Artificial Intelligence, Child Safety, and Digital Governance: Evaluating Policy Effectiveness in Preventing Online Trafficking
This paper was presented at the Silicon Valley Peace Conference 2026 in Palo Alto, California, United States, held September 17–20, 2026, under the theme “AI’s Role in Community Development and Peacebuilding in Developing Nations,” and is published as part of the Proceedings of the Silicon Valley Peace Conference 2026.
Abstract
This research analyzes national, regional, and global artificial intelligence (AI) policy frameworks that govern the use of AI to prevent online trafficking and exploitation of children and youth, with a focus on India and similar developing country contexts. It addresses three central questions: (1) In what ways do existing AI safety regulations and national AI strategies mitigate or inadvertently exacerbate trafficking risks in digital environments? (2) What comparative gaps exist in policy design, implementation, and institutional coordination across jurisdictions? (3) How can participatory and polycentric policy models enhance equitable AI deployment while supporting peacebuilding and community resilience? The study is grounded in critical political economy and feminist governance frameworks, which analyze the intersections of digital infrastructures with structural inequalities, linguistic marginalization, and uneven state capacity. The study employs a comparative policy analysis of major regulatory and governance instruments, including India’s Digital Personal Data Protection Act (2023), the European Union’s AI Act (2024), and United Nations guidelines on child online protection and ethical AI. The analysis integrates both qualitative and quantitative methods. Legislative and policy documents are systematically coded to evaluate risk classification systems, regulatory safeguards, accountability mechanisms, and provisions for equitable access. These results are triangulated with empirical implementation data from national crime statistics, international law enforcement datasets, and documented AI-assisted anti-trafficking interventions. Case studies examine India’s use of AI-enabled tools such as Garuda Drishti for detecting trafficking- linked digital advertisements, as well as AI-supported child protection interventions by the National Commission for Protection of Child Rights, which facilitated the resolution of over 26,000 complaints and the rescue of more than 2,300 children. Global applications, such as Thorn’s Spotlight tool, have enabled the identification of over 26,000 victims through predictive analytics and natural language processing. Regional assessments across South Asia reveal fragmented yet emerging AI-enabled coordination among law enforcement and civil society actors. Preliminary findings suggest that although AI introduces new risks, such as algorithmic targeting, deepfake-enabled coercion, and exploitation of low-literacy and marginalised populations, it also substantially improves detection, investigative efficiency, and victim identification when implemented within accountable governance frameworks. However, existing policy regimes are largely shaped by high-income regulatory contexts and insufficiently address Global South priorities, including multilingual inclusion, rural connectivity, and institutional capacity limitations. The principal contribution of this study is the development of a novel evaluation model for AI safety regulations in anti-trafficking contexts. Informed by polycentric governance principles, the model integrates multi-level institutional coordination and participatory metrics to assess policy effectiveness beyond formal compliance. It introduces indicators for detection accuracy across linguistic groups, mitigation of false positive risks, equity in deployment, and stakeholder participation, including youth and survivor engagement. The model also advances hybrid public and private governance mechanisms, supporting coordinated ethical audits, accountability safeguards, and localized policy adaptation. By linking AI governance effectiveness to child protection, social stability, and community resilience, this study offers a scalable evaluation model to inform national AI strategies and global regulatory cooperation. It proposes policy pathways for equitable, accountable, and peace-oriented AI deployment in developing nations.
Keywords: child safety, digital governance, online trafficking, AI deployment
Introduction
Human trafficking represents the worst form of crime against vulnerable and marginalised populations, internationally recognised as modern-day slavery. Victims endure abuses violating basic tenets of human rights and dignity. However, trafficking does not occur in vacuity. It is cradled in a web of legitimate and illegitimate industries. Innovation within these industries supports traffickers to expand operations with impunity while evading law enforcement detection (Manoraj, 2021).
Human interaction and communication has been transformed by the technological push in the last decade. Technology now gives us the ability to reach millions of people, get information and change reality online. This has changed our behaviour and interaction with external stimuli. This is the case with artificial intelligence (AI), which is transforming the war against human trafficking.
AI brings new vectors of harm as it allows traffickers to weaponise technology and amplify their operations. Traffickers use AI to identify and target victims, gather intelligence about vulnerabilities, communicate with large groups of potential victims through culturally and linguistically tailored messages, and develop customized exploitation strategies while avoiding detection (The Impact of Artificial Intelligence on Trafficking in Persons, 2026).
At the same time, AI is transforming the way states, civil society and digital platforms detect, investigate and respond to online harms. AI systems scour troves of social media content for suspicious ads, and flag reports of child exploitation for human review. These capabilities have the potential to allow for faster detection and more targeted interventions, especially in low-resource investigative environments.
This paper explores national, regional and global policy frameworks that regulate the use of AI in preventing online trafficking and exploitation of children and youth, with a focus on India and other developing countries. The analysis answers three main questions: (1) how do existing AI safety regulations and national AI strategies mitigate or exacerbate trafficking risks in digital environments? (2) What are the gaps in policy design, implementation, and institutional coordination across jurisdictions? (3) How can participatory and polycentric policy models support equitable AI deployment, while also supporting community resilience? This study uses critical political economy and feminist governance frameworks to explore the interaction between digital infrastructures and structural inequalities, linguistic marginalisation and uneven state capacity.
The paper describes the conceptual and theoretical framework that underpins the analysis followed by an overview of the context of human trafficking, online trafficking and AI. It then charts the policy landscape of key regulatory instruments such as India’s Digital Personal Data Protection Act (2023), the European Union’s AI Act (2024) and United Nations guidelines on child online protection and ethical AI. Case studies are presented to explore AI-enabled tools used to combat human trafficking in India and South Asia.
Preliminary results indicate that although AI introduces additional risks, including algorithmic targeting, coercion through deepfakes, and economic exploitation of low-literacy and marginalized populations, it also has the potential to greatly improve detection, investigative efficiency, and victim identification when used within accountable frameworks of governance. However, current policy environments are dominated by regulatory settings in high-income contexts and do not sufficiently take into account priorities relevant to the Global South such as multilingual inclusion, rural connectivity and capacity constraints of institutions.
The primary contribution of this paper is the development of a evaluation model for AI safety regulations in anti-trafficking contexts. The model follows principles of polycentric governance, such as multi-level institutional coordination and participatory metrics, to evaluate policy performance beyond formal compliance. It suggests indicators for detection accuracy by linguistic group, risk mitigation of false positives, equity of deployment and stakeholder participation, including engagement with youth and survivors. The model further promotes hybrid public-private governance mechanisms that allow for coordinated ethical audits, accountability safeguards and adaptation of policies to local contexts.
The paper links the effectiveness of AI governance to child protection, social stability, and community resilience. It provides a scalable evaluation model that can inform national AI strategies and global regulatory cooperation. It proposes policy pathways to advance equitable, accountable and peace- promoting AI deployment in developing countries.
Conceptual and Theoretical Framework
Digital infrastructures structure power as much as they enable services. The digital systems being built today are shaping how countries govern, how they execute and deliver services (PIB, 2026). The result is a political economy of AI in which platform architectures and data regimes materially shape who is visible, whose harms are legible, and what interventions are possible (UNESCO, 2021). Models trained largely on major-language corpora reinforce linguistic exclusion. Urbanized, centralized data sources are the bedrock for detection pipelines that render rural experiences invisible. Analytics designed for high-signal, high-bandwidth environments break down in environments where connectivity is intermittent. These failures are not technical footnotes; they are social determinants of who is protected and who is exposed (Nugraha & Shinta, 2024).
Critical political economy highlights the ways datafied governance interacts with poverty, caste, gender and spatial marginality. It looks into profound links between politics and economics that affect the world regarding resources and power (Scherrer et al., 2023; Sustainability Directory, 2025). The framework examines the interaction between the market and the government that influences society, environment and sustainability (Sustainability Directory, 2025). The objective is to move beyond simplistic economic models and to confront the reality of power dynamics and political choices that have a significant impact on economic outcomes (Scherrer et al., 2023; Sustainability Directory, 2025).
Trafficking and online exploitation are part of hierarchies that predate the internet. Training data and institutional deployment decisions transfer social bias to algorithmic risk scores, content-prioritisation heuristics, and automated moderation pipelines (Gorwa et al., 2020). Where marginalised castes, ethnic minorities and impoverished communities are closely surveilled or have limited legal remedy, automated flags may simply stigmatise rather than deliver protective outcomes. In contrast, where governance is rooted in local knowledge – community reporting channels, survivor networks, multilingual moderation – AI systems enhance existing capacities.
Private platforms and tech companies are critical nodes in the shaping of risk and remedy (Rahman et al., 2024; Suzor, 2019; Zuboff, 2019). They control evidence generation and sharing pipelines. They decide content visibility and algorithmic demotion (Gillespie, 2018; Noble, 2018; Roberts, 2019). These decisions are informed by business models, legal obligations and technological constraints (Rahman et al., 2024; Suzor, 2019; Zuboff, 2019). Law enforcement, NGOs and private firms often collaborate in anti-trafficking work, facilitating rapid data sharing and investigative leads, but raising questions of consent, transparency and accountability when detection logics are driven by private priorities (L’Hoiry et al., 2024; Latonero, 2011). Where multinational firms implement a standardised content policy across divergent jurisdictions, the asymmetry is stark (Rahman et al., 2024; Suzor, 2019; Zuboff, 6 2019). Moderation policies calibrated for metropolitan English-speaking users fail for regionally nuanced content, resulting in missed detections and algorithmic misclassification (L’Hoiry et al., 2024; Noble, 2018; Suzor, 2019). The policy terrain therefore, should be co-produced by state regulation and external actors such as NGOs and survivor groups (L’Hoiry et al., 2024; OSCE, 2020, 2023; Rahman et al., 2024).
Feminist governance understands protection as political and contested, not technical or benevolent (D’Ignazio & Klein, 2020; Sawer et al., 2023). It reshapes political, organizational and institutional structures, through the integration of feminist values: inclusivity, intersectionality, shared leadership (Sawer et al., 2023). Feminist scholars critique security narratives that emphasize detection over autonomy and long-term well-being, highlighting how “protection” obscures surveillance practices that reproduce gendered control and paternalism (D’Ignazio & Klein, 2020; UNICEF Innocenti, 2025).
Children and youth are not passive objects of rescue protocols; they are political subjects with rights, agency and situated knowledge (UNICEF Innocenti, 2026). A feminist-informed governance position requires survivor-centredness, informed consent where possible, and reparative practices beyond incident removal to rehabilitation and community reintegration (Hester et al., 2025). Gender- and age- specific harms matter because exploitation often depends on social norms (Reeves et al., 2018). When AI systems regard victims as data points rather than as subjects with rights, interventions risk creating secondary harm – stigmatisation, retraumatisation or punitive entanglement (UNICEF, 2025; Vanberghen, 2024).
Polycentric governance is supportive of transnational, cross-sectoral challenges as an alternative to monolithic regulatory solutions (Stephan et al., 2019). Several independent decision-making centers operate independently but are aligned within common broader frameworks (Stephan et al., 2019). Polycentric architecture distributes authority amongst national regulators, sectoral agencies, law enforcement, platforms, NGOs, and survivor collectives. Polycentric architecture builds in redundancies that decrease single point failures and allows local knowledge to calibrate interventions (Thiel, 2023). These systems create co-regulatory spaces and negotiated protocols for data sharing, redress and oversight (Ostrom, 2009; Stephan et al., 2019). Polycentricity offers adapted responses to cross-border trafficking that take account of linguistic and cultural differences, yet still enable international investigative cooperation (Ostrom, 2009; Thiel, 2023).
The Global South critique questions which harms are given priority for attention, whose languages are supported to develop models, and whether design considerations include low-bandwidth, offline modes of detection (Arora, 2024; Couldry & Mejias, 2020; Mhlambi & Tiribelli, 2023). It asks for a reorientation of risk frameworks to incorporate linguistic diversity, infrastructural precarity and variation in the capacity of states to audit and enforce rules (Okola et al., 2022; UNESCO, 2021). Safety- by-design must be joined by equity-by-design: procedural equity in participation, distributive equity in resource allocation, and epistemic equity in whose knowledge informs algorithmic models (Arora, 2024; Aneja & Shrivastava, 2026; UNESCO, 2021).
The theoretical frame considers AI systems as embedded interventions whose effects depend on data ecologies, power asymmetries, institutional forms and inclusion of those most affected.
Methodology
This study employs comparative policy analysis of major regulatory instruments, including India’s Digital Personal Data Protection Act (2023), the European Union’s AI Act (2024), and United Nations guidelines on child online protection and ethical AI. Legislative and policy documents undergo systematic coding to evaluate risk classification systems, regulatory safeguards, accountability mechanisms, and provisions for equitable access.
The selection of instruments is logical and purposive. There are two reasons for the main locus being India’s DPDP Act and associated national AI and child-protection strategies. India is a case in point of concentrated challenges (large platform user bases, high levels of linguistic diversity, heterogeneous levels of connectivity, multi-tiered administrative structures that shape implementation). The EU AI Act is a high-capacity regulatory benchmark, with its risk-classification architecture and procedural safeguards providing a contrasting model for assessing transferability and fit. International instruments such as the UN guidelines on child online protection, UNICEF guidance on AI and children’s rights and UNESCO’s recommendations on the ethics of AI set global normative baselines that national frameworks purport to align with or deviate from. The choice of these instruments allows for analytic triangulation across levels of national law, super-national regulation and trans-national norms, which is necessary to investigate domestic policy fidelity and cross-border coordination in anti- trafficking responses.
The sources of data are primary legal texts, administrative records and practice-oriented documentation. Policy texts include the DPDP Act, AI strategy documents, the EU AI Act and its delegated instruments, and UN multilateral guidance. Implementation datasets include national crime records and law enforcement summaries. Practice based evidence is derived from operational documentation and impact reports of AI assisted interventions such as Garuda Drishti & Say Help app. Supplementary material includes reports from NGOs, investigative journalism on deployments and results, and peer-reviewed or grey-literature assessments of pilot projects. In cases where official administrative data are still restricted, triangulation draws on vetted secondary sources.
Coding is developed along six analytical dimensions, in successive iterative phases: risk classification, prescribed safeguards, accountability mechanisms, equity provisions, participatory processes and institutional coordination. Comparative Matrices examine jurisdictional variations. The case-study synthesis connects the coded policy outputs and implementation metrics, evaluating the alignment of normative prescriptions with observable outcomes. The administrative crime data do not sufficiently reflect the scale of trafficking and online exploitation. The study is cautious about absolute numbers. The interpretive approach interprets power asymmetries and voices of affected communities through lenses of critical political economy and feminist governance.
Literature Review and Mapping the Policy Landscape
Human Trafficking & Online Trafficking
Trafficking means the actions and the processes undertaken to illegally transport people for exploitation. Article 3 of the Palermo Convention defines human trafficking as: “(a) Trafficking in persons shall mean the recruitment, transportation, transfer, harbouring or receipt of persons, by means of the threat or use of force or other forms of coercion, of abduction, of fraud, of deception, of the abuse of power or of a position of vulnerability or of the giving or receiving of payments or benefits to achieve the consent of a person having control over another person, for the purpose of exploitation. Exploitation shall include, at a minimum, the exploitation of the prostitution of others or other forms of sexual exploitation, forced labour or services, slavery or practices similar to slavery, servitude or the removal of organs; […]”(United Nations Convention against Transnational Organized Crime, 2000). Therefore, the three key elements that must be present for a situation of trafficking in persons (adults) to exist are: (i) the Action (what is done); (ii) through means of (how it is done); and (iii) goals (why it is done) (United Nation, 2019).
The global scale and persistence of the crime is further illustrated by the 2024 UNODC Global Report on Trafficking in Persons, which reports that Member States identified 202,478 trafficked persons between 2020 and 2023 (UNODC, 2024). The number of victims detected was 25 per cent higher in 2022 than in 2019, a post-pandemic rebound in detection as vulnerabilities linked to poverty, conflict and climate-induced displacement deepened. Now, children represent 38% of the victims detected. This is a dramatic shift in the age profile of trafficking and raises serious concerns about the weakening of child protection systems (UNODC, 2024).
Disaggregating the data by age and gender reveals very gendered patterns of victimization. Most victims detected are women and girls, primarily in sexual exploitation, whereas boys and men are more likely to be victims of forced labour (UNODC, 2024; Rami et. al., 2025). Forty-two per cent of victims were trafficked for forced labour and 36 per cent for sexual exploitation, confirming that forced labour is at least co-equal relative to sexual exploitation in several regions (UNODC, 2024). Trafficking for forced criminality, like coercion into online scams, has moved from the margins to the mainstream (UNODC, 2024).
Approximately 70% of convicted traffickers are men, 28% are women and 74% are linked to organised crime structures (UNODC, 2024). In South Asia the picture is different, with 55% of identified victims trafficked for forced labour and 30% for sexual exploitation (UNODC, 2024). The victims are 44% women, 15% men, 24% boys and 17% girls (UNODC, 2024). The region is primarily a destination for domestic and intra-regional trafficking, but victims can be found in 36 countries worldwide (UNODC, 2024). The number of countries where data is collected has increased from 26 in 2009 to 129 in 2024, but trafficking is still grossly underreported due to the clandestine nature of the crime (UNODC, 2024; UNODC, 2018; Manoraj, 2021).
The existing challenges in combatting human trafficking are augmented with technology, where the traffickers now possess the ability to locate, recruit, coerce and control their victims online. Online trafficking or technology facilitated trafficking is a new tool used by traffickers in each stage of the trafficking process (Manoraj, 2021). Mark Latonero in 2012 defined technology-facilitated trafficking
Figure 1
Elements of Human Trafficking (Greiman & Bain, 2012)
as “[…] refers to the social and technical ecosystem wherein individuals use information and communication technologies to engage in human trafficking and related behaviours. Digital and networked technologies impact visibility, coordination, transaction, exchange, and organization. These technologies, therefore, can impact various aspects of trafficking, from grooming, recruitment, and control of victims, to advertising, movement, and financial transactions.” (Latonero, 2012). Whereas, International Centre for Missing & Exploited Children in its report “'Studies In Child Protection: Technology-Facilitated Child Sex Trafficking” has defined “technology-facilitated child sex trafficking” as one which “denotes that the Internet and related technologies may enable/assist with the trafficking of children for sexual purposes. These technologies may include ICTs such as hardware and software devices and programs (e.g., personal computers, scanners, digital cameras, multimedia programs, and communications equipment); email, video conferencing, wireless networks, and mobile services; communication networks (e.g., newsgroups, web message and bulletin boards, chatrooms, File Transfer Protocol (FTP), Peer-to-Peer (P2P) networks); encryption; digital currencies, and other tools.” (International Centre for Missing & Exploited Children, 2018).
Technology now touches every link in the human trafficking chain, from recruitment to payment, vastly extending the reach and control of traffickers. Perpetrators use online platforms, messaging apps and social media to identify, groom and recruit (or procure) victims, whose services are then sold to buyers in increasingly anonymised digital marketplaces. When recruited, victims are subjected to intensive control and compulsion. They are forced to perform labour or sexual acts, kept on call, and prevented from reporting their exploitation to law enforcement, NGOs or family members through the use of physical and psychological abuse, social isolation, debt bondage, substance dependency, coercion, fraud and the withholding of identity documents. Technology also mediates advertising and customer acquisition, with online platforms hosting ads, reviews, chats and other communications – especially for sexual services – that allow traffickers to attract and engage with buyers at scale. Exploitative practices include sexual services, forced labour, forced begging, coerced criminal activity and organ removal – all of which can be coordinated, monitored and optimised using digital tools. These activities are supported by financial transactions – payments, transfers and laundering – that increasingly take place through online banking, e-wallets and other digital payment systems, which make money flows faster and harder to trace across jurisdictions (UNODC, 2019). Moreover, the platforms for online communication have turned into virtual venues for the spread of online sexual abuse material (OSCE, 2020). The use of technology in trafficking has created new business opportunities like live streaming of sexual acts to large audiences and the digitalization of large-scale pornography of trafficked victims.
Previously, technologies such as self-destructing messaging apps, secondary or anonymous profiles (e.g. finstagrams), end-to-end encrypted channels, location-based services, and face-altering filters provided traffickers with powerful tools to recruit, groom and conceal victims. Offenders are now building on these affordances with AI techniques such as deepfakes, sentiment analysis and behavioural profiling to target and manipulate individuals at scale (Manoraj, 2021). Generative AI also reduces the barriers to creating synthetic sexual abuse material, non-consensual intimate imagery and hyper-realistic personas that can be used across platforms without the original recorded abuse, making detection, evidentiary standards and victim identification all the more difficult. The transition from relatively simple obfuscation tools to sophisticated AI-driven systems is altogether a critical inflection point in technology-facilitated trafficking.
Artificial Intelligence
Artificial intelligence (AI) denotes a family of computational techniques—especially machine learning and deep neural networks—that learn statistical patterns from large-scale data to perform tasks such as perception, prediction, and decision-making with limited explicit programming (UNESCO, 2021). Its economic footprint is no longer speculative: estimates place the global AI market at about USD 198 billion in 2023, with projections of more than USD 1.8 trillion by 2030, implying compound annual growth rates upward of 30% in some segments (FAIST, 2024; Nova One Advisor, 2025). Beneath these revenue curves lies an equally dramatic expansion of computational infrastructure; recent analyses suggest that aggregate AI computing capacity from specialized chips has been growing by roughly 3.3 times per year since 2022, corresponding to a doubling time of about seven months (You et. al., 2026).
Such scaling accelerates scientific discovery and productivity, yet it also amplifies concerns about labour displacement, concentration of technological power, and cross-border inequality, as advanced economies and a handful of firms dominate frontier model development (UNCTAD, 2025).
Normatively, AI is increasingly framed as a site of contestation over human rights, environmental sustainability, and epistemic justice. UNESCO’s Recommendation on the Ethics of Artificial Intelligence, adopted by 194 member states, attempts to anchor AI governance in principles of human dignity, transparency, and accountability while acknowledging the structural risks of bias, surveillance, and exclusion (UNESCO, 2021). AI thus crystallizes a broader political question: who designs, owns, and audits these systems, and under what institutional arrangements their benefits and harms are distributed (UNESCO, 2021).
AI in India
India's AI ecosystem has moved from experimentation to institution-building with unusual speed. The IndiaAI Mission, approved with approximately ₹10,300 crore over five years, anchors this transition by funding a national GPU grid, domestic semiconductor capacity, and open data platforms designed to democratise access to compute and datasets (Ministry of Electronics and Information Technology, 2025; PIB, 2025). A common facility of about 18,693 GPUs, offered at subsidised rates well below global benchmarks, now underwrites model training for universities, startups, and public agencies and remains slated for further expansion (Ministry of Electronics and Information Technology, 2025; Times of India, 2026). This infrastructure agenda deliberately couples industrial policy with the Prime Minister's rhetoric of "technology for all," positioning AI as a general-purpose capability rather than a niche export industry (PIB, 2025).
Within this architecture, the IndiaAI Safety Institute (AISI) functions as the system's normative core. Set up under the "Safe and Trusted AI" pillar, AISI operates as a hub-and-spoke network linking IITs, IIITs, NITs, defence universities, and civil-society partners to produce India-specific tools for watermarking, deepfake detection, machine unlearning, synthetic data generation, bias mitigation, and stress-testing of high-risk systems (IndiaAI, 2025). The mandate extends beyond technical research. AISI advises regulators, articulates risk taxonomies, and contextualises international standards for India's multilingual, unequal, and data-scarce environments (IndiaAI, 2025; Data Security Council of India, 2025). It aims to embed safety-by-design and equity-by-design into the AI value chain rather than bolt them on as ex post compliance requirements.
The institute has started eight “Responsible AI” projects that lay out an emerging governance toolkit for the whole lifecycle of model development, deployment, and oversight. IIT Jodhpur is at the forefront of research in machine unlearning in generative foundation models, developing ways to selectively delete training data so that models can “forget” specific individuals without having to be retrained from scratch (PIB, 2024a). IIT Roorkee develops synthetic data pipelines that generate realistic yet privacy-preserving datasets to mitigate bias in data distributions and downstream machine-learning pipelines (Mahajan, 2025). NIT Raipur introduces this bias-mitigation agenda to applied health-care settings where skewed datasets translate into unequal care along caste, gender, and regional lines (PIB, 2024a; Mahajan, 2025). DIAT Pune develops explainable and privacy- preserving AI frameworks for security applications; a consortium of IIT Delhi and IIIT Delhi tests robust privacy-enhancing machine-learning models with cryptographic protection (Mahajan, 2025). “Nishpaksh”, an ethical certification toolkit that makes non-discrimination operational into auditable metrics, is developed by IIIT Delhi (Mahajan, 2025). Track-LLM of Amrita Vishwa Vidyapeetham focuses on large language models, with a focus on transparency, risk assessment, and context-aware evaluation (PIB, 2024a). Taken together, these two projects are an ambitious experiment in building an indigenous safety stack that combines formal guarantees with participatory oversight.
The legal and regulatory scaffold around AI remains dense. Through the IT Act and 2021 Intermediary Rules, the government obliges platforms to remove unlawful content within tight timelines (PIB, 2024). The Digital Personal Data Protection Act 2023 and 2025 Rules introduce heightened safeguards for children, including verifiable parental consent and bans on behavioural tracking (PolicyEdge, 2026; PIB, 2026). India's AI Governance Guidelines recommend risk assessment frameworks and sector- specific obligations (Data Security Council of India, 2025).
India now ranks first globally in AI skill penetration, with AI-related competencies roughly 2.8 times the world average and talent concentration growing more than 260% since 2016 (Jeevanandam, 2024; Jeevanandam, 2024a). Indigenous models such as BharatGen and Sarvam-M exemplify an emerging sovereign AI stack encoding linguistic diversity into model design (PIB, 2025a). India's AI trajectory fuses ambitious state-led initiatives with expanding talent and insistence on multilingual, public- interest applications.
Global Governance on AI
The EU’s Artificial Intelligence Act (herein after “EU AI Act, 2024”) provides an architecture linking risk classification with compliance obligations in a coercive regulatory posture. It sets out the “high-risk” AI systems and creates inflexible requirements—data governance, documentation, human oversight, robustness and accuracy standards, and conformity assessments—all of which are directly relevant to systems used in the detection of trafficking and child protection (European Commission, 2024). The Act’s requirements for fundamental rights impact assessments and transparency are intended to tackle the harms that can be caused by algorithmic surveillance. Regarding anti-trafficking tools, the EU approach would require deployers to demonstrate accuracy, explainability and human-in-the-loop safeguards prior to operationalization. The EU framework’s strengths are its procedural depth and the institutional capacity of member states to carry out conformity assessments and sanctions. But this prescriptive approach assumes regulatory resources, administrative bandwidth, and standardized data ecosystems that many developing contexts do not have. As such, the AI Act is a clear example of challenges of policy transfer – it provides a strong framework of expectations, but well-defined safeguards do not always align with operational needs in low-capacity settings (European Commission, 2024; OECD, 2024).
The UN and multilateral guidance operate differently, providing normative guardrails rather than coercive standards. The UNESCO Recommendation on the Ethics of Artificial Intelligence and UNICEF’s policy guidance on AI and children emphasize rights-based design, participatory governance, and safeguards that are tailored to the developmental needs of children (UNESCO, 2021; UNICEF Innocenti, 2025). The documents highlight ethical principles such as non-discrimination, privacy and transparency, and urge States to conduct child-centred impact assessments and promote the meaningful participation of young people and caregivers. At the UN level, AI is framed in human rights terms, with an emphasis on international cooperation to protect children from online harms. But these norms are often high-level and non-binding and often fail to address operational details specific to trafficking detection, such as evidentiary standards for AI-flagged leads, protocols for cross- jurisdictional data sharing, or measures to prevent algorithmic re-traumatization. The normative scaffolding is useful, but it needs to be translated into operational guidance that is sensitive to context, particularly in countries where institutional enforcement and technical capacity are heterogeneous.
Some of these gaps are filled by international policing and sectoral bodies through trafficking-specific operational guidance. INTERPOL’s transnational efforts and projects to break the cycle of harm are centered around intelligence sharing, capacity building for investigative units and tools that can surface cross-border networks (INTERPOL, 2025). Operational work shows the value of algorithm leads and pattern analysis in disrupting trafficking networks. But they also reveal enduring tensions: the need for actionable intelligence by law enforcement can be at odds with privacy-preserving demands, and a lack of trust between agencies often hinders timely data sharing. International cooperation, strengthened by standardised mutual legal assistance protocols, interoperable data formats and joint training, is more effective in translating AI-assisted leads into prosecutions and victim support (EU- Lisa & Eurojust, 2022). Where collaboration breaks down, AI outputs are siloed and limited in scope.
The regional dynamics in South Asia is a mixed and changing scenario. India’s neighbors, Bangladesh, Nepal, Pakistan, Sri Lanka and others, have heterogeneous policy trajectories (Tech Global Institute, 2023; Mohanty, 2026). Some states have adopted nascent digital safety programs and cybercrime strategies; others are focusing on legislative oversight of platforms and flows of data without commensurate investment in children’s safety or regulation specific to artificial intelligence. Cross- border coordination is inconsistent, featuring ad hoc collaborations, information exchanges facilitated by NGOs, regional workshops, and limited connectivity between investigators and advocates. However, systematic regional architectures for harmonised AI governance in child protection are not yet developed. Language diversity, migratory labour corridors and informal recruitment practices make the region particularly susceptible to the dynamics of transnational trafficking. Political, legal and infrastructural barriers impede such integration, but cohesive regional mechanisms, such as standard data-sharing agreements, mutual capacity-building programmes and interoperable investigative platforms, would significantly improve responses (Mohanty, 2026).
The policy gaps analysis shows recurring themes in the comparative table on India, transnational and global AI policy documents (see table 1). India has statutory scaffolding and a proactive innovation posture but operational challenges in translating law into equitable, localized practice exist. The EU provides a coherent, risk-based blueprint with enforceable obligations, but assumes institutional capacities not always in place. UN guidance provides normative guidance but needs to be operationalised for trafficking specific use cases. South Asia’s emergent policies are a patchwork, highlighting the need for regional coordination and for evaluation frameworks that acknowledge infrastructural, linguistic, and institutional diversity. Successful AI governance in anti-trafficking will require hybrid methods: enforceable standards where feasible; normative guidance when needed; and polycentric locally adapted mechanisms for participation, oversight and regional cooperation (UNICEF, 2025; INTERPOL, 2025).
AI in Human Trafficking
Artificial intelligence has become central to contemporary anti-trafficking practice, transforming investigative methods for locating victims, tracing financial flows, and analyzing the social ecologies underlying exploitation. Simultaneously, traffickers leverage similar capabilities to automate grooming, conceal illicit proceeds, and industrialize abuse, often surpassing the effectiveness of traditional regulatory tools. Accordingly, a comprehensive analysis of AI and human trafficking must treat the technology as a complex and contested infrastructure, rather than as a neutral instrument.
Figure 2
Use of AI in Combatting Human Trafficking (Bagnall, 2025)
Figure 3
Comparison of AI Governance and Policy (Source: Author's Own Development)
Traffickers' Appropriation of AI and Generative Technologies
The same affordances which empower investigators, lower exploitation costs. OSCE's "New Frontiers" brief warns that criminal organisations are adopting generative AI to reduce costs or decrease detection probability (Bennett et al., 2024). Generative chatbots produce convincing recruitment scripts tuned to local idioms and gender norms. Deepfake tools lower thresholds for producing deceptive pornography, weakening victims' ability to withdraw consent (Bennett et al., 2024).
OSCE highlights scam compounds in Southeast and South Asia where workers are deceived into online fraud operations; AI systems optimise these by automating customer interactions (Bennett et al., 2024). In India, traffickers exploit encrypted messaging, AI-assisted editing, and automated advertising to identify vulnerable individuals (Tiwari, 2025). As generative models become accessible, marginal costs of producing synthetic abuse material approach zero while detection remains intensive.
The asymmetry is stark for children. OSCE and UNICEF warn generative AI can fabricate child sexual abuse material without contact offences, overwhelming hash-matching systems (Bennett et al., 2024; UNICEF, 2026). Thorn shows how tools turbo-charge grooming, as predators generate iterative deepfakes coercing children (Inabnett, 2026). Without provenance technologies and watermarking, power tilts toward organised abuse.
Data Gaps, Global South Epistemologies and AI
Davis characterises the Global South "data crisis" as colonial artefact distorting trafficking governance. ILO, IOM, and UNICEF surveys rely on formal labour markets and state-centric reporting, under- representing informal economies and communities with authority mistrust (Davis, 2024). AI built on skewed baselines risks entrenching epistemic injustice. Davis argues participatory data practices could invert this by centring survivor testimonies and local knowledge systems (Davis, 2024). Nontraditional intelligence, when co-produced with communities, reveals risk ecologies formal surveys miss. This requires shared dataset ownership, community veto over surveillance-increasing uses, and funding for Global South organisations to shape AI tools (Davis, 2024).
AI-Augmented Policing and Victim Identification
Commercial tools such as Veritone's IDentify and Track illustrate how AI extends law-enforcement reach by analysing vast digital evidence volumes impossible to review manually (Inabnett, 2026). Computer-vision models match faces across missing-person databases, CCTV feeds, and family- supplied images, while "human-like object" detection enables tracking without personally identifiable information, crucial where facial recognition is constrained (Inabnett, 2026). These systems support cross-border cooperation by correlating signals from different jurisdictions.
Sabo (2019) presents the use of the SAS Viya framework to orchestrate multiple data sources at strategic, operational and tactical levels with the help of AI. Sabo (2019) shows how machine-learning models can be applied to US State Department Trafficking in Persons reports to extract information about who is trafficked, for what purpose and under which governance arrangements, and then geospatially visualise patterns to guide interventions. The framework processes social media traces, migration datasets and national incident reports, providing multi-layered risk pictures (Sabo, 2019).
This is complemented by civil society innovation deployments like Thorn’s Spotlight and Solis platforms automatically scrape and organize online commercial sex ads to help officers identify children and connect cases across platforms (Bennett, 2024; Inabnett, 2026). The Polaris Project uses AI to analyze hotline data, which allows them to identify new trends in recruitment and victim profiles by transforming unstructured stories into actionable intelligence (Inabnett, 2026). For instance, “Traffick Jam” by Carnegie Mellon University uses machine learning to detect anomalies in posting patterns of escort ads that might suggest coercion (Inabnett, 2026).
Financial Intelligence, Nontraditional Data and Network Disruption
Financial institutions limit their understanding of trafficking as a financial crime and leave the resolution to police. Presently, machine-learning models analyse transactional flows to identify unusual login behaviour, cross-border transfer spikes or structured cash deposits that suggest funnel accounts laundering exploitation profits (Bagnall, 2025). Predictive analytics use historical case data, and socio-economic indicators to produce risk scores, leading to enhanced due diligence (Bagnall, 2025).
Bagnall (2025) distinguishes between “traditional intelligence” which relies on negative news and corporate registries, and “nontraditional intelligence” which is based on open-source information on websites, blogs and dark-web forums. This layer is noisy, and consists of granular clues about brothels and recruitment pages, minus the AI. Generative-AI systems now consume heterogeneous content (both traditional and non-traditional information), producing vector embeddings for trafficking typologies and semantic searches, and revealing linkages between aliases and phone numbers
(Bagnall, 2025). Further, external identifiers are used to link bank records in algorithms that resolve entities. While network-analysis tools chart links across jurisdictions.
Sabo’s framework shows how AI can be used to improve existing data models on trafficking, and to create new models considering the context, like armed conflict or the collapse of governance, that enable exploitation. Anti-trafficking AI should continue to evolve understanding of structural vulnerabilities (Sabo, 2019)..
Children, AI Governance and Anti-Trafficking
Children occupy exposed positions. Aitken shows AI increasingly mediates children's learning and service access, while children remain absent from design decisions (Aitken et al., 2025). Grimes describes this "digital childhood" where AI profiles children from infancy, yet governance leans on parental consent (Grimes, 2025). UN Committee on the Rights of the Child's General Comment No. 25 clarifies all Convention rights apply digitally, calling on states to regulate targeted advertising and automated decision-making (Committee on the Rights of the Child, 2021). ITU-UNICEF urge child- rights impact assessments and child-friendly complaint mechanisms (UNICEF, 2026).
AI-supported child protection interventions by the National Commission for Protection of Child Rights (NCPCR) in India illustrate how algorithmic tools can materially strengthen national child-safeguarding systems when embedded in an institutional reform agenda. Leveraging AI-enabled scanning and automated triage across sixteen integrated digital portals, NCPCR has, over a six‑month period, resolved approximately 26,000 child-rights complaints and facilitated the rescue of more than 2,300 vulnerable children from trafficking and other high‑risk situations, alongside the repatriation of over 1,000 children to their home districts (The Print, 2025; Habib, 2025). These interventions mark a shift from reactive, case-by-case policing toward more proactive, data-driven detection of online child sexual abuse material and related harms, while also demonstrating the scale that can be achieved when AI tools are coupled with coordinated action by state and district authorities.
For trafficking, AI tools detecting grooming must integrate children's experiences and agency, not cast them as data points (UNICEF, 2026). Algorithmic misclassification can mislabel consensual sexuality as exploitation or miss covert abuse. Rights-based design requires transparent model behaviour and appeal pathways children navigate.
Toward a Balanced Governance Agenda
The current literature highlights a central paradox: while AI enhances the capacity to detect trafficking, it also expands traffickers' capabilities and intensifies surveillance (Inabnett, 2026; Sabo, 2019; Tiwari, 2025). Child-rights scholars caution that generative AI may industrialize exploitation if governed solely by market forces (Bennett et al., 2024; UNICEF, 2026; Aitken et al., 2025; Grimes, 2025). An effective response requires rebalancing power in data production, prioritizing Global South perspectives, embedding AI within human-rights frameworks, and designing tools that utilize nontraditional intelligence without perpetuating colonial surveillance practices. Practitioners must address AI's dual- use nature, recognizing that harms are predictable outcomes of design decisions. Only through such measures can AI meaningfully shift the balance against traffickers.
Case Study
Case Study 1 - “Say Help”
Case Overview
In late October 2025, Delhi Police dismantled an alleged trafficking racket operating under the guise of a wellness spa in the Paharganj area of Central Delhi (Mishra & Ojha, 2025). The operation resulted in the rescue of six women and the arrest of one accused. The catalyst for the police intervention was an alert generated through the AI-enabled “Say Help” safety application, a voice-activated emergency platform designed primarily for women’s safety and public security. The premises in question, Galaxy Spa near Today Hotel on DBG Road, had already come under police scrutiny based on prior intelligence suggesting it was being used as a front for organised trafficking. A woman trapped inside the facility used the Say Help app to send an SOS during the night of October 30–31, 2025 (Mishra & Ojha, 2025). The alert transmitted live video, audio, and GPS location data to emergency contacts and the police control room. Acting on this intelligence, a decoy officer entered the spa with marked currency and confirmed the activity through a pre-arranged missed call (Mishra & Ojha, 2025). The operation was reported as a successful instance of AI-assisted anti-trafficking intervention.
App Features and Technical Architecture
Say Help is a global voice-activated emergency platform developed by US-based tech professional Mahesh Salgaonkar in 2023 (Rizwan, 2025). The app was inspired by a case in the United States involving two teenage girls who were raped and murdered, where privacy laws prevented police from accessing a victim’s phone for years despite the presence of critical evidence (Rizwan, 2025). The core design of Say Help is that users can trigger an alert through a custom voice command or a single push button, and the application immediately transmits live video, audio, and GPS details to emergency contacts and law enforcement.
Upon installation, users register and add trusted contacts to their emergency network. The app requests access to the camera and microphone, with permission options including “only this time” or “while using the app,” though users are urged toward the latter for continuous functionality. Once launched, the app can run quietly in the background. Alerts can be triggered by tapping “Send alert” inside the app or by using the voice command “Help, help, help.” The voice trigger works only after users tap “Activate” once; after that, they can close the app and use the voice command as long as the phone’s microphone remains accessible (Mishra & Ojha, 2025).
The app supports up to five custom trigger phrases in any language. Each phrase is registered by repeating it three times, and the system accepts it if it recognises the phrase correctly at least twice. When triggered, the phone begins streaming without any visible sign. A link is sent instantly to emergency contacts and the control room, opening into a dashboard that displays the user’s details, the live video feed, and a real-time map tracking movement and speed until the alert is closed. Say Help is reportedly being used in 18 countries, including Mexico, India, Chile, Peru, the UAE, Kuwait, and Colombia, with alerts from women and senior citizens prioritised (Rizwan, 2025).
Law Enforcement Implementation and Operational Workflow
Say Help was integrated with the Central district police in Delhi in September 2025. A small control room is established behind Daryaganj Police Station with three computers and an 11-member team that manages a steady stream of beeping alerts. On average, 40 pings reach the app control room every day, many of which are test alerts from people experimenting after hearing about the app. The app has been downloaded over 5,000 times in India.
The operational workflow is manual and human-mediated. When an alert appears genuine, the team manually identifies the nearest police station and shares a link mirroring the one sent to emergency contacts. They can track police movement in real time as officers head towards the user. Accidental triggers are frequent, often more than 10 a day, and only one or two alerts usually warrant being forwarded to the police control room (Rizwan, 2025). The developer’s team reviews each alert and shuts down those that appear frivolous. The video feed cannot be downloaded, and police can access footage only through a request under proper legal procedure. Once an alert is closed, even the Say Help team cannot access the history. Further, only one authorised person in Delhi Police can search a specific alert.
Use of AI and Technical Challenges
Say Help was originally designed as a broader safety and emergency-response platform, and its developer has been attempting to add AI features for automated routing and crime-type detection. These planned AI functions were intended to help the system interpret incidents more intelligently, including identifying harassment, assault, domestic violence, or similar risks from live audio and video. However, the planned AI features have hit significant roadblocks, both technical and contextual.
A major setback came from Delhi Police’s geo-fencing design. Ideally, police boundaries are mapped precisely, but in Delhi, these borders are thick 100-metre buffer zones. The AI struggled to interpret these ambiguous border zones confidently. It was noticed that most crimes were committed within these border regions. As the AI could not confidently map these areas, alerts from these areas could not be automatically routed and the team presently undertakes manual tracking.
Another planned feature is crime-type detection, where AI analyses video and audio to infer whether an incident involves harassment, assault, domestic violence, or something else. This requires extensive training across accents, dialects, and noisy environments. The technical limitation underscores how digital systems stumble when the social world resists clean categorization. Without training data that reflects local speech patterns, noisy environments, and linguistic diversity, the system may work well in privileged contexts but poorly in the settings where protection is most needed.
Criticism and Governance Concerns
The deployment of Say Help has sparked major concerns around consent, surveillance and the political economy of AI-enabled protection. Though the app may have helped the police rescue six women that night, it also demonstrates a bigger trend of how safety technologies tend to be positioned (Rizwan, 2025). Critics argue that tools like Say Help can risk framing complex issues like trafficking as things that can be solved by quick police response. This framing turns structural harms about law and order when safety is built around surveillance and enforcement, deeper questions of survivor justice, long- term support and prevention tend to fall to the wayside (Rizwan, 2025). Further, it is pointed that even with real-time alerts the end point is still the police—and systemic delays may not change. (Rizwan, 2025) This critique is an important one because protection systems tend to succeed narrowly on the metric of interception but not on larger questions of dignity, trust, and institutional repair.
The app also raises concerns about covert monitoring and consent. The app can run in the background and requires access to the camera and microphone once the user taps “Activate,” so those who are less digitally experienced may not realize they’ve left it in a state where it is still listening for voice triggers even when they don’t mean for it to. The app can also be used in the event of mutual fights or disputes which could lead to unnecessary police intervention or recording of bystanders without their consent. These are not edge cases of technical but of governance issues. A surveillance apparatus that secretly films people with little transparency can develop coercive power more rapidly than it develops accountability. This concern is heightened in a society where marginalized communities already experience disproportionate surveillance and weaker access to legal redress.
Finding
This case demonstrates that AI-enabled safety tools can improve detection and response capacity, but their policy value depends critically on safeguards, accountability, equity, participation, and institutional coordination. The Say Help case illustrates a socio-technical arrangement where rescue, monitoring, consent, and control become tightly entangled. The app clearly helped police act quickly in one trafficking case, but it also reveals the governance gap between technical capability and rights- based, context-sensitive protection. The correct unit of analysis is not the app alone but the entire socio-technical arrangement around it, including police capacity, vendor incentives, legal safeguards, and the lived vulnerabilities of users. Measures around protection of case history and accessibility matter because they indicate an effort to build procedural restraint into the platform. Still, restraint does not erase asymmetry. It is contended that the architecture still concentrates evidentiary power in a small circle of institutional actors. The result is a safety regime that promises empowerment while embedding asymmetrical visibility. This case supports the broader finding that technological fixes cannot replace survivor-centred services, community-based prevention, or durable legal accountability.
CASE STUDY – GARUDA DRISHTI
Case Overview
Garuda Drishti (“Eagle Eye”) is an advanced social media monitoring and cyber intelligence project launched by Maharashtra Chief Minister Devendra Fadnavis in Nagpur on August 10, 2025. Originally designed to detect fraudulent activities, communal hate, and fake news, the system was developed by Nagpur Police to curb social media misuse. In a ground-breaking initiative, Nagpur Police upgraded Garuda Drishti to tackle digital trafficking as part of their ongoing Operation Shakti mission against flesh trade networks. This marks the first time in Maharashtra that an AI-driven software was adapted specifically to address human trafficking and prostitution rackets (Bose, 2025). In parallel, Nagpur Police deployed eight AI-enabled facial recognition cameras at Jadhav Chowk near Ganeshpeth Bus Stand, a known trafficking hotspot. This has resulted in three raids taking place in just one month at the Jadhav Chowk surveillance zone (Bose, 2025). According to additional reporting, the operation sealed 23 premises for running sex rackets in a separate enforcement wave (Bose, 2026). Another report indicates that Operation Shakti rescued 85 women from flesh trade rackets in Nagpur (Loksatta, 2025).
App Features and Technical Architecture
Garuda Drishti is an AI-driven software system that scours social media platforms like Facebook and Instagram for signs of illicit activities. The system detects fake job scams, escort services, and fraudulent advertisements that are markers of online exploitation (Bose, 2025). This cutting-edge feature accelerates the identification of online crimes, enabling rapid case registration under the Bharatiya Nyaya Sanhita (BNS) and the Immoral Traffic (Prevention) Act (ITPA). The deployed cameras are loaded with data on trafficking survivors, suspects, and racketeers, and they send real-time alerts to the nearest police station facilitating swift action. As per news reports, the AI-driven facial recognition system is helping police locate missing women and children while tracking repeat offenders with precision (Bose, 2025). The initiative is supported by a private NGO and leverages data from past operations to map trafficking-prone areas.
Use of AI and Technical Challenges
Garuda Drishti started as a general social media monitoring project, looking for fake news, communal hate and fraudulent activity. Police adapted it to their anti-trafficking work, expanding its function to scan Facebook and Instagram for escort advertisements, fake job offers, and other markers of online exploitation (Bose, 2025). That change matters because it shows how a general-purpose surveillance tool morphs into a targeted law-enforcement instrument. The same architecture that tracks digital misinformation can also track bodies, vulnerabilities and illicit markets. In the process the boundaries between cyber policing and social control are blurred.
The facial recognition system is being enhanced with past operational data and NGO support to improve relevance and reduce false positives. On the one hand, the system's accuracy can be improved by the inclusion of local intelligence. The new system also takes into account prior policing decisions, on the other hand. If past raids were examples of bias, selective attention, or uneven complaint-registration, the model may reproduce those distortions at scale. AI does not wipe out institutional memory. It does this automatically. That’s its strength and its danger.
But social media monitoring for “fraudulent advertisements” or “fake job scams” can easily catch ambiguous speech, satire and informal labor promotion that the algorithm can’t reliably differentiate. Facial recognition systems bring another layer of uncertainty in terms of identification, particularly in dense urban environments where lighting, pose and camera angle can all distort. A false match in the trafficking context is not a technicality. This can lead to detention, stigmatization or coercive intervention. Speed is nothing without accuracy.
Criticism and Governance Concerns
The Garuda Drishti and Operation Shakti case brings to light serious concerns about governance and due process. The system does not just "discover" trafficking, but classifies and prioritizes forms of visibility. Platform scraping privileges content that looks legible to digital systems. Face recognition privileges those already keyed into the database. Communities that leave fewer searchable digital traces, or that move through the city in ways that are not easily mapped by algorithmic systems, risk remaining peripheral to the system’s gaze. The result is patchy coverage. The technology could be very useful for repeat cases in data rich corridors but less useful where social exclusion, informal labor, migration or linguistic diversity make recognition more difficult.
There’s still little accountability. Who checks the system? Who will set it right? What are the standards for facial recognition matches, social media flagging or alert escalation? unclear procedures for accountability, AI-powered policing can turn from protecting the public into opaque risk management. The initiative reflects administrative ambition, inter-agency coordination, and a willingness to invest in advanced tools. But the more important question is about governance structures and oversight mechanisms on which the information is scarce.
Finding
Garuda Drishti and Operation Shakti demonstrate that AI-enabled policing tools can improve detection capacity and operational speed, but they also reveal critical governance gaps in accountability, equity, participation, and institutional coordination. The case illustrates how general- purpose surveillance infrastructure becomes repurposed for targeted law enforcement, narrowing the boundary between cyber policing and social control. The system shows measurable utility in rescue operations and hotspot monitoring, yet it simultaneously raises due process concerns, bias risks, and surveillance asymmetries. This case supports the broader finding that technological fixes cannot replace survivor-centred services, community-based prevention, or durable legal accountability.
Findings
Artificial intelligence is expanding the arsenal of tools for identifying trafficking patterns, but its use for child and youth protection creates a parallel architecture of risk. Indeed, evidence from India, Europe and international guidance show that AI does not only detect harm. It transforms the space in which harm is identified, prioritized and addressed (UNESCO, 2021; UNICEF Innocenti, 2026). This reshuffling spreads vulnerability but develops efficiencies. Systems that identify suspicious advertisements and flag exploitation at the same time misunderstand social context, amplify biases and direct state attention to populations already subject to disproportionate surveillance. AI is an ambivalent infrastructure: it renders some victims more visible but makes others newly legible to coercive oversight.
AI creates two linked dangers here: algorithmic targeting and automation bias. Predictive systems inherit observational bias from historical data, so they often classify poor youth, migrants, and welfare-visible groups as “at risk” because those populations generate more institutional traces, not because they face greater victimization. Models trained on complaint-heavy urban or English- language datasets therefore overflag communities already close to surveillance while missing victims in informal labor networks or linguistically marginalized spaces (OECD, 2024).
The problem deepens when investigators trust these outputs too readily. False positives can trigger raids, child removals, or reputational harm, and those effects often outlast the investigation itself. AI 22
Figure 4
Comparative Insights From Case Studies (Source: Author's Own Development)
then does more than misclassify. It builds durable pathways of suspicion around marginalized communities, and those pathways are difficult to unwind.
Deepfake-enabled grooming exacerbates this problem by changing the method of abuse. Generative AI makes the technical creation of synthetic sexualized imagery and manipulated videos used to coerce victims easier. The smugglers no longer rely only on physical access. They coordinate reputational assault at scale and ensnare young people in cycles of fear before institutions record abuse (UNESCO, 2021). Regulatory regimes acknowledge synthetic harms, but operational systems lack the ability to detect them. Exploiters adapt faster than the institutions that are supposed to stop them.
Regulatory gaps are still wide. Most frameworks treat AI safety in general ethical terms, rather than trafficking-specific harms. India’s Digital Personal Data Protection Act provides a baseline for child data protection but offers no trafficking-focused guidance on how AI should classify or audit cases especially those involving minors. The EU AI Act mandates high-risk duties, but leaves victim-centred safeguards in multilingual contexts insufficiently prescribed. Internationally, UNESCO and UNICEF offer guidance on ethics but lack operational specificity around false positives, survivor-informed audits, or cross-border evidence sharing.
Implementation gaps remain acute. AI-assisted triage works better in some Indian districts, but outcomes still vary sharply because staff capacity, data quality, and institutional interoperability differ across regions. Where cybercrime units are staffed and technically competent, AI adds real value; elsewhere, the same software operates inside a weak ecosystem, constrained by bandwidth and limited operational support. The result is an illusion of scale that conceals fragile implementation.
Coordination failures deepen the problem. Anti-trafficking governance brings together child- protection commissions, police, regulators, and platforms, yet their priorities rarely align. Siloed mandates slow referrals, fragment responsibility, and prevent AI-generated leads from moving smoothly from moderation to law enforcement to rehabilitation. Even when the system flags correctly, institutional handoffs often break down.
Protection is uneven geographically. Among the other weaknesses are multilingual blind spots. AI works best in English. Trafficking communications spread in mixed dialects and coded vernaculars. Victims who speak a less represented language are more likely to be overlooked. The rural-urban divide exacerbates inequalities. Urban centres have better connectivity and specialty units. These do not occur in country areas. Protection becomes an element of the institutional landscape, not just the algorithm.
AI improves protection in certain situations. The evidence suggests that strong human in the loop systems are a basic safeguard. Having analysts review leads improves the quality of detection and fairness of the process. The strongest results appear to be where governance combines technical precision with accountability. Tools like Thorn Spotlight show that AI can be a powerful aid to victim identification, with the support of structured agreements and documented human review and operational feedback (Thorn, 2024).
The results are not a binary assessment of AI. The value of AI depends on the appropriateness of governance structures. AI improves detection when systems are multilingual, supervised, audited and integrated with care infrastructures. But it can also be destructive when used as a substitute for robust institutional arrangements. This distinction should inform the design of future policy.
Evaluation Model
This paper presents a proposed evaluation model which stems from a simple but outstanding problem: current AI governance frameworks do not adequately assess AI systems in anti-trafficking and child protection settings, prioritizing formal compliance, abstract ethics, or technical performance in isolation. Voluntary, broadly applicable AI risk frameworks, such as the NIST AI Risk Management Framework that emphasizes trustworthiness, governance, measurement, and risk management, are not meant for institution-specific assessment in the context of trafficking (NIST, 2023). UNICEF’s child- centred AI guidance offers a rights-based orientation, emphasizing protection, inclusion, fairness, privacy, transparency and meaningful participation of children, but it is mostly normative guidance and not a field-ready regulatory scoring tool (UNICEF, 2026). Similarly, UNESCO’s ethics framework provides foundational principles such as proportionality, fairness, accountability, human oversight and non-discrimination, but it is deliberately high-level and does not say how one should compare an Indian anti-trafficking deployment with an EU regulatory regime or a regional South Asian coordination mechanism (UNESCO, 2021). The difference is no concept. It is a lack of evaluation.
This paper proposes a model for treating anti-trafficking AI regulation as a governance problem that is measured across contexts, rather than a compliance exercise limited to whether a statute mentions transparency or whether a vendor publishes an impact statement. So there are four design principles. First the model should be comparable across contexts. A framework that cannot meaningfully compare India, the European Union, and lower-capacity South Asian jurisdictions will obscure the very unevenness that this paper seeks to illuminate. Second, the model has to be sensitive to the realities of the Global South, including issues such as multilinguality, connectivity constraints in rural areas, institutional fragmentation and asymmetrical state capacity. Third, it must foreground equity and participation, for child protection systems that are blind to caste, class, gender, age and linguistic exclusion will reproduce the same hierarchies they claim to mitigate. Fourth, it has to be compatible with polycentric governance. AI in anti-trafficking is not just made up of the state, but also the platforms, telecom infrastructures, NGOs, survivor networks, auditors and channels of cross-border policing. This distributed authority structure must be taken into account in any serious model (Ostrom, 2009; OECD, 2024; UNICEF, 2026).
The proposed model adopts five pillars with equal weighting: Effectiveness, Technical Robustness, Governance and Adherence, Equity and Inclusion, and Assurance and Adaptation. Each pillar captures a dimension that existing frameworks tend to isolate from the others. Together they provide a more complete account of whether AI regulation in anti-trafficking contexts works, for whom it works, under what institutional conditions, and at what social cost.
The first pillar, Effectiveness, asks whether AI-enabled regulation and deployment actually reduce harm. Effectiveness draws from the findings of this study, especially the gap between formal legal safeguards and operational outcomes. The key metrics are Risk Reduction Evidence, Cost- Effectiveness, and Comparative Benchmarking. Risk Reduction Evidence measures the number and quality of causal or quasi-causal studies showing that a regulatory or technological intervention reduces trafficking or exploitation incidence. Cost-Effectiveness captures the cost per victim identified or protected, forcing evaluators to examine whether resource use aligns with public value. Comparative Benchmarking assesses how a jurisdiction performs relative to peer systems with similar or superior institutional capacity. Data sources include implementation reports, law-enforcement datasets, independent evaluations, and public budgets.
The second pillar, Technical Robustness, evaluates whether a system performs reliably under real- world conditions rather than under idealized testing assumptions. The core metrics are Detection Accuracy, Adversarial Robustness, Multilingual Performance, and Future-Proofing. Detection Accuracy measures the proportion of validated trafficking-related cases correctly identified by the system. Adversarial Robustness measures how well the system resists evasion tactics such as coded language, image manipulation, and platform migration. Multilingual Performance measures whether the system functions across major and minor languages, dialects, and transliterated scripts rather than defaulting to English or dominant-language corpora. Future-Proofing examines whether the regulatory and technical design includes pathways to respond to emerging threats such as deepfakes, synthetic child sexual abuse material, or AI-generated grooming content. Data come from audit logs, bias tests, and language-specific evaluations.
The third pillar, Governance and Adherence, shifts attention from models to institutions. The four principal metrics are Implementation Feasibility, Staff Empowerment, Compliance Monitoring, and Escalation Protocols. Implementation Feasibility measures whether a jurisdiction has budget, training, staffing, and infrastructure sufficient to sustain AI deployment rather than merely announce it. Staff Empowerment measures whether caseworkers, analysts, or regulators possess authority to override automated outputs, halt unsafe deployments, or refer cases upward. Compliance Monitoring measures whether independent or quasi-independent audits occur at regular intervals and whether findings influence practice. Escalation Protocols evaluate the speed and clarity with which high-risk cases move from algorithmic identification to human review, referral, and response. Data sources include staffing records, audit reports, case-management systems, and agency guidelines.
The fourth pillar, Equity and Inclusion, sits at the normative centre of the model. The metrics are Rural Deployment, Marginalized Focus, Survivor Co-Design, and Capacity Equity. Rural Deployment measures parity between urban and rural access to AI-enabled protection systems. Marginalized Focus measures whether systems identify, protect, and avoid disproportionately harming groups that already face structural disadvantage, including low-income communities, caste-marginalized groups, ethnic minorities, and gender-nonconforming youth where relevant and ethically measurable. Survivor Co-Design measures whether survivors, youth representatives, and civil-society actors participate in the governance, review, or redesign of systems. Capacity Equity measures whether training, infrastructure, and institutional support extend beyond metropolitan centres. Data sources include deployment maps, meeting records, training rosters, demographic complaint data, and stakeholder interviews.
The fifth pillar, Assurance and Adaptation, draws on NIST’s trustworthiness orientation, UNESCO’s principles of accountability and transparency, and the paper’s critique of static policy models. Its metrics are Independent Audit, Transparency Score, Adaptation Speed, and Global Benchmarking. Independent Audit measures whether a third party verifies safety, fairness, and compliance claims. Transparency Score captures the availability and intelligibility of public-facing information about system goals, performance, oversight, and complaints processes. Adaptation Speed measures how quickly institutions update policy or technical safeguards in response to new threats such as AI- generated exploitation materials. Global Benchmarking measures alignment with internationally recognized standards without collapsing local context into imported templates.
The model uses a simple scoring logic. Each metric is scored from 1 to 5, where 1 indicates absence or severe weakness and 5 indicates strong, credible, and context-sensitive performance. Metrics within each pillar are averaged, and the five pillar scores are then equally weighted to generate an overall score.
AI SAFETY FRAMEWORK
Precise, Measurable Definitions With Data Sources and Calculation Methods
PILLAR 1: EFFECTIVENESS (20%)
Figure 5
AI Safety Framework – Assurance & Adaptation (Source: Author's Own Development)
Figure 6
AI Safety Framework – Equity & Inclusion (Source: Author's Own Development)
Figure 7
AI Safety Framework – Governance & Adherence (Source: Author's Own Development)
Figure 8
AI Safety Framework – Technical Robustness (Source: Author's Own Development)
Figure 9
AI Safety Framework – Effectiveness (Source: Author's Own Development)
SCORING METHODOLOGY. Pillar Score = Average of 4 metrics (1-5 scale) Overall Score = (P1×0.20) + (P2×0.20) + (P3×0.20) + (P4×0.20) + (P5×0.20) 2.0-2.9 = DEVELOPING 3.0-3.9 = ADEQUATE 4.0-5.0 = EXEMPLARY
Recommendation
The findings carry immediate policy implications because governance quality, not technical sophistication alone, determines whether AI-assisted anti-trafficking interventions protect victims or reproduce inequality. National governments, regional institutions, private platforms, and civil society shape that outcome. The evaluation model developed in this paper should function as a decision instrument, not an academic accessory. It can guide budgeting, procurement, inter-agency design, regulatory review, and cross-border cooperation by shifting attention from broad promises to measurable performance. This matters especially in developing-country contexts, where AI strategies foreground innovation while leaving victim-centred oversight under-specified (UNICEF, 2025; OECD, 2024).
For national governments, especially India and comparable states, the first task is institutional integration. AI evaluation should sit inside national AI strategies, child online protection policies, cybercrime protocols, and procurement rules. It should not remain a detached compliance annex. If governments adopt the proposed model, agencies must demonstrate not only that systems work, but also for whom they work, under what safeguards, and with what accountability. That requirement fits India's policy landscape, where data protection, telecom regulation, cybercrime enforcement, AI governance, and child welfare overlap yet rarely share a common evaluative language.
Institutional capacity is decisive. Governments should allocate dedicated budgets for multilingual model testing, bias audits, child-rights impact assessments, and rural deployment support instead of treating such work as optional refinement. Training should extend beyond technical teams. Police units, child-protection officers, district administrators, and judicial actors need practical literacy in how AI classifies risk, how errors travel through workflows, and how automated outputs can be challenged. Cross-agency task forces should become standard, especially for cases involving minors, platform evidence, and rapid referral. Those task forces should include law enforcement, child commissions, education ministries, social welfare, telecom regulators, and digital-governance agencies. Trafficking prevention is not policing alone. A case flagged online may require rescue, counselling, school reintegration, legal support, and family-level intervention in rapid succession. Fragmented institutions cannot deliver that chain of care effectively (UNICEF, 2025; UNESCO, 2021).
Multilingual, rural-focused, and intersectional deployment must become regulatory obligations. Systems trained mainly in English or dominant national languages miss trafficking signals embedded in dialects, transliterated scripts, and local codes. Rural underdeployment compounds the problem by leaving areas with weak institutional presence reliant on tools never calibrated for their linguistic and infrastructural conditions. Governments should require region-specific validation before scaling tools nationally. Procurement standards should mandate testing across multiple languages, low-bandwidth environments, and socially diverse user groups. Child-protection systems should integrate intersectional review, examining whether caste, ethnicity, migration status, disability, gender identity, or informal labour conditions increase misclassification or non-detection. Without that review, AI becomes a machine of administrative invisibility where exploitation is hardest to see (UNICEF, 2025).
Regional and international bodies face a parallel challenge. They need to move beyond coordination rhetoric toward usable regional compacts. South Asia lacks a cohesive framework linking AI governance, child online protection, and anti-trafficking cooperation in ways reflecting cross-border realities. Labour migration, platform-mediated recruitment, linguistic plurality, and informal trafficking routes cross national boundaries, yet governance remains siloed. A shared regional template should benchmark readiness, identify capacity gaps, and harmonize minimum safeguards across countries. States could agree on common standards for human review, data-sharing documentation, linguistic testing, and audit disclosure while preserving national legal variation (ITU, 2023; Internet Society, 2017).
Global guidance from the UN, OECD, and related bodies already supplies principles of accountability, transparency, fairness, and human rights. The problem lies in translation. International institutions should adapt these norms for developing-country conditions rather than assuming high-capacity models travel unchanged. That means operational guidance for low-resource environments, regional peer learning, and funding for audit ecosystems, multilingual evaluation, and child-rights-based AI 28 review. Regional plans emerging in Southeast Asia around child online protection offer a useful precedent because they combine multi-stakeholder coordination, cross-border referral logic, and child-centred co-creation rather than relying only on national enforcement silos (Safe Online, 2023; APLE Cambodia, 2025). South Asia needs similar architecture adapted to its political and linguistic terrain.
Platforms, private firms, and technical actors cannot remain peripheral. They control critical infrastructures of visibility. They hold data, shape recommendation systems, and set practical limits on what investigators and child-protection agencies can see. Their responsibilities should extend beyond reactive cooperation. Platforms should participate in polycentric governance arrangements that include formal data-sharing protocols, independent auditing, plain-language transparency reporting, and child-rights impact review. OECD principles already stress accountability, traceability, transparency, and meaningful human oversight; platforms should operationalize them in trafficking and child-protection contexts, not merely cite them in policy statements (OECD, 2024). That includes disclosure of detection systems, escalation criteria, error management, and redress for wrongful flags.
Co-development of safeguards with survivors, youth representatives, and civil society should become standard practice. This is not symbolic. Survivors and frontline advocates understand exploitative patterns, coded communication, platform vulnerabilities, and institutional harms in ways engineering teams often do not. Their participation improves legitimacy and accuracy. Ethical audit commitments should become stricter. Vendors should permit independent third-party review of high-risk systems used in child-protection and trafficking settings, including tests for bias, multilingual performance, false positives, and unintended harms. Proprietary opacity cannot stand as a blanket defense where children's safety and civil liberties are at stake (UNICEF, 2026; UNESCO, 2024).
Civil society, survivors, and youth need stronger institutional standing. Advocates should ask whether systems have been audited, whether rural deployment parity has improved, whether local languages appear in validation datasets, and whether survivor participation has shaped escalation protocols. This shifts debate away from AI salvation toward governance. Civil society groups should conduct shadow evaluations, publish scorecards, and join oversight councils. Survivor groups and youth networks should influence design from the outset, especially in reporting interfaces, trauma-sensitive workflows, and institutional response pathways (UNICEF, 2025).
Meaningful participation requires infrastructure. Governments and donors should fund translation, accessibility support, digital literacy, and safe consultation mechanisms so participation does not become a privilege of the already empowered. Participation must alter decisions. Otherwise, it becomes decorative governance.
AI regulation in anti-trafficking contexts should be judged not by high-level principles or isolated rescue statistics, but by whether institutions deliver equitable, accountable, multilingual, and survivor- centred protection across uneven realities.
Conclusion
This paper shows that AI neither automatically protects nor inherently threatens children and youth in digital trafficking environments. It improves detection and response only when institutions embed it within human review, clear escalation pathways, multilingual capacity, and accountable coordination. Where those conditions exist, AI strengthens triage, sharpens lead generation, and supports victim identification; where they do not, it amplifies misclassification, surveillance, and exclusion (Thorn, 2025; UNESCO, 2021; UNICEF, 2026).
The comparative analysis also confirms that policy maturity alone cannot solve the problem. India’s legal architecture has advanced in data governance and AI institutionalization, but trafficking-specific safeguards remain uneven. The European model offers more elaborate compliance structures, yet it assumes administrative capacity that many developing contexts lack. International norms provide ethical direction, but they remain too abstract to resolve language gaps, audit deficits, redress failures, and fragmented institutional responsibility (OECD, 2024; UNESCO, 2021; UNICEF Innocenti, 2025).
The evaluation model presented in this paper assesses AI safety regulations across five pillars. It is designed to be adaptable across regions and types of harm, necessitating recalibration for local languages, legal institutions, and patterns of exclusion. The model prioritizes accountability, participation, and institutional realities over abstract aspirations.
The paper therefore argues for a polycentric and participatory model of AI governance. Protection in online trafficking cases requires more than centralized surveillance or technical efficiency. It requires shared authority among states, platforms, child-protection bodies, law-enforcement agencies, NGOs, and survivors themselves. That model better supports equity, resilience, and peacebuilding because it reduces the institutional violence that often accompanies opaque technological intervention.
The broader implication is clear: AI can expand protective capacities only when governance is substantive, localized, and grounded in rights-based principles. Absent these conditions, AI may increase state surveillance without enhancing protection. Ethical considerations must be integral to regulation, shaping design, oversight, and implementation from the outset.
References
Aitken, M., Mahomed, S., Briggs, B., & Atabey, A. (2025, September 23). How should children’s rights be integrated into AI governance? UNESCO. https://www.unesco.org/en/articles/how-should-childrens- rights-be-integrated-ai-governance
Aneja, A., & Shrivastava, A. (2026). Balancing innovation and equity: The Global South’s role in shaping AI copyright policy. BRICS Law Journal, 13(1), 56–75. https://www.bricslawjournal.com/jour/article/download/1572/350
APLE Cambodia. (2025, November 21). 2025 ASEAN ICT Forum on Child Online Protection: Empowering Children and Strengthening ASEAN’s Commitment. APLE Cambodia. https://aplecambodia.org/2025- asean-ict-forum-on-child-online-protection-empowering-children-and-strengthening-aseans- commitment/
Arora, P. (2025). Creative data justice: a decolonial and indigenous framework to assess creativity and artificial intelligence. Information Communication and Society, 28(13), 2231–2247. https://doi.org/10.1080/1369118X.2024.2420041
Bagnall, C., Krupena, S., Kemp, C., & Massimi, F. (2025). AI’s Role in Combating Human Trafficking in the Financial Sector. Retrieved May 26, 2026, from https://www.acams.org/en/opinion/ais-role-in- combating-human-trafficking-in-the-financial-sector
Bennett, P., Cucos, R., & Winch, R. (2024). New Frontiers The use of Generative AI to facilitate trafficking in persons [Policy Brief]. OSCE Office of the Special Representative and Co-ordinator for Combating Trafficking in Human Beings. https://cthb.osce.org/sites/default/files/f/documents/7/d/579715.pdf
Bose, S. (2025, August 28). Cops unleash AI arsenal to crush human trafficking, prostitution rackets. The Times of India. https://timesofindia.indiatimes.com/city/nagpur/cops-unleash-ai-arsenal-to-crush-human- trafficking-prostitution-rackets/articleshow/123551729.cms
Bose, S. (2026). Cops seal 23 premises for running sex rackets. The Times of India. https://timesofindia.indiatimes.com/city/nagpur/cops-seal-23-premises-for-running-sex- rackets/articleshow/127952560.cms
Committee on the Rights of the Child. (2021). General comment No. 25 (2021) on children’s rights in relation to the digital environment. https://www.right-to-education.org/resource/general-comment-no-25-2021- children-s-rights-relation-digital-environment
Couldry, N., & Mejias, U. A. (2020). The costs of connection: How data is colonizing human life and appropriating it for capitalism. Social Forces, 99(1), page 6. https://doi.org/10.1093/sf/soz172
D’Ignazio, C., & Klein, L. F. (2020). Data Feminism. https://doi.org/10.7551/mitpress/11805.001.0001
Data Security Council of India. (2025). Summary of India AI governance guidelines. https://www.dsci.in/files/content/knowledge-centre/2025/summary-of-india-ai-governence- guidelines.pdf
Davis, J. (2024). Bridging the gaps: Human trafficking data in the Global South. https://www.jarrettdavis.org/blog/bridging-the-gaps-37fj3
EU-Lisa & Eurojust. (2022). Artificial Intelligence supporting cross-border cooperation in Criminal Justice. EU- Lisa & Eurojust. https://www.eurojust.europa.eu/publication/artificial-intelligence-supporting-cross- border-cooperation-criminal-justice
European Commission. (2024). The EU artificial intelligence act, 2024. European Commission. https://artificialintelligenceact.eu/
FAIST. (2024). The global AI market is expected to reach $1.81 trillion by 2030. https://www.faistgroup.com/news/global-ai-market-2030/
Gillespie, T. (2018). Custodians of the internet: Platforms, content moderation, and the hidden decisions that shape social media. Yale University Press.
Gorwa, R., Binns, R., & Katzenbach, C. (2020). Algorithmic content moderation: Technical and political challenges in the automation of platform governance. Big Data & Society, 7, 205395171989794. https://doi.org/10.1177/2053951719897945
Greiman, V., & Bain, C. (2012). The emergence of cyber activity as a gateway to human trafficking. International Journal of Cyber Warfare and Terrorism, 12(1). http://blogs.bu.edu/ggreiman/files/2013/10/ICIW2013JournalonInfoWarfareGREIMAN.pdf
Grimes, S. (2025, January 23). Responsible AI and Children: Towards a Rights-Based Approach to AI Governance. Joan Ganz Cooney Center. https://joanganzcooneycenter.org/2025/01/23/responsible-ai- and-children/
Habib, A. (2025, November 20). Arunachal child rights panel reports 26,000 cases disposed, 2,300 children rescued in 6 months. India Today NE. https://www.indiatodayne.in/arunachal-pradesh/story/arunachal- child-rights-panel-reports-26000-cases-disposed-2300-children-rescued-in-6-months-1306315-2025-11- 20
Hester, M., Williamson, E., Eisenstadt, N., Abrahams, H., Aghtaie, N., Bates, L., Gangoli, G., Robinson, A., Walker, S. J., McCarthy, E., Matolcsi, A., & Mulvihill, N. (2025). What Is Justice? Perspectives of Victims- Survivors of Gender-Based Violence. Violence against women, 31(2), 570–597. https://doi.org/10.1177/10778012231214772
Inabnett, K. (2026, January 27). AI Revolutionizes the Fight Against Human Trafficking. Veritone. Veritone. https://www.veritone.com/blog/ai-public-safety-human-trafficking/
IndiaAI. (2025). India takes the lead: Establishing the IndiaAI Safety Institute for responsible AI innovation. https://indiaai.gov.in/article/india-takes-the-lead-establishing-the-indiaai-safety-institute-for- responsible-ai-innovation
International Centre for Missing & Exploited Children. (2018). Studies in child protection: Technology- facilitated child sex trafficking. Author. https://www.icmec.org/wp- content/uploads/2018/12/Technology-Facilitated-Child-Sex-Trafficking_final_11-30-18.pdf
International Telecommunication Union (ITU). (2023). Child Online Protection Asia and the Pacific Concluding Workshop Conference Report. https://www.itu.int/en/ITU-D/Regional- Presence/AsiaPacific/Documents/Child%20Online%20Protection%20Asia%20and%20the%20Pacific%20C oncluding%20Workshop%20Conference%20Report.pdf
Internet Society. (2017). Mapping Online Safety in Asia Pacific. https://www.icmec.org/wp- content/uploads/2018/08/Mapping-Online-Safety-in-11-Asian-Countries-Internet-Society.pdf
INTERPOL. (2025). Preventing the cycle of harm. https://www.interpol.int/en/Crimes/Crimes-against- children/Projects/Preventing-the-Cycle-of-Harm
Jeevanandam, N. (2024). India tops global AI skill penetration and talent concentration rates. IndiaAI. https://indiaai.gov.in/news/india-tops-global-ai-skill-penetration-and-talent-concentration-rates
Jeevanandam, N. (2024a). India leads global AI talent and skill penetration. IndiaAI https://indiaai.gov.in/article/india-leads-global-ai-talent-and-skill-penetration
L’Hoiry, X., Moretti, A., & Antonopoulos, G. A. (2024). Human trafficking, sexual exploitation and digital technologies. Trends in Organized Crime, 27(1), 1–9. https://doi.org/10.1007/s12117-024-09526-4
Latonero, M. (2011). Human trafficking online: The role of social networking sites and online classifieds. University of Southern California, Annenberg Center on Communication Leadership & Policy.
Latonero, M. (2012). Technology and human trafficking: The rise of mobile and the diffusion of technology- facilitated trafficking. USC Annenberg Center on Communication Leadership & Policy. https://technologyandtrafficking.usc.edu/files/2012/11/HumanTrafficking2012_Nov12.pdf
Loksatta. (2025, August 28). Operation Shakti: Nagpur police rescue 85 women from flesh trade racket. Loksatta. https://www.loksatta.com/nagpur/operation-shakti-nagpur-police-rescue-85-women-from- flesh-trade-racket-rub-99-5762486/
Mahajan, M. (2025). India’s AI Safety Institute: Ensuring Ethical and Safe AI Applications. Linkedin. https://www.linkedin.com/pulse/indias-ai-safety-institute-ensuring-ethical-safe-applications-czt5c/
Manoraj, V. (2021). Hidden Tsunami: Technology Facilitated Trafficking and Responsibility of Online Platforms in India. Sewa International. Delhi. https://www.sewainternational.org/RESOURCES/Publications/Human_Trafficking_Report.pdf
Mhlambi, S., Tiribelli, S. (2023). Decolonizing AI Ethics: Relational Autonomy as a Means to Counter AI Harms. Topoi 42, 867–880 https://doi.org/10.1007/s11245-022-09874-2
Ministry of Electronics and Information Technology. (2025). With robust and high end common computing facility in place, India all set to launch its own safe & secure indigenous AI model at an affordable cost. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2097709®=3&lang=2
Mishra, H., & Ojha, A. (2025, November 6). AI-based safety app helps bust immoral trafficking racket at Delhi spa, 6 rescued. India Today. India Today. https://www.indiatoday.in/cities/delhi/story/ai-based-safety- app-helps-bust-immoral-trafficking-racket-at-delhi-spa-6-rescued-2814618-2025-11-06
Mohanty, A. (2026). AI Governance in South Asia: Shared Priorities, Challenges & Future Trajectories [Report]. IIT Madras, CeRAI & WSAI IITM. https://cerai.iitm.ac.in/docs/AI-Governance-In-South-Asia.pdf
NIST. (2023). AI Risk Management Framework. https://doi.org/10.6028/NIST.AI.100-1. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. New York University Press.
Nova One Advisor. (2025). Artificial intelligence (AI) market share analysis report, 2023–2032. https://www.novaoneadvisor.com/report/artificial-intelligence-market
Nugraha, T., & Shinta, A. (2024, April 29). Ethical & Responsible AI in Developing Countries: Strategies for Progress. Modern Diplomacy. https://moderndiplomacy.eu/2024/04/29/ethical-responsible-ai-in- developing-countries-strategies-for-progress/
NYU Steinhardt. (2024). AI across sectors: Bridging technology and peacebuilding. https://steinhardt.nyu.edu/ihdsc/ai-across-sectors-bridging-technology-and-peacebuilding
OECD. (2024). AI Principles. https://www.oecd.org/en/topics/sub-issues/ai-principles.html
Okolo, C. T., Dell, N., & Vashistha, A. (2022). Making AI explainable in the Global South: A systematic review. In Proceedings of the 5th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies (COMPASS ’22) (pp. 439–452). Association for Computing Machinery. https://doi.org/10.1145/3530190.3534802
Organization for Security and Co-operation in Europe (OSCE). (2020). Leveraging innovation to fight trafficking in human beings: A comprehensive analysis of technology tools. OSCE Office of the Special Representative and Co-ordinator for Combating Trafficking in Human Beings & Tech Against Trafficking. https://resources.osce.org/sites/default/files/f/documents/9/6/455206_1.pdf
Ostrom, E. (2009). A polycentric approach for coping with climate change. Available at SSRN 1934353.
PolicyEdge. (2026). Lok Sabha: Legal safeguards against AI‑related harms. https://www.policyedge.in/p/lok- sabha-legal-safeguards-against-ai-related-harms
Press Information Bureau (PIB). (2025a). Dr. Jitendra Singh hails "BharatGen" as India’s first sovereign multilingual and multimodal AI driven Large Language Model. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2194204®=3&lang=2
Press Information Bureau (PIB). (2026, March 6). India’s Digital Public Infrastructure. https://www.pib.gov.in/www.pib.gov.in/Pressreleaseshare.aspx?PRID=2235812
Press Information Bureau. (2024). Government taking measures to ensure safe and trusted artificial intelligence. https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=2036440®=3&lang=2
Press Information Bureau. (2024a). India commits to safe and trusted AI with new EoI initiative. https://www.pib.gov.in/PressReleaseDetail.aspx?PRID=2065579®=3&lang=1
Press Information Bureau. (2025). India’s AI revolution. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2108810®=3&lang=2
Press Information Bureau. (2026a). Legal safeguards in place to prevent potential harms that may arise from AI and related technologies. https://www.pib.gov.in/PressReleaseDetail.aspx?PRID=2238191®=3&lang=2
Rahman, H. A., Karunakaran, A., & Cameron, L. D. (2024). Accountability into account in the management of platforms. Academy of Management Annals, 18(1), 251–294. https://doi.org/10.5465/annals.2022.0090
Rami, F., Searight, H., Dryjanska, L., & Hall, M. (2025). Modern Slavery and the United Nations Sustainable Development Goals. Canadian Psychology, 66(4), 325-339.
Reeves, J., Soutar, E., Green, S., & Crowther, T. (2018). Children and Young People’s Vulnerabilities to Grooming. In Ş. Çetinkaya (Ed.), Contemporary Perspective on Child Psychology and Education. InTech. https://doi.org/10.5772/intechopen.71672
Rizwan, H. (2025, December 15). Inside The App That Sends Police Live Video The Moment Someone Says “Help.” https://www.decodeinternet.in/decode/inside-the-app-that-sends-police-live-video-the- moment-someone-says-help-30260
Roberts, S. T. (2019). Behind the screen: Content moderation in the shadows of social media. Yale University Press.
Sabo, T. (2019). An Artificial Intelligence Framework on SAS Viya to Counter International Human Trafficking. SAS Global Forum, Dallas, Texas.
Safe Online. (2023). Learnings, partnership and collaboration for a safe digital future. https://safeonline.global/learnings-partnership-and-collaboration-for-a-safe-digital-future/
Sawer, M., Banaszak, L. A., True, J., & Kantola, J. (Eds.). (2023). Handbook of feminist governance. Edward Elgar Publishing. https://doi.org/10.4337/9781800374812
Scherrer, C., Garcia, A., & Wullweber, J. (Eds.). (2023). Handbook on critical political economy and public policy. Edward Elgar Publishing. doi.org
Stephan, M., Marshall, G. R., & McGinnis, M. (2019). An introduction to polycentricity and governance. In A. Thiel, W. Blomquist, & D. E. Garrick (Eds.), Governing complexity: Analysing and applying polycentricity (pp. 21–44). Cambridge University Press. doi.org
Sustainability Directory. (2025, November 28). Critical Political Economy. Pollution Sustainability Directory. https://pollution.sustainability-directory.com/term/critical-political-economy/
Suzor, N. P. (2019). Lawless: The secret rules that govern our digital lives. Cambridge University Press. https://www.cambridge.org/core/books/abs/lawless/contents/F7BBE591DAFD294EA28659EBD769FCBD
Tech Global Institute. (2023, May 20). South Asia is Split on the Race to Regulate AI. https://techglobalinstitute.com/announcements/blog/digital-freedom-network-is-a-coalition-to-protect- and-promote-fundamental-rights/
The Impact of Artificial Intelligence on Trafficking in Persons. (2026). U.S. Department of State. https://www.state.gov/wp-content/uploads/2026/04/The-Impact-of-Artificial-Intelligence-Accessible- 4.16.2026.pdf
The Print. (2025). NCPCR rescued 2,300 children across India in six months: https://theprint.in/india/ncpcr- rescued-2300-children-across-india-in-six-months-official/2788232/
Thiel, A. (2023). Polycentric Governing and Polycentric Governance. In Frank Gadinger, and Jan Aart Scholte (eds), Polycentrism: How Governing Works Today (Oxford, 2023; online edn, Oxford Academic, 22 June 2023). https://doi.org/10.1093/oso/9780192866837.003.0005
Thorn. (2024). Spotlight and Solis: Technology to find children faster and deter predators.
Tiwari, A. (2025). Smart solutions to dark crimes: Exploring AI’s potential in preventing human trafficking. Indian Journal of Integrated Research in Law, 5(2), 1815–1825. https://ijirl.com/wp- content/uploads/2025/04/SMART-SOLUTIONS-TO-DARK-CRIMES-EXPLORING-AIS-POTENTIAL-IN- PREVENTING-HUMAN-TRAFFICKING.pdf
UNCTAD. (2025). AI market projected to hit $4.8 trillion by 2033, emerging as a dominant frontier technology. https://unctad.org
UNESCO. (2021, November 23). Recommendations on the ethics of artificial intelligence. UNESCO. SHS/2023/PI/H/1
UNICEF Innocenti. (2025). Guidance on AI and children: Recommendations for AI policies and systems that uphold child rights (Version 3.0). https://www.unicef.org/innocenti/reports/policy-guidance-ai-children
UNICEF Innocenti. (2026). Growing with Rights: Understanding and supporting the evolving capacities of the child. https://www.unicef.org/innocenti/media/12651/file/UNICEF-Innocenti-Growing-Rights-report- 2026.pdf
UNICEF. (2026). Need for safe, inclusive and empowering AI framework for children. https://www.unicef.org/india/press-releases/need-safe-inclusive-and-empowering-ai-framework- children
United Nations Convention against Transnational Organized Crime, November 15, 2000, https://treaties.un.org/pages/viewdetails.aspx?src=treaty&mtdsg_no=xviii-12&chapter=18&clang=en
United Nations Office on Drugs and Crime (UNODC). (2024). Global report on trafficking in persons 2024 (p. 176). United Nations. https://www.unodc.org/documents/data-and-analysis/glotip/2024/GLOTIP2024_BOOK.pdf
United Nations. (2019). What is Human Trafficking?. Human Rights and Human Trafficking Fact Sheet no. 36 (p. 3). New York and Geneva. https://www.unodc.org/unodc/en/human-trafficking/what-is-human- trafficking.html
UNODC. (2018). Global Report on Trafficking in Persons 2018. (United Nations publication). https://www.unodc.org/unodc/en/data-and analysis/glotip.html
UNODC. (2019). Trafficking in Persons & Smuggling of Migrants: Technology Facilitating Trafficking in Persons. : https://www.unodc.org/cld/en/education/tertiary/tip-and-som/module-14/key-issues/technology- facilitating-trafficking-in-persons.html
Vanberghen, C. (2024, February 24). Deepfake Dilemma: Balancing Regulation, Responsibility, and Rights in the AI Era. Modern Diplomacy. https://moderndiplomacy.eu/2024/02/24/deepfake-dilemma-balancing- regulation-responsibility-and-rights-in-the-ai-era/
You, J., Somala, V., Edelman, Y., & Emberson, L. (2026). Global AI computing capacity is doubling every 7 months. Epoch AI. https://epoch.ai/data-insights/ai-chip-production
Zuboff, S. (2019). The age of surveillance capitalism: The fight for a human future at the new frontier of power. PublicAffairs. https://www.hbs.edu/faculty/Pages/item.aspx?num=56791
