume. That is more than double PayPal’s annual volume and rapidly approaching a significant share of global payment flows.</p><p>The year 2025 marked a regulatory inflection point. The United States passed the GENIUS Act, the European Union implemented MiCA, and jurisdictions from Singapore to Hong Kong established comprehensive licensing frameworks. For the first time, major economies have clear rules governing how stablecoins can operate, who can issue them, and what reserves must back them. This regulatory clarity is accelerating adoption by traditional financial institutions, including <a href="https://www.jpmorgan.com/payments/newsroom/jpm-coin-usd-deposit-token-institutional-clients">JPMorgan</a>, <a href="https://www.citigroup.com/global/news/press-release/2025/citi-integrates-citi-token-services-with-24-7-usd-clearing-real-time-cross-border-payments-liquidity-management">Citi</a>, and <a href="https://corporate.visa.com/en/solutions/crypto/stablecoins.html">Visa</a>.</p><p>What makes this moment significant is not just the scale of adoption, but the nature of the infrastructure being built. Stablecoins represent something new: programmable, always-on, globally accessible digital money that operates on open protocols. They sit at the intersection of public and private, state and market, centralized and decentralized. Understanding them requires a framework that goes beyond traditional financial analysis to examine their technical architecture, governance models, and implications for <a href="https://www.undp.org/digital/digital-public-infrastructure">Digital Public Infrastructure</a>.</p><p>This article provides that framework. We will examine what makes stablecoins technically distinct, how different governance models are emerging, what the new regulatory landscape looks like, and how development practitioners might think about their potential and risks.</p><h3>The technical architecture: open protocols, standards, and composability</h3><p>To understand why stablecoins matter, it helps to contrast them with how digital money has traditionally worked. When you send a bank transfer, you are not actually moving money in real time. You are sending a message through a network (like SWIFT) requesting that another bank update their ledger. The actual settlement, the moment money truly changes hands, happens later, often through central bank systems that only operate during business hours. While stablecoins work differently. They are tokenized cash issued on public blockchains. When you send a stablecoin, the transaction settles in minutes or seconds, 24 hours a day, anywhere in the world. The blockchain serves as a shared ledger that all participants can verify, eliminating the need for correspondent banks to maintain nostro accounts and reconcile balances.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/904/1*d4xp8AunlT_GpoCT5XZzqw.png" /></figure><h4>The stablecoin stack</h4><p>This stack creates both opportunities and challenges. The open nature means anyone can build on top of it without permission. But it also means fragmentation. There are dozens of blockchains, each with different trade-offs in speed, cost, and security. Moving value between them requires bridges or cross-chain protocols, adding complexity and risk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OjpfHUSnx8Xn1hDFahSX0A.png" /></figure><h3>The governance spectrum: state, market, and hybrid models</h3><p>Not all stablecoins are created equal, and the differences matter for policymakers. We can identify three distinct models, each with different implications for monetary sovereignty, financial stability, and innovation.</p><h4>CBDCs: state-issued digital currency</h4><p>Central Bank Digital Currencies are the official digital version of a nation’s fiat currency. They are legal tender, backed by the central bank, and typically issued on permissioned (not publicly accessible) ledgers. Examples include the People’s Bank of China’s e-CNY and the Eastern Caribbean Central Bank’s DCash.</p><p>CBDCs offer maximum monetary sovereignty and financial stability. They are risk-free assets, direct liabilities of the central bank. But they also raise concerns about surveillance, disintermediation of commercial banks, and the capacity of central banks to manage retail payment systems.</p><h4>Regulated private stablecoins: the market approach</h4><p>Fiat-backed stablecoins like <a href="https://www.circle.com/usdc">USDC</a> and <a href="https://tether.to/en/transparency/">USDT</a> are issued by private entities, backed by reserves of fiat currency or equivalents, and operate on public blockchains. They are not legal tender, but they are increasingly regulated as electronic money or payment instruments.</p><p>This model prioritizes innovation and global reach. <a href="https://www.circle.com/usdc">Circle</a>, for example, holds licenses in 46 US states, the EU, UK, Singapore, Japan, and other jurisdictions. <a href="https://tether.to/en/transparency/">Tether</a> maintains over $183 billion in circulation with daily transparency reporting. But the model also introduces counterparty risk: users must trust that the issuer actually holds the reserves they claim.</p><h4>Bank-issued tokenized deposits: the hybrid model</h4><p>Some banks are issuing their own digital tokens representing deposits. <a href="https://www.jpmorgan.com/payments/newsroom/jpm-coin-usd-deposit-token-institutional-clients">JPMorgan’s JPM Coin</a> (JPMD) is a USD deposit token available on Base, the Ethereum Layer 2 blockchain. <a href="https://www.citigroup.com/global/news/press-release/2025/citi-integrates-citi-token-services-with-24-7-usd-clearing-real-time-cross-border-payments-liquidity-management">Citi Token Services</a> offers tokenized internal liquidity transfers integrated with 24/7 USD clearing.</p><p>These hybrid models combine the regulatory familiarity of banking with the efficiency of blockchain settlement. They are typically permissioned, available only to institutional clients, and backed by regulated bank deposits. They offer less openness than public stablecoins but more regulatory certainty.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*juyP2-7ywrRt6Hzu" /></figure><h3>Cryptos within DPI</h3><p>In most DPI frameworks, core building blocks already include digital identity, real-time payment rails, and interoperable data layers. Stablecoins can be positioned as a programmable settlement layer that operates on top of, or alongside, these components, enabling faster and more flexible value transfer while remaining anchored to public oversight and existing financial institutions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6EQDc414OiqBfH-1caVhhg.png" /></figure><p>From a governance and political perspective, the central concern is monetary sovereignty. Stablecoins embedded in DPI must not undermine a state’s control over its currency, capital flows, or financial stability. This implies a model where issuance, reserve management, and redemption are tightly regulated, and where public authorities retain visibility into systemic risks. In this sense, stablecoins function best as regulated digital representations of fiat value, used for settlement efficiency, rather than as parallel private monies competing with national currencies.</p><p>Legally and regulatorily, successful integration depends on clarity and coordination. Stablecoins intersect with multiple legal domains, including payments law, securities regulation, consumer protection, and cross-border supervision. For policymakers, the priority is not to create entirely new regimes, but to adapt existing financial regulations to tokenized forms of money. Clear rules on licensing, reserve backing, disclosure, and liability are essential to ensure trust, protect users, and enable cross-border interoperability between DPI systems.</p><p>Technologically, stablecoins introduce a layered architecture that aligns naturally with DPI principles. At the bottom sits the infrastructure layer, including blockchains, custody providers, and cybersecurity systems. Above this is a governance and trust layer, where compliance, identity, and auditability are enforced. On top, application layers enable public and private use cases such as remittances, government payments, or aid disbursement. This modular structure allows governments to adopt innovation incrementally, without committing to a single vendor or protocol.</p><p>Market structure and competition also matter. While stablecoins can lower barriers to entry and reduce reliance on legacy correspondent banking networks, they can also concentrate power if a small number of issuers or platforms dominate. Embedding stablecoins into DPI therefore requires open standards, interoperability requirements, and safeguards against excessive market concentration, ensuring that public infrastructure remains neutral and contestable.</p><p>Finally, social impact should frame all policy decisions. Stablecoins can improve inclusion, transparency, and efficiency, particularly in cross-border payments and public transfers. However, these benefits only materialize if users can safely access wallets, convert to local currency, and understand the systems they rely on. Without attention to usability, consumer protection, and digital literacy, the integration of crypto into DPI risks reinforcing existing inequalities rather than reducing them.</p><h3>Use cases and impact pathways</h3><p>Stablecoins are already being used in ways that matter for development. Understanding these use cases helps clarify where the technology offers genuine advantages and where the hype exceeds the reality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3go-eE71L54UfdV7v5FG6g.png" /><figcaption>Architecture of Blockchain embedded in DPI</figcaption></figure><h4>Cross-border payments and remittances</h4><p>This is the clearest use case. Traditional cross-border payments are slow (3–5 days), expensive (average 6–3% for remittances), and limited by declining correspondent banking relationships. Stablecoins can settle in minutes at a fraction of the cost.</p><p>The scale is significant. McKinsey estimates that stablecoins already facilitate about 3% of the $200 trillion in global cross-border payments. For remittance corridors where both sender and receiver have access to digital wallets, the savings can be substantial.</p><p><a href="https://www.circle.com/cpn">Circle Payments Network</a> exemplifies this approach, connecting banks and payment providers to enable 24/7 settlement via USDC. The network includes partners across Latin America, Asia, and Africa, facilitating payments that would be slow or impossible through traditional rails.</p><h4>Treasury and B2B payments</h4><p>Businesses are increasingly using stablecoins for treasury management and cross-border supplier payments. The benefits include faster settlement, reduced working capital requirements, and 24/7 availability.</p><p><a href="https://www.jpmorgan.com/payments/newsroom/jpm-coin-usd-deposit-token-institutional-clients">JPMorgan’s JPM Coin</a> targets this use case, offering institutional clients near-instant settlement on public blockchain. <a href="https://www.citigroup.com/global/news/press-release/2025/citi-integrates-citi-token-services-with-24-7-usd-clearing-real-time-cross-border-payments-liquidity-management">Citi Token Services</a> enables corporate clients to manage liquidity across multiple banks in real time.</p><h4>Humanitarian and aid disbursements</h4><p>Several NGOs have piloted stablecoin-based aid disbursements, citing speed, transparency, and reduced leakage compared to cash distributions. The ability to track funds on-chain and verify receipt provides accountability benefits.</p><p>However, this use case also highlights risks. Recipients need safe ways to convert stablecoins to local currency. Volatility in crypto-collateralized stablecoins can wipe out value. And the technical complexity of blockchain systems can create barriers for vulnerable populations.</p><h3>Risk factors to monitor</h3><p>Development practitioners should be aware of several risks:</p><p><strong>Volatility in non-fiat-backed stablecoins.</strong> Crypto-collateralized and algorithmic stablecoins have experienced significant depegging events. The 2022 collapse of Terra’s UST wiped out billions in value.</p><p><strong>Regulatory uncertainty.</strong> The landscape is evolving rapidly. Today’s compliant operation may face new restrictions tomorrow.</p><p><strong>Technical infrastructure requirements.</strong> Using stablecoins requires internet access, digital devices, and technical knowledge that may exclude vulnerable populations.</p><p><strong>Digital literacy gaps.</strong> Without proper understanding, users may fall victim to scams, lose private keys, or make irreversible errors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LEcAs9D6H0YIzY5g" /></figure><h3>Landscape: Regulatory inflections</h3><p>The regulatory landscape for stablecoins shifted dramatically. Frameworks that define what stablecoins are, who can issue them, and how they must operate.</p><h4>The United States: GENIUS Act</h4><p>The Guiding and Establishing National Innovation for U.S. Stablecoins (GENIUS) Act, passed in July 2025, creates the 1st comprehensive federal framework for payment stablecoins in the US. Key provisions include:</p><p><strong>Reserve requirements:</strong> Issuers must hold high-quality liquid assets, safeguarded by qualified custodians</p><p><strong>Transparency:</strong> Reserve disclosures and published redemption policies</p><p><strong>Consumer protection:</strong> Bankruptcy provisions ensuring stablecoin holders have priority claims on reserve assets</p><p><strong>Licensing:</strong> Pathway for both federal and state-chartered issuers</p><p><strong>Restrictions:</strong> Prohibition on offering yield or interest to avoid competing with bank deposits</p><p>The Act represents a shift from defensive caution to strategic engagement. As one analysis noted, policymakers who once dismissed blockchain projects as distractions are now building frameworks around them.</p><h4>The European Union: MiCA</h4><p>The Markets in Crypto-Assets Regulation (MiCA), implemented in December 2024, provides a comprehensive framework across the EU. For stablecoins (classified as “e-money tokens”), MiCA requires:</p><ul><li>Issuance only by regulated e-money institutions or credit institutions</li><li>Full backing by fiat currency reserves</li><li>Right of redemption at par value</li><li>Robust governance and risk management frameworks</li></ul><p><a href="https://www.circle.com/usdc">Circle</a> was the first major global stablecoin issuer to achieve MiCA compliance, securing an Electronic Money Institution license from France’s ACPR.</p><h4>Asia-Pacific developments</h4><p>Singapore and Hong Kong have emerged as leaders in stablecoin regulation. Singapore’s Monetary Authority requires Major Payment Institution licenses for stablecoin issuers. Hong Kong’s Stablecoin Ordinance, passed in May 2025, mandates licensing for all HKD-backed stablecoin issuers with strict reserve and audit requirements.</p><h4>Implications for development contexts</h4><p>For emerging markets, these frameworks present both opportunities and challenges. On one hand, regulatory clarity enables safer participation in the global stablecoin ecosystem. On the other hand, compliance requirements may be burdensome for smaller institutions or developing regulatory capacities.</p><p>The risk of regulatory arbitrage is real. If major jurisdictions impose strict requirements while others do not, activity may migrate to less regulated venues. This creates challenges for consumer protection and financial stability. It also raises questions about how development organizations can support capacity building for effective stablecoin regulation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b3f6f449050c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Securing the Frontier: A Technical Deep Dive into AI Trust & Safety]]></title>
            <link>https://medium.com/@dtr0105/securing-the-frontier-a-technical-deep-dive-into-ai-trust-safety-99dfc42649f1?source=rss-13709fdd2256------2</link>
            <guid isPermaLink="false">https://medium.com/p/99dfc42649f1</guid>
            <category><![CDATA[trustworthy-ai]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[ai-infrastructure]]></category>
            <dc:creator><![CDATA[Tianrong Ding Terisa]]></dc:creator>
            <pubDate>Mon, 23 Feb 2026 04:11:51 GMT</pubDate>
            <atom:updated>2026-02-23T04:11:51.214Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial intelligence is now making decisions that affect millions of lives. From healthcare diagnostics to autonomous vehicles, AI systems operate in domains where errors carry catastrophic consequences. Yet these systems often function as black boxes, their decision-making processes opaque even to their creators. This transparency gap creates a fundamental challenge: how do we trust systems we cannot fully understand or audit?</p><p>This technical review examines the cutting-edge approaches, theories, and implementations shaping AI trust and safety in 2026. We will explore the frameworks that make AI systems interpretable, the standards governing safety-critical applications, and the emerging technologies that promise to bridge the gap between AI capabilities and human accountability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/705/1*AE4ZoUN883GW1WNPlbKXuw.png" /></figure><h3>The Five Pillars of Trustworthy AI</h3><p>Trustworthy AI rests on five pillars that together create systems humans can rely on. These pillars provide a framework for evaluating AI systems before deployment in high-stakes environments.</p><p><strong>Explainability</strong> sits at the core. It means AI systems can articulate why they made specific decisions in terms humans understand. When a loan application is denied or a medical diagnosis is suggested, stakeholders need explanations they can evaluate and challenge. Without explainability, AI becomes an oracle rather than a tool, demanding faith instead of enabling informed judgment. <a href="https://www.sciencedirect.com/science/article/pii/S2405959525001584">Research from ScienceDirect</a> shows that explainability is particularly critical in cybersecurity applications where analysts must verify whether AI detections represent genuine threats or false positives.</p><p><strong>Bias and fairness</strong> addresses the reality that AI systems trained on historical data often replicate historical inequities. Fairness means ensuring AI decisions don’t discriminate against protected groups and that outcomes are equitable across populations. This requires both technical solutions (detecting and mitigating bias in training data) and governance processes (ongoing monitoring for discriminatory patterns).</p><p><strong>Reproducibility</strong> ensures that AI systems produce consistent results across different runs and environments. If an AI model produces different outputs when retrained on the same data, or behaves differently in production than in testing, it can’t be trusted for critical decisions. Reproducibility requires careful version control, documentation of training procedures, and standardized evaluation protocols.</p><p><strong>Sustainability</strong> considers the environmental and long-term viability of AI systems. Training large models consumes enormous energy resources. Sustainable AI means optimizing models for efficiency, considering the carbon footprint of AI deployments, and ensuring systems can be maintained and updated over their operational lifetime.</p><p><strong>Transparency</strong> encompasses openness about AI capabilities, limitations, data sources, and decision-making processes. Transparent AI systems clearly communicate what they can and can’t do, what data they were trained on, and how they should be appropriately used. This enables informed decisions about when AI is appropriate and when human judgment should prevail.</p><p>These five pillars align closely with UNDP’s approach to <a href="https://www.undp.org/digital/blog/how-digital-standards-are-helping-undp-build-inclusive-digital-societies">human-centric digitalization</a>. The organization’s <a href="https://www.undp.org/digital/news/un-releases-universal-dpi-safeguards-framework-promote-safe-and-inclusive-digital-public-infrastructure">Universal DPI Safeguards</a> framework emphasizes that technology should serve people rather than the reverse. Trustworthy AI is not a technical nicety but a prerequisite for inclusive digital transformation that leaves no one behind.</p><h3>Explainable AI: Decoding the Black Box</h3><p>The challenge of explainable AI (XAI) has become more urgent as deep learning models grow increasingly complex. Modern neural networks may contain billions of parameters, making it nearly impossible to trace how specific inputs lead to specific outputs. This “black box” nature creates problems in high-stakes domains where decisions must be defensible. <a href="https://www.sciencedirect.com/science/article/pii/S2405959525001584">Academic research</a> documents the growing urgency of this challenge across healthcare, finance, and public safety applications.</p><p>The XAI field has developed several approaches to this challenge. <strong>Intrinsic explainability</strong> involves building models that are interpretable by design. Decision trees, rule-based systems, and linear regression models provide transparency because their decision logic is explicit. While these models may sacrifice some predictive power compared to deep learning, they offer the advantage of being fully auditable. In regulated industries like healthcare and finance, interpretable models often remain the preferred choice despite their limitations. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC10879008/">Healthcare AI ethics research</a> highlights how regulatory requirements drive demand for interpretable models in clinical settings.</p><p><strong>Post-hoc explanation methods</strong> take a different approach. Rather than changing the model architecture, these techniques analyze trained models to generate explanations after the fact. <a href="https://github.com/marcotcr/lime">LIME</a> (Local Interpretable Model-agnostic Explanations) works by approximating complex models with simpler, interpretable ones in the neighborhood of specific predictions. <a href="https://shap.readthedocs.io/en/latest/">SHAP</a> (SHapley Additive exPlanations) uses game theory to attribute predictions to specific input features, showing which factors most influenced a decision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*snh2TgqRMlfYaEOp" /></figure><p>Attention mechanisms in neural networks provide another window into model behavior. By visualizing which parts of an input the model “focused on” when making a decision, attention maps can reveal whether models are using relevant features or spurious correlations. However, attention visualization has limitations: high attention doesn’t always mean high importance, and attention maps can be misleading.</p><p>The distinction between <strong>local and global explanations</strong> matters for practical applications. Local explanations focus on individual predictions, answering “why did the model make this specific decision?” Global explanations characterize overall model behavior, showing which features matter most across all predictions. Both are necessary: local explanations for debugging specific errors, global explanations for understanding systemic biases. The <a href="https://www.sciencedirect.com/science/article/pii/S2405959525001584">XAI research literature</a> emphasizes that effective deployment requires both types of explanations working together.</p><p>The cybersecurity domain illustrates XAI’s practical value. The <a href="https://www.sciencedirect.com/science/article/pii/S2405959525001584">ScienceDirect review</a> highlights how explainable AI helps security analysts understand why systems flag certain network traffic as suspicious. Without explanations, analysts cannot verify whether AI detections represent genuine threats or false positives. XAI tools like <a href="https://www.ibm.com/products/watson-openscale">IBM Watson OpenScale</a>, <a href="https://interpret.ml/">Microsoft InterpretML</a>, and <a href="https://cloud.google.com/explainable-ai">Google Cloud Explainable AI</a> now provide enterprise-grade explanation capabilities.</p><p>The global XAI market reflects growing demand for interpretability. According to <a href="https://www.sciencedirect.com/science/article/pii/S2405959525001584">market research</a>, the XAI market is projected to grow from $6.4 billion in 2024 to $34.6 billion by 2033, representing a compound annual growth rate of 18.4%. This growth signals recognition across industries that AI adoption requires explainability.</p><h3>Safety-critical AI Systems</h3><p>Safety-critical AI operates in domains where failures cause catastrophic harm: autonomous vehicles, aircraft collision avoidance, railway signaling, medical devices, and industrial robotics. These applications demand reliability standards far exceeding those of typical AI systems.</p><p>The safety gap between AI capabilities and requirements is stark. <a href="https://dl.acm.org/doi/full/10.1145/3626314">Traditional functional safety standards</a> like IEC 61508 require failure probabilities as low as 10^-9 per hour of operation, equivalent to one dangerous failure every 114 years. Current AI systems typically achieve 99% accuracy, which sounds impressive until you realize it means one failure per 100 decisions. For a self-driving car making thousands of decisions per hour, 99% accuracy is catastrophically inadequate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fa9K2tDCcxibUEhl" /></figure><p>This gap has driven development of new standards specifically for AI-based safety systems. <a href="https://www.iso.org/standard/70939.html">ISO/PAS 21448</a> (Safety of the Intended Functionality, or SOTIF) addresses the challenge of ensuring autonomous systems behave safely even when operating correctly according to their specifications. <a href="https://ul.org/news/ul-4600-standard-safety-autonomous-products">UL 4600</a> provides a framework for evaluating the safety of autonomous systems that rely on machine learning. These standards represent attempts to adapt traditional safety engineering to AI’s unique characteristics. The <a href="https://dl.acm.org/doi/full/10.1145/3626314">ACM Computing Surveys review</a> of AI for safety-critical systems provides comprehensive coverage of these standards.</p><p>Several key properties define trustworthy ML systems for safety-critical applications:</p><p><strong>Auditability</strong> enables independent examination of development and verification processes. Safety assessors must be able to review how systems were built, tested, and validated. <a href="https://dl.acm.org/doi/full/10.1145/3626314">Safety standards research</a> identifies auditability as a core requirement for ML systems in safety-critical applications.</p><p><strong>Data quality</strong> ensures training data is free from defects and possesses desired characteristics. <a href="https://dl.acm.org/doi/full/10.1145/3626314">Safety-critical systems research</a> shows that poor training data is a leading cause of AI system failures in production.</p><p><strong>Monitorability</strong> allows operators to discriminate correct from incorrect behavior in real-time. Safety-critical systems need mechanisms to detect when they are operating outside their design parameters.</p><p><strong>Provability</strong> provides mathematical guarantees that certain properties are satisfied. Formal verification techniques can prove that systems will not enter unsafe states under specified conditions.</p><p><strong>Robustness</strong> ensures systems perform their intended function even with abnormal or unknown inputs. This includes handling sensor failures, adversarial attacks, and environmental conditions not seen during training.</p><p>Runtime monitoring represents a critical safety mechanism. “Circuit breaker” systems can halt autonomous operations when anomalies are detected, returning control to human operators or safe default states. The <a href="https://dl.acm.org/doi/full/10.1145/3626314">ACM Computing Surveys review</a> of AI for safety-critical systems emphasizes that runtime monitoring is essential because pre-deployment testing cannot cover all possible operational scenarios.</p><h3>Emerging Technical Approaches</h3><p>The field of AI trust and safety continues to evolve rapidly. Several emerging approaches show particular promise for addressing current limitations.</p><p><strong>Autonomous AI agents</strong> present new trust challenges. Unlike traditional AI systems that respond to individual queries, autonomous agents can take sequences of actions over extended periods, interacting with multiple systems and making complex decisions. Multi-agent systems, where multiple AI agents interact, introduce emergent behaviors that are difficult to predict from individual agent specifications. Research highlights the need for new governance frameworks that can handle distributed responsibility across agent interactions.</p><p><strong>Federated learning</strong> enables AI model training across distributed data sources without centralizing sensitive data. This is particularly valuable in healthcare and financial services, where privacy regulations restrict data sharing. However, federated learning introduces new trust challenges: how do you verify that remote training nodes are providing accurate gradients rather than attempting to poison the model? How do you maintain explainability when models are trained across thousands of devices?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OrtE-TczkDWCWuUJ" /></figure><p><strong>Constitutional AI</strong> represents an attempt to build ethical constraints directly into AI systems. Rather than training models solely to maximize task performance, constitutional AI trains models to follow explicit ethical principles. <a href="https://www.anthropic.com/research/constitutional-classifiers">Anthropic’s research</a> in this area explores how AI systems can be trained to recognize and refuse harmful requests while remaining helpful for legitimate applications.</p><p><strong>Circuit tracing</strong> and mechanistic interpretability aim to understand the internal computations of neural networks at a granular level. <a href="https://transformer-circuits.pub/2025/attribution-graphs/methods.html">Research from Anthropic</a> and other labs has demonstrated techniques for identifying specific circuits within large language models that handle particular capabilities. This work promises to move beyond correlational explanations to genuine understanding of how AI systems process information.</p><p><strong>Deliberative alignment</strong> trains AI systems to reason explicitly about safety considerations before producing outputs. <a href="https://arxiv.org/abs/2412.16339">OpenAI’s research</a> in this area shows that training models to deliberate about potential harms can reduce harmful outputs without excessive refusals of benign requests. This approach treats safety not as a filter applied to outputs but as an integral part of the reasoning process.</p><p>The pace of research in these areas is accelerating. According to analysis from <a href="https://www.lesswrong.com/posts/Wti4Wr7Cf5ma3FGWa/shallow-review-of-technical-ai-safety-2025-2">LessWrong’s 2025 technical AI safety review</a>, major AI labs now employ hundreds of researchers working on safety and alignment. OpenAI’s safety teams include Alignment, Safety Systems, Interpretability, Preparedness, and Model Policy groups. Anthropic has dedicated teams for Scalable Alignment, Interpretability, Control, and Alignment Stress-Testing. Google DeepMind maintains teams for Amplified Oversight, Interpretability, and Frontier Safety Risk Assessment.</p><h3>Implementation System &amp; Tools</h3><p>Translating trust and safety research into practice requires tools and frameworks that development teams can actually use. Several open-source and commercial platforms have emerged to address this need.</p><p>The <a href="https://vectorinstitute.ai/ai-trust-and-safety/resources-projects-tools-frameworks-and-best-practices/">Vector Institute</a> has developed several tools for trustworthy AI implementation. <a href="https://unbias-mkdocs.readthedocs.io/en/latest/"><strong>UnBIAS</strong></a> is a text analysis and debiasing toolkit that assesses and corrects biases in textual data. <a href="https://github.com/VectorInstitute/flex_model"><strong>FlexModel</strong></a> provides interpretability capabilities for distributed large language models running across multiple GPUs. <strong>CyclOps</strong> offers evaluation and monitoring tools specifically designed for healthcare machine learning applications. <strong>FL4Health</strong> enables federated learning research in clinical settings, addressing the challenge of training on distributed medical data while preserving privacy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R0LQwlXiywMEKer_" /></figure><p>Commercial platforms have also emerged to address enterprise needs. <a href="https://www.ibm.com/products/watson-openscale">IBM Watson OpenScale</a> provides real-time AI model assessment and explanation capabilities integrated into security frameworks. <a href="https://interpret.ml/">Microsoft InterpretML</a> offers explainable machine learning models applicable to malware detection and threat classification. <a href="https://cloud.google.com/explainable-ai">Google Cloud Explainable AI</a> provides feature attribution and interpretation techniques deployable on cloud AI models.</p><p>Risk management frameworks help organizations systematically address AI safety. <a href="https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf">OpenAI’s Preparedness Framework</a> establishes evaluation protocols for models with potentially dangerous capabilities. <a href="https://deepmind.google/blog/strengthening-our-frontier-safety-framework/">Google DeepMind’s Frontier Safety Framework</a> defines thresholds for when models require enhanced safety measures. Anthropic’s Responsible Scaling Policy commits the organization to specific safety investments as models become more capable.</p><p>These frameworks share common elements: they establish evaluation criteria for identifying risky capabilities, define safety measures proportional to risk levels, and commit organizations to specific actions before deploying powerful systems. While implementation details vary, the underlying principle is consistent: safety should scale with capability.</p><h3>Governance and Outlook</h3><p>Technical solutions to AI trust and safety must operate within policy and governance frameworks. The regulatory landscape for AI is evolving rapidly, with different jurisdictions taking different approaches.</p><p>The United Kingdom has adopted a <a href="https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper">pro-innovation framework</a> built on five principles: safety and security, appropriate transparency and explainability, fairness, accountability and governance, and contestability and redress. Rather than creating a new AI regulator, the UK approach empowers existing sectoral regulators to apply these principles within their domains. This framework emphasizes context-based regulation rather than technology-specific rules, recognizing that the same AI technology can pose very different risks in different applications.</p><p>The European Union’s <a href="https://artificial-intelligence-act.com/">AI Act</a> takes a risk-based approach, classifying AI applications into prohibited, high-risk, and limited-risk categories with corresponding regulatory requirements. High-risk applications, including AI used in critical infrastructure, education, and law enforcement, face strict requirements for risk management, data quality, transparency, and human oversight.</p><p>GDPR’s “right to explanation” provides individuals with the right to understand automated decisions affecting them. While the precise scope of this right remains debated, it has driven investment in explainable AI among organizations operating in Europe.</p><p>The <a href="https://www.undp.org/digital/news/un-releases-universal-dpi-safeguards-framework-promote-safe-and-inclusive-digital-public-infrastructure">Universal DPI Safeguards</a> framework offers a complementary perspective focused on human rights and inclusive development. This framework emphasizes that <a href="https://www.undp.org/digital/digital-public-infrastructure">digital public infrastructure</a>, including AI systems, should be designed to protect human rights, promote equity, and ensure that technological benefits are broadly shared. This approach aligns with the organization’s broader mission of eradicating poverty while protecting the planet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lr80nwVXzqXoYvSZ" /></figure><p>Looking forward, several challenges and opportunities will shape AI trust and safety. Real-time explainability for autonomous systems remains an open problem: how do you provide meaningful explanations when decisions must be made in milliseconds? Standardizing safety evaluations across industries could enable more consistent practices and easier comparison of different approaches. Addressing the “half-life of AI governance knowledge” (estimated at 24–36 months by AI governance research) requires building adaptive governance frameworks that can evolve as quickly as the technology itself.</p><p>Research gaps remain, particularly regarding AI trust and safety in developing economies. Most research focuses on contexts with abundant computational resources, skilled technical personnel, and established regulatory institutions. Extending trustworthy AI to resource-constrained environments is both a technical challenge and a development imperative.</p><h3>Building Trust in AI Systems</h3><p>AI trust and safety is not a single technology but a multidisciplinary field combining machine learning research, safety engineering, ethics, and governance. The five pillars of trustworthy AI provide a framework for evaluating systems before deployment.</p><p>For development practitioners, the key insight is that trustworthy AI is achievable but requires intentional investment. It can’t be bolted on after development but must be built in from the start. Organizations deploying AI in high-stakes contexts should evaluate systems against established frameworks, implement appropriate monitoring and oversight mechanisms, and maintain human accountability for AI-assisted decisions.</p><p>The cutting-edge technologies reviewed here represent significant progress, but the field continues to evolve. Organizations should monitor developments in autonomous agent governance, federated learning, constitutional AI, and mechanistic interpretability. These emerging approaches promise to address current limitations and enable new applications of AI in service of development goals.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99dfc42649f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Open-Source Engine: Accelerating AI and Digital Development Through Collaborative Frameworks]]></title>
            <link>https://medium.com/@dtr0105/the-open-source-engine-accelerating-ai-and-digital-development-through-collaborative-frameworks-a807aea62952?source=rss-13709fdd2256------2</link>
            <guid isPermaLink="false">https://medium.com/p/a807aea62952</guid>
            <category><![CDATA[international-development]]></category>
            <category><![CDATA[ecosystem-building]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[digital-infrastructure]]></category>
            <dc:creator><![CDATA[Tianrong Ding Terisa]]></dc:creator>
            <pubDate>Mon, 23 Feb 2026 01:16:37 GMT</pubDate>
            <atom:updated>2026-02-23T01:16:37.468Z</atom:updated>
            <content:encoded><![CDATA[<p>The rapid advancement in artificial intelligence and digital technologies is undeniably fueled by a powerful, often unsung hero: the open-source ecosystem. Far from being a chaotic free-for-all, this ecosystem operates on a sophisticated interplay of technical mechanisms and guiding principles, accelerated by strategic funding and policy frameworks.</p><p>At its core, the open-source ecosystem thrives on <strong>collaboration and transparency</strong>. Openly sharing allows developers worldwide to inspect, modify, and distribute it. This shared ownership fosters a collective intelligence that rapidly identifies bugs, develops new features, and iterates on existing solutions at a pace unmatched by proprietary models.</p><p><strong>Key Mechanisms Driving Open-Source Acceleration:</strong></p><ol><li><strong>Distributed Version Control Systems (DVCS):</strong> Tools like Git and platforms like GitHub, GitLab, and Bitbucket are the circulatory system of open-source development. They enable geographically dispersed teams to work on the same codebase simultaneously, track changes, merge contributions, and revert to previous versions seamlessly. This distributed nature significantly reduces bottlenecks and empowers continuous integration and delivery.</li><li><strong>Modular Architectures and APIs:</strong> Open-source projects often adopt modular designs, breaking down complex systems into smaller, interchangeable components. This, coupled with well-defined Application Programming Interfaces (APIs), allows developers to build upon existing solutions without needing to understand the entire system from scratch. This reusability is a cornerstone of rapid prototyping and innovation in AI, where pre-trained models and libraries (e.g., TensorFlow, PyTorch) are frequently leveraged.</li><li><strong>Community-Driven Peer Review:</strong> Unlike closed development cycles, open-source projects benefit from a constant stream of peer review. Every line of code, every proposed feature, and every bug fix is subject to scrutiny from a diverse community of experts. This rigorous process significantly enhances code quality, security, and robustness.</li><li><strong>Licensing Models:</strong> Open-source licenses (e.g., MIT, Apache, GPL) define the terms under which software can be used, modified, and distributed. While varied, they all uphold the fundamental principle of freedom to use and distribute, ensuring that innovation remains accessible and shareable.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_3Ho4XAEc4ThG0G5f5I6Yw.png" /><figcaption>The open-source engine: A framework for accelerating AI &amp; digital advancement</figcaption></figure><p><strong>Catalytic Roles:</strong></p><ol><li><strong>Standards:</strong> Open standards are critical for interoperability and prevent vendor lock-in. In AI, initiatives like ONNX (Open Neural Network Exchange) allow for model portability across different frameworks, fostering greater flexibility and collaboration. Governments and industry consortia play a vital role in promoting and sometimes mandating the adoption of open standards.</li><li><strong>Development Funding:</strong> While often driven by volunteer efforts, significant funding — from philanthropic organizations, government grants, and corporate sponsorship — is essential for sustaining large-scale open-source projects. This funding supports core maintainers, infrastructure, and strategic initiatives that might not otherwise be viable.</li><li><strong>Policy Frameworks:</strong> Governments and regulatory bodies are increasingly recognizing the strategic importance of open source. Policies encouraging the use of open-source software in public sector procurement, mandating open data standards, and investing in open-source talent development are becoming more common. Furthermore, intellectual property policies that balance protection with the promotion of open innovation are crucial for a healthy ecosystem.</li></ol><p>The synergistic combination of these mechanisms, principles, and external accelerators creates a self-reinforcing loop that drives exponential growth in AI and digital development. It’s a testament to the power of collective intelligence and a model for how complex societal challenges can be tackled through collaborative innovation.</p><h3>The technical foundation: Open source as interoperable building blocks</h3><p>At its core, open source provides the interoperable building blocks that make DPI possible. The Linux Foundation, steward of foundational technologies from Linux to Kubernetes to PyTorch, has identified five essential factors for successful open source DPI projects: Strong technical foundations, Active developer communities; Effective communication and branding; Long-term sustainability and; Large-scale adoption by enterprises and governments.</p><p>These factors reinforce each other: Strong technical foundations attract developers. Active communities improve the code. Clear communication drives adoption. Adoption creates sustainability. And sustainability enables further technical investment.</p><p>The approach works on 2 levels: Horizontally, organizations like the Linux Foundation steward fundamental building blocks: operating systems, container orchestration, programming languages, and AI frameworks. Vertically, sector-specific stakeholders address particular domains. Financial institutions focuses on financial services; Energy providers addresses power grids; And other foundations tackle telecommunications and education, etc.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/653/1*teBf6-x9yGNpUTXLpm_8vw.png" /></figure><p>This layered architecture matters for governments. It means they are not starting from scratch. They can combine proven components rather than building monolithic systems. They benefit from the “many eyes” security model, where global communities identify and resolve vulnerabilities faster than any single vendor could. And they gain transparency, which builds the public trust essential for widespread adoption of digital services.</p><h3>Policy Enablement: How technical openness creates strategic options</h3><p>Technical openness creates policy options that closed systems cannot match. This is particularly evident in the growing emphasis on digital sovereignty. Closed systems create dependency, while open systems create capacity. Open source acts as a “critical force multiplier that enables Europe to do more with less.” This perspective is gaining traction globally. Governments are recognizing that proprietary systems, however convenient in the short term, create long-term strategic vulnerabilities. When critical infrastructure depends on foreign vendors, so does national autonomy.</p><p>The policy response has 5 components. First, leverage public sector buying power to scale open source AI. Government procurement is not just about acquiring technology. It shapes market dynamics and signals strategic priorities. When governments specify open standards and open source solutions, vendors adapt.</p><p>Second, mobilize funding to incubate and maintain the open source AI stack. Sustainable open source requires investment. This means supporting not just initial development, but ongoing maintenance, security updates, and community building.</p><p>Third, facilitate access to computing infrastructure for open source AI research and development. AI training requires significant computational resources. Ensuring broad access prevents concentration of AI capabilities in a few large corporations.</p><p>Fourth, unlock data for open source AI development while protecting privacy. Data is the fuel for AI. Creating open datasets, with appropriate safeguards, enables broader participation in AI development.</p><p>Fifth, build capacity in public and private sectors to leverage open source. Technology alone is insufficient. People need skills to deploy, customize, and maintain open systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GayNftVC4ft11GINz1ihqg.png" /></figure><p>This is where institutional innovations like Open Source Programme Offices (OSPOs) become critical. OSPOs provide the organizational structure for governments to engage systematically with open source. They coordinate contributions, manage compliance, build community relationships, and ensure sustainability. Kenya recently launched its National OSPO at the Global DPI Summit in Cape Town, establishing a central coordinating body for government digital assets. Trinidad and Tobago has established an OSPO at the University of the West Indies.</p><p>The <a href="https://www.undp.org/digital/osee">OSEE initiative</a>, led by ITU and UNDP with EU support, is catalyzing this trend. Over three and a half years, OSEE is enhancing capacity for local and regional actors to adopt open source through training frameworks, technical facilities, and a global knowledge hub.</p><h3>A framework for understanding country approaches</h3><p>Countries adopting open source DPI typically fall along two dimensions: <strong>maturity of digital ecosystem</strong> (nascent vs. advanced) and <strong>primary motivation</strong> (efficiency vs. inclusion). This creates four quadrants that help explain strategic choices:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YXtSNYZMkSVlaRM3" /><figcaption>Framework for understanding country approaches to opensource DPI based on ecosystem maturity and primary motivation</figcaption></figure><p>Singapore represents the “advanced ecosystem + efficiency focus” quadrant, using open source to optimize government operations and export proven solutions globally. Brazil occupies the “nascent-to-advanced ecosystem + inclusion focus” quadrant, leveraging open source to close service gaps for marginalized populations while building toward systemic transformation.</p><h3>Singapore: Exporting proven solutions through the Digital Gateway</h3><p>Singapore launched the <a href="https://www.digitalgateway.gov.sg/">Singapore Digital Gateway (SGDG)</a> in September 2025, consolidating over 30 digital resources including open-source tools, governance frameworks, and training programs for international sharing. This platform embodies Singapore’s philosophy of treating DPI expertise as a global public good rather than proprietary advantage.</p><p>The tools available through SGDG demonstrate the depth of Singapore’s open source commitment:</p><ul><li><strong>FormSG</strong>: A secure digital form builder already adopted by Cambodia and Sri Lanka for government services</li><li><strong>Isomer</strong>: A static site generator that reduces website launch time from months to 36 hours</li><li><strong>OpenAttestation</strong>: A blockchain-based framework for tamper-resistant document verification</li><li><strong>Purple HATS</strong>: An automated accessibility testing tool accredited as a Digital Public Good</li></ul><p>Singapore’s <a href="https://www.open.gov.sg/">Open Government Products (OGP)</a> team, which joined the Digital Public Goods Alliance in March 2025, operates on a “tech for the public good” philosophy. Their approach emphasizes experimentation, iteration, and willingness to pivot. The annual “Hack for Public Good” hackathon prototypes solutions that often become production systems.</p><p>The results are measurable. Singapore’s open source legal tool summarizes case law in minutes rather than hours. The Maritime Digital Twin, launched in March 2025, creates a dynamic virtual model of the Port of Singapore using open data standards. The Open Digital Platform orchestrates real-time data across smart districts.</p><p>What distinguishes Singapore’s approach is systematic knowledge transfer. Through partnerships with the World Bank and UNDP, Singapore co-creates training programs that help other countries adapt these tools. Cambodia has implemented five Singapore tools and contributed enhancements back to the community. This is the virtuous cycle in action: build, share, adapt, improve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JbR7RC-e9fZGGGep" /><figcaption>Singapore’s suite of open-source digital tools available through the Digital Gateway for global adaptation</figcaption></figure><h3>Brazil: Scaling inclusion through participatory digital infrastructure</h3><p>Brazil’s approach centers on using open source DPI to address inequality while building institutional capacity across federal, state, and municipal levels. The <a href="https://www.gov.br/">Gov.br platform</a> now serves 156 million users with over 4,200 online services, representing one of Latin America’s most mature GovTech implementations. Several initiatives illustrate Brazil’s inclusion-focused model:</p><p><strong>Verifiable Credentials for Rural Landowners</strong>: Launched in 2025, this program uses the DPI as a Service (DaaS) framework to issue digital credentials to 7 million rural landowners. These credentials enable seamless document sharing with banks and loan providers, addressing a critical barrier to financial inclusion for agricultural communities.</p><p><strong>Participatory Budgeting at Scale</strong>: Piauí state’s digital participatory budgeting platform has engaged 160,000+ citizens in voting on over R$80 million in public investments. The <a href="https://brasilparticipativo.gov.br/">Brasil Participativo</a> platform generated 8,254 citizen proposals with 4 million accesses during the recent four-year planning cycle.</p><p><strong>Environmental DPI</strong>: The Cadastro Ambiental Rural (CAR) serves as digital public infrastructure for environmental management, supporting Brazil’s climate strategy ahead of <a href="https://www.weforum.org/stories/2025/04/brazil-govtech-digital-public-infrastructure-development/">COP30 in 2025</a>. MapBiomas provides annual land cover mapping as a Digital Public Good using open licenses.</p><p>Brazil’s <a href="https://softwarepublico.gov.br/">Software Público Brasil</a> portal enables sharing and reuse of public software across government levels. The Computers for All program promotes free and open-source software in affordable computers. The ProInfo initiative brings open source educational technology to schools nationwide.</p><p>What distinguishes Brazil’s approach is multi-level governance. While Singapore operates primarily at the national level with centralized coordination, Brazil coordinates across federal, state, and municipal governments. The Tribunal de Contas da União enables citizen monitoring of roadworks with geo-tagged feedback. Niterói’s multichannel system schedules vaccines, childcare, and social assistance via app, web, and WhatsApp.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zCgNZ3u2PDpIB3Pf" /><figcaption>Brazil’s inclusion-focused DPI model demonstrating scale across multiple governance levels and areas</figcaption></figure><p>Both countries demonstrate that open source DPI is not one-size-fits-all. Singapore’s efficiency-focused model works because of advanced technical capacity and centralized governance. Brazil’s inclusion-focused model succeeds through distributed innovation and participatory design. The common thread is treating digital infrastructure as a public good that improves through sharing and adaptation.</p><p>The inclusion dimension extends beyond national borders. Open source enables localization for languages and contexts that proprietary vendors might ignore. Africa’s diverse linguistic and cultural landscape requires adaptable technology. Open source provides the flexibility to meet these needs.</p><p>Safeguards are critical to this vision. The Universal DPI Safeguards Framework, developed through multi-stakeholder processes, ensures that digital public infrastructure respects human rights, protects privacy, and promotes inclusion. Transparency in open source systems makes such safeguards auditable and enforceable.</p><h3>Closing the loop: The self-reinforcing ecosystem</h3><p>The four dimensions, technical, policy, economic, and social, form a self-reinforcing cycle. Technical innovation creates capabilities that enable policy confidence. Policy confidence attracts economic investment. Economic investment delivers social impact. Social impact creates demand for further technical innovation.</p><p>The <a href="https://www.undp.org/digital/news/50-5-milestone-event-illustrates-how-implement-new-global-digital-compact-safely-and-inclusively">50-in-5 campaign</a> exemplifies this approach. By supporting 50 countries to implement DPI within 5 years, the campaign is accelerating the virtuous cycle at scale. Technical assistance, peer learning, and resource mobilization combine to help countries leapfrog traditional development pathways.</p><p>Several trends will shape this ecosystem in coming years. China’s growing leadership in open-source AI models is creating new options for countries seeking alternatives to Western-dominated technology stacks. Agentic systems requiring interoperability will drive demand for open standards. Data commons for public interest AI will expand access to training data. And new institutional models like Sovereign Tech Funds will provide sustainable financing for open source maintenance.</p><h3>Building open digital infrastructure for the public good</h3><p>The case for open source in AI and digital public infrastructure is no longer theoretical. It is demonstrated in India’s startup ecosystem, Singapore’s Digital Gateway, Brazil’s participatory platforms, Kenya’s OSPO launch, and countless other examples worldwide.</p><p>The virtuous cycle is real. Technical foundations enable policy innovation. Policy innovation attracts economic investment. Economic investment delivers social impact. And social impact creates demand for further technical advancement.</p><p>For policymakers, the imperative is clear. Invest in open source capacity. Establish OSPOs. Support communities. And prioritize interoperability and transparency in procurement decisions.</p><p>For development practitioners, the opportunity is equally clear. The tools and frameworks exist. The knowledge is accumulating. The partnerships are forming. What remains is execution at scale.</p><p>The airport analogy holds. Just as airports enable airlines to serve passengers, open digital infrastructure enables innovators to serve citizens. The question is not whether to build this infrastructure, but how to build it right: open, interoperable, and in service of the public good.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a807aea62952" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI trust and safety strategies across multiple intervention level]]></title>
            <link>https://medium.com/@dtr0105/ai-trust-and-safety-strategies-across-multiple-intervention-level-eb611245bd21?source=rss-13709fdd2256------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb611245bd21</guid>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[ai-policy]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[trusted-ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Tianrong Ding Terisa]]></dc:creator>
            <pubDate>Tue, 17 Feb 2026 14:55:35 GMT</pubDate>
            <atom:updated>2026-02-17T14:55:35.611Z</atom:updated>
            <content:encoded><![CDATA[<p>AI systems are now making decisions that affect millions of lives, from credit approvals to medical diagnoses to criminal sentencing. Yet a 2024 Mozilla Foundation study found that 57% of companies using AI had no formal incident response plan in place, despite more than half reporting at least one AI-related incident in the past year.</p><p>The challenge isn’t that organizations don’t care about AI trust and safety. It’s that most don’t know where to start or how to coordinate their efforts across different organizational layers. Technical teams build monitoring tools. Compliance teams write policies. Operations teams design workflows. But these efforts often happen in isolation, creating gaps that undermine the entire program.</p><p>This guide presents a practical framework for implementing AI trust and safety strategies across three intervention levels: system and tooling, process, and policy. Each level addresses different aspects of the challenge, and more importantly, they work together to create comprehensive protection.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HCebKO6CtgUFk3Jk" /></figure><h3>Understanding the challenge of AI trust and safety</h3><p>Before diving into solutions, let’s clarify what we’re actually trying to achieve. AI trust and safety encompasses two distinct but related domains.</p><p><strong>AI security</strong> protects systems from technical threats: adversarial attacks that manipulate model outputs, data poisoning that corrupts training sets, prompt injection attacks that bypass safety controls, and unauthorized access to model weights or training data. These are fundamentally cybersecurity challenges applied to AI systems.</p><p><strong>AI safety</strong> prevents unintentional harm: biased decisions that discriminate against protected groups, hallucinated information presented as fact, unpredictable behavior in novel situations, and misalignment between model objectives and human values. These are unique to machine learning systems that learn patterns from data rather than following explicit programming.</p><p>The distinction matters because the interventions differ. Security breaches typically have clear attack vectors and can be addressed through technical controls. Safety failures are often emergent properties of complex systems and require ongoing monitoring, human oversight, and governance frameworks.</p><p>Organizations need both. A secure system can still produce harmful outputs. A safe system can still be compromised by attackers. The three-level framework addresses both domains in an integrated way.</p><h3>Level 1: System and tooling interventions</h3><p>System-level interventions form the technical foundation of AI trust and safety. These are the automated controls, monitoring systems, and validation tools that operate continuously without human intervention.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Re7PqIYPuQplMZz_" /></figure><p><strong>AI trust validation</strong> tools continuously assess model performance, fairness, and robustness. These systems track metrics like demographic parity across different user groups, accuracy drift over time, and vulnerability to adversarial inputs. Platforms like Swept AI provide automated validation that runs continuously rather than as one-time certification exercises.</p><p><strong>Technical controls</strong> include encryption for data at rest and in transit, access management systems that enforce least-privilege principles, network segmentation that isolates AI infrastructure, and data loss prevention tools that block sensitive information from reaching external models. These are extensions of standard cybersecurity practices adapted for AI-specific risks.</p><p><strong>Red teaming and adversarial testing</strong> proactively probe systems for vulnerabilities before attackers do. This includes automated fuzzing that generates unexpected inputs, human red teams that attempt creative attacks, and specialized tools like Microsoft’s Counterfit framework for testing AI model robustness.</p><p><strong>Explainability and transparency tools</strong> help humans understand why AI systems make specific decisions. These range from simple feature importance scores to sophisticated techniques like LIME and SHAP that explain individual predictions. Transparency is both a safety requirement (catching problematic reasoning) and a regulatory necessity under frameworks like the EU AI Act.</p><p>When implementing system-level interventions, organizations should integrate with existing security infrastructure rather than building parallel systems. Most enterprises already have SIEM platforms, identity management systems, and compliance monitoring tools. AI-specific controls should feed into these established workflows.</p><p>UNDP’s <a href="https://www.undp.org/digital/ai">AI for Development</a> initiative emphasizes building these technical foundations with a human-centered approach, ensuring that monitoring and validation tools serve broader development goals rather than becoming ends in themselves.</p><h3>Level 2: Process interventions</h3><p>Process interventions define how humans interact with AI systems and respond when things go wrong. Even the most sophisticated technical controls cannot fully automate trust and safety. Human judgment remains essential.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U7tZ419EEnhVUwhs" /></figure><p><strong>Incident management workflows</strong> provide structured procedures for detecting, classifying, and responding to AI failures. A 2023 case study illustrates why this matters: a language model used by a U.S. public agency generated false legal interpretations in responses to citizen queries. Because the agency had clear incident response procedures, they quickly paused the system, notified the public, reviewed past interactions, and retrained the model on corrected data. Without those processes, the incident could have persisted for months.</p><p>Key components of AI incident management include:</p><ul><li><strong>Detection systems</strong> that monitor model outputs for anomalies, bias indicators, or unexpected behavior patterns</li><li><strong>Classification frameworks</strong> that assign severity levels based on potential harm and affected populations</li><li><strong>Notification protocols</strong> that alert the right stakeholders (technical teams, legal, communications) based on incident type</li><li><strong>Containment procedures</strong> for temporarily pausing systems or limiting access when risks are identified</li><li><strong>Investigation processes</strong> that identify root causes in data, model architecture, or deployment configuration</li><li><strong>Remediation workflows</strong> for updating models, correcting processes, or adding safeguards</li><li><strong>Documentation requirements</strong> that capture lessons learned and update policies accordingly</li></ul><p><strong>Human-in-the-loop (HITL) requirements</strong> ensure that AI systems never operate as fully autonomous decision-makers. Georgia’s government guidelines provide a clear standard: “AI tools should not be the only agents involved in any decision-making.” This means humans review both inputs to and outputs from AI systems, with clear accountability for final decisions.</p><p><strong>Continuous validation processes</strong> go beyond initial testing to monitor systems throughout their operational life. AI models degrade over time as the world changes. Data drift, concept drift, and emerging edge cases all require ongoing attention. Validation processes should include scheduled re-assessment, trigger-based reviews when metrics exceed thresholds, and periodic red teaming exercises.</p><p><strong>Cross-functional governance committees</strong> bring together technical, legal, ethical, and business stakeholders to make decisions that no single function can make alone. These committees typically set risk tolerance levels, approve high-risk use cases, review incident reports, and ensure that AI deployments align with organizational values.</p><p>UNDP’s <a href="https://www.undp.org/digital/standards">Digital Standards</a> provide frameworks for establishing these governance processes, with particular attention to inclusive decision-making that represents diverse stakeholder perspectives.</p><h3>Level 3: Policy interventions</h3><p>Policy interventions establish the governance framework within which technical and process interventions operate. Policies answer fundamental questions: What risks are acceptable? Who is accountable when things go wrong? How do we ensure compliance with regulations?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fJASBA52P_AoC3Hu" /></figure><p><strong>Responsible AI policies</strong> articulate organizational commitments to ethical AI development and deployment. Microsoft’s Responsible AI Standard, for example, is built on six core principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. These principles translate into specific requirements that product teams must meet before deployment.</p><p><strong>Risk assessment methodologies</strong> provide systematic approaches to evaluating AI systems before deployment. The EU AI Act’s four-tier risk classification (unacceptable, high, limited, minimal) offers one framework. NIST’s AI Risk Management Framework provides another. Effective risk assessment considers not just technical performance but potential impacts on individuals, communities, and society.</p><p><strong>Regulatory alignment</strong> ensures that organizational practices meet legal requirements. Key frameworks include:</p><ul><li><strong>EU AI Act</strong>: Risk-based regulation with strict requirements for high-risk systems</li><li><strong>NIST AI Risk Management Framework</strong>: Voluntary guidance widely adopted as best practice</li><li><strong>ISO/IEC 42001</strong>: International standard for AI management systems</li><li><strong>Sector-specific regulations</strong>: FDA guidance for medical AI, financial services regulations for algorithmic trading</li></ul><p><strong>Transparency and explainability requirements</strong> mandate that organizations document how AI systems work and communicate this to affected individuals. This includes model cards that summarize system capabilities and limitations, plain-language explanations of automated decisions, and public reporting on AI use and incident statistics.</p><p><strong>Accountability structures</strong> assign clear responsibility for AI system outcomes. This includes designated AI ethics officers, escalation paths for ethical concerns, and mechanisms for affected individuals to challenge automated decisions or seek redress.</p><p>Policy interventions often fail when they are developed in isolation from technical realities. Policies that ignore implementation constraints are ignored in practice. Effective policy development involves technical teams from the start and includes regular review cycles that incorporate lessons from operational experience.</p><p>UNDP’s <a href="https://www.undp.org/digital/transformations">Digital Transformation Framework</a> emphasizes the importance of aligning AI policies with broader digital governance strategies, ensuring that AI-specific interventions support rather than conflict with organizational transformation goals.</p><h3>Integrating the three intervention levels</h3><p>The three intervention levels are not independent options to choose from. They are interdependent components of a comprehensive trust and safety program. Integration failures create dangerous gaps.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vMLHalVYvx60UgBn" /></figure><p><strong>System controls enforce policy requirements.</strong> If your policy prohibits using certain data types for training, technical controls should block that data from reaching training pipelines. If your policy requires fairness testing, validation tools should automatically run those tests and block deployment when thresholds aren’t met.</p><p><strong>Processes trigger technical safeguards.</strong> Incident response workflows should automatically activate containment controls when certain thresholds are crossed. Human review processes should feed data back into training sets to improve model performance.</p><p><strong>Policies inform system design.</strong> Security architecture decisions should reflect policy risk tolerance. Monitoring metrics should align with policy-defined fairness criteria. Access controls should implement policy-defined data governance rules.</p><p>Common integration pitfalls include:</p><ul><li><strong>Over-reliance on technical controls</strong> without human oversight for edge cases and novel situations</li><li><strong>Policies that ignore technical feasibility</strong>, creating requirements that teams cannot practically implement</li><li><strong>Processes that bypass automated safeguards</strong> in the name of efficiency, undermining security investments</li><li><strong>Siloed teams</strong> that optimize their level without understanding impacts on other levels</li></ul><p>Effective integration requires cross-functional governance structures with authority to make trade-offs across technical, operational, and policy domains. It also requires shared metrics that measure program effectiveness holistically rather than optimizing individual levels in isolation.</p><p>UNDP’s <a href="https://www.undp.org/digital/ai-trust-and-safety-re-imagination-programme-building-frameworks-future">AI Trust and Safety Re-imagination Programme</a> provides resources for organizations developing integrated approaches, with particular attention to contexts where technical infrastructure may be limited but governance innovation can compensate.</p><h3>Implementation roadmap for organizations</h3><p>Building a comprehensive AI trust and safety program doesn’t happen overnight. Here’s a phased approach that organizations can adapt to their specific context.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZdK92b0_RbARUCX9" /></figure><p><strong>Phase 1: Assessment (Months 1–2)</strong></p><p>Start with a clear-eyed assessment of your current state. Map existing AI systems across the organization. Identify which systems make high-stakes decisions affecting individuals. Catalog existing technical controls, incident response capabilities, and policy frameworks. This isn’t about achieving perfection; it’s about understanding where you are so you can prioritize effectively.</p><p>Risk identification should focus on systems that affect vulnerable populations, make decisions that are difficult to reverse, or operate in regulated domains. Not all AI systems require the same level of trust and safety investment.</p><p>Stakeholder mapping identifies who needs to be involved in governance decisions. This typically includes technical teams, legal and compliance, business units using AI, and representatives of affected communities.</p><p><strong>Phase 2: Foundation (Months 3–6)</strong></p><p>Deploy core technical controls for high-risk systems first. This includes basic monitoring, access controls, and input validation. Don’t try to build everything at once; start with controls that address your highest-priority risks.</p><p>Establish incident response workflows with clear ownership and escalation paths. Run tabletop exercises to test these workflows before you need them in a real crisis.</p><p>Draft initial policies that articulate principles and high-level requirements. These will evolve, but having a starting point enables consistent decision-making.</p><p><strong>Phase 3: Integration (Months 7–12)</strong></p><p>Connect the levels through automation. Technical controls should feed data to incident response systems. Incident reports should trigger policy reviews. Policy changes should drive technical roadmap priorities.</p><p>Implement continuous monitoring that tracks metrics across all three levels. This includes technical performance, process adherence, and policy compliance.</p><p>Conduct your first full-cycle review, assessing what’s working, what’s not, and where to invest next. AI trust and safety is not a project with an end date; it’s an ongoing operational commitment.</p><p>Maturity indicators help track progress. Early-stage programs are reactive, responding to incidents after they occur. Mature programs are proactive, identifying and mitigating risks before they manifest. The most advanced programs are predictive, using data to anticipate emerging risks and adapt defenses accordingly.</p><p>UNDP’s <a href="https://www.undp.org/digital/dra">Digital Readiness Assessment</a> provides diagnostic tools that organizations can use to evaluate their current capabilities and identify priority areas for investment.</p><h3>Building responsible AI for sustainable development</h3><p>AI trust and safety isn’t just about risk mitigation. It’s about building systems that genuinely serve human needs and contribute to sustainable development goals.</p><p>The three-level framework provides a structure, but the specific implementation must be adapted to context. A financial services firm in a major market faces different challenges than a government agency in a developing country. A healthcare AI system requires different safeguards than a content recommendation engine.</p><p>What remains constant is the need for coordination across technical, process, and policy domains. Organizations that invest in only one level while neglecting others will find their trust and safety programs undermined by gaps they didn’t anticipate.</p><p>The investment is worth it. Organizations with mature AI trust and safety programs can deploy AI more confidently, respond to incidents more effectively, and build trust with users, regulators, and the public. In an environment where AI incidents can destroy reputation and value overnight, comprehensive trust and safety isn’t a cost center. It’s a competitive advantage.</p><p>UNDP’s <a href="https://www.undp.org/digital/ai-trust-and-safety">AI Trust and Safety</a> resources offer guidance for organizations at all stages of this journey, with particular attention to contexts where responsible AI can accelerate progress toward the Sustainable Development Goals. The <a href="https://www.undp.org/digital/ai">AI for Development</a> initiative provides additional resources for building AI systems that serve the needs of the Global Majority.</p><p>The frameworks and strategies outlined here are starting points, not endpoints. The field of AI trust and safety is evolving rapidly as both the technology and our understanding of its impacts mature. Organizations that build adaptive, integrated programs today will be best positioned to navigate the challenges and opportunities ahead.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb611245bd21" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>