Taxonomy & Glossary

This page defines the classification framework used across the AI Hub tracker. It draws on the DCI AI Hub Taxonomy (v2, March 2026), integrating the AI Hub Definitional Framework and discussions across partners during the Taxonomy workshop.

The taxonomy provides a structured vocabulary for documenting AI use cases in social protection, ensuring consistent classification across countries, programmes, and implementation types. The full taxonomy document, including detailed mapping matrices between AI capabilities, use case categories, and social protection functions, will be available for download once finalised.

1.1 AI Capabilities

Core AI model capabilities — the main technical functions and tasks the model performs on the data. The taxonomy broadly distinguishes between analytical AI tasks (which analyse existing data to produce a score, label, grouping or flag) and generative tasks (which produce new content such as text, synthetic data or translations).

Analytical AI Tasks

Prediction (including forecasting) 34 cases

Models that estimate outcomes based on patterns learned from historical data. Prediction models output a numeric value or probability that quantifies an outcome of interest (e.g. expected cost, predicted demand, risk score). Forecasting is a subtype where the value is explicitly tied to a future time horizon.

In social protection, prediction outputs typically serve as advisory inputs — informing human judgment, programme design, or downstream analytical processes rather than directly determining entitlements or access to services.

Classification 66 cases

Produces a category or label (e.g. eligible/not eligible, priority tier), often with associated probabilities that can be thresholded into decisions. In practice, the same modelling approaches as for prediction are frequently used — the distinction lies in the form of the output and governance implications.

In social protection, assigning a categorical label — particularly where it gates access to entitlements or services — typically carries higher decision criticality than producing a numeric estimate, and may require stronger human oversight, auditability of threshold-setting, and access to appeal.

Clustering (similarity and grouping) 10 cases

Organizes data based on how alike items are and how they naturally form coherent segments. Includes similarity scoring and nearest-neighbour matching (often via embeddings) as well as unsupervised clustering that produces group assignments and profiles.

In social protection, clustering supports population segmentation for programme design, deduplication of beneficiary records across registries, and matching individuals to similar profiles for case management or service referral.

Anomaly and Change Detection 30 cases

Models that identify rare, unusual, or suspicious observations that deviate from expected patterns, often where labelled examples of problematic cases are scarce or constantly evolving. Outputs are anomaly scores or alerts that prioritise items for human review.

In social protection, anomaly detection is widely used for fraud and error detection in payment systems, flagging irregular transaction patterns or duplicate claims for human review, as well as monitoring data quality across registries.

Perception and Extraction from Unstructured Inputs 53 cases

Models that transform unstructured data — text, images, audio, or video — into structured observations. Outputs include extracted fields from documents (OCR), entities and relations from text, objects/regions in images, and transcripts from speech.

In social protection, extraction capabilities enable processing of identity documents, application forms and case notes into structured data that can feed eligibility assessments, grievance triage or compliance checks.

Ranking and Decision Systems 25 cases

Produce a prioritized list of options or a selected action/policy designed to achieve an objective under constraints. Includes learning-to-rank and recommender approaches as well as optimization-oriented decision models. While these may use predictions or classifications as inputs, they are distinct because their primary output is what to do (or what to do first).

In social protection, ranking and decision systems support case prioritisation across large caseloads, resource allocation across geographic areas, and recommendation of tailored service or benefit packages.

Generative AI Tasks

LLMs for Content Creation, Transformation and Modality Conversion 20 cases

Intra-modal generation: Models that produce or edit content within the same modality — drafting, rewriting, summarization, translation, and correction (text-to-text), or image/audio enhancement and modification.

Cross-modal generation: Models that express content in a different modality from the input — text-to-speech, text-to-image, image captioning, and related cross-modal transformations.

In social protection, generative capabilities support beneficiary-facing chatbots and voicebots, multilingual translation of programme information, drafting of decision notices and case summaries, and accessibility tools such as text-to-speech.

Synthetic Dataset Generation 2 cases

Models that produce datasets or collections of records designed to approximate relevant statistical properties of real data. The goal is typically utility for testing, simulation, augmentation, or privacy-aware sharing.

In social protection, synthetic data generation supports privacy-preserving sharing of beneficiary data for research, stress-testing of eligibility models under varied scenarios, and augmentation of training datasets where real data is scarce.

1.2 Social Protection AI Use Case Categories

An organic clustering of the key roles AI can play in social protection functions. These categories describe what AI does in generic terms — each has different practical applications depending on the phase of the delivery chain and the type of programme. Most use cases combine several AI capabilities from Section 1.1 rather than being limited to a single one. To see how these use cases map onto specific social protection system functions at each level, see Section 1.3 below.

Vulnerability, Needs and Risk Assessment (including Predictive Analytics) 32 cases

Identifies who and where support is most needed by estimating poverty, vulnerability, or risk levels and segmenting populations to guide policy, design and operational decisions.

Examples: Geospatial poverty mapping, predictive vulnerability scoring, non-take-up detection, outreach segmentation and optimisation.

Indirectly affects rights and entitlements. Governance sensitivity escalates where vulnerability scores are used as direct inputs to individual eligibility determination.

Trend and Shock Forecasting 7 cases

Anticipates future trends, shocks or crises to enable early planning and preventive or adaptive measures, including anticipatory action. Forecasting is about predicting what will happen while vulnerability assessment estimates who is or will be affected.

Examples: Climate and hazard forecasting/EWS, conflict and displacement early warning, demographic and socio-economic trend prediction.

Primarily informational. Governance sensitivity escalates where forecasts directly trigger anticipatory transfers or automatic benefit adjustments.

Identification, Verification and Record Linkage 13 cases

Confirms identity, validates claims or credentials, and links records across fragmented systems to create a unified and reliable view of individuals or households.

Examples: Biometric verification, deduplication, automated evidence validation, cross-registry linkage, entity resolution.

May directly affect rights and entitlements. Governance sensitivity is highest for biometric systems, given documented cases of exclusion among elderly, disabled and manual-labour populations.

Decision Support for Eligibility and Benefits 14 cases

Assesses whether individuals or households qualify for programmes and what level or type of support they should receive, through rule-based or model-assisted decision support. Only encompasses cases where AI directly supports the eligibility determination.

Examples: Scoring, benefit package configuration, service tailoring, enrolment decision support.

Directly affects rights and entitlements. This is the highest-criticality use case category. Requires strongest human oversight (HITL), auditability of threshold-setting, and accessible appeal mechanisms.

Data Quality and Anomaly Detection 12 cases

Prepares, validates and enhances data, while detecting unusual or suspicious patterns, prioritising cases for data quality checks or investigation to protect programme integrity. Focuses on data and system integrity.

Examples: Transaction anomaly detection, data consistency checks, predictive integrity risk scoring, network/collusion analytics, audit prioritisation.

Governance sensitivity is high where anomaly scores trigger automatic actions such as payment suspension. Documented cases of harmful outcomes (e.g. Dutch childcare benefits scandal) underscore the need for human review before adverse action.

User Communication and Interaction 30 cases

Facilitates two-way information exchange with beneficiaries/applicants or staff by providing guidance, explanations, and personalised messages through automated or assisted channels.

Examples: Chatbots/voicebots, translation, explainable decision notices, personalised messaging.

Low decision criticality but high visibility. Governance sensitivity increases where automated communication replaces rather than supplements human service channels.

Matching and Recommendation 12 cases

Connects individuals to suitable benefits, services, support opportunities, jobs, or training by analysing profiles, preferences, and relevant data sources.

Examples: Case management, job matching, skills profiling, career pathway guidance, programme recommendation.

Moderate criticality. Governance sensitivity escalates where matching outputs directly determine programme referrals or service access without caseworker review.

Operational and Process Automation 23 cases

Streamlines administrative workflows through two distinct sub-categories:

(a) Document processing and generative staff assistance: Extracting, digitising and routing information from documents, and generating drafts, summaries or case file syntheses.

(b) Workload and resource forecasting: Predicting operational demand, caseload volumes and staffing requirements.

Primarily internal and operational. Governance sensitivity escalates where automated workflow routing affects processing times in ways that materially delay access to benefits.

Compliance and Integrity 29 cases

Monitors adherence to programme rules, requirements and conditions, manages complaints and appeals. Focuses on beneficiary and programme-rule adherence (vs. data quality which focuses on data/system integrity).

Examples: Conditionality verification, behavioural compliance analytics, complaint triage and pattern detection.

May directly affect rights and entitlements. Documented cases of harmful outcomes (e.g. Robodebt in Australia). Requires human review before any adverse action affecting entitlements.

Policy Analysis, Learning and M&E 6 cases

Generates insights to improve policy and programme design and implementation, improving system performance over time.

Examples: Policy microsimulation, counterfactual analysis, model performance and bias/fairness monitoring, trend monitoring, improved analytics/dashboards.

Primarily internal and operational. Governance sensitivity increases where bias and fairness evaluations reveal systemic discrimination, requiring transparent reporting and corrective action.

1.3 Social Protection System Functions

While Section 1.2 describes what AI does in generic terms (e.g. "vulnerability assessment", "decision support"), this section describes where in the social protection system those AI use cases are applied. The same use case category can serve different functions depending on the level — for example, "vulnerability assessment" supports policy-level needs analysis at the policy level, eligibility criteria design at the programme design level, and individual case prioritisation at the delivery chain level. The full taxonomy document (forthcoming) includes a detailed matrix mapping every use case category against each system function.

Policy

Legal and policy frameworks — Strategic vision, legal frameworks, policy coherence
Financing — Budget planning, fiscal sustainability, risk-informed financing
Coordination and governance + Technical and functional capacities — Inter-agency coordination, staff capacity building, governance oversight

Programme Design

Benefits and service package — Designing optimal combinations of benefits and services
Eligibility criteria and qualifying conditions — Defining and calibrating eligibility thresholds
Level, value, frequency and duration of support — Setting benefit levels and payment schedules

Implementation / Delivery Chain

Outreach/communications/sensitisation — Reaching underserved populations, targeted communication
Registration — Data collection, deduplication, registration prioritisation
Management of contributions and withdrawals — Contribution processing, employment verification (social insurance)
Assessment of needs/conditions + enrolment — Identity verification, eligibility assessment, enrolment
Provision of payments/services — Payment delivery, KYC compliance, service delivery
Accountability mechanisms — Grievance redressal, complaint triage, appeals
Case management — Caseload prioritisation, service coordination, compliance monitoring
Profiling, job matching and support services — Skills assessment, job matching, career guidance (labour market)
Monitoring and evaluation — Performance analytics, data quality monitoring, bias detection

2.1 AI Lifecycle Stages

Adapted from established governance frameworks (OECD, 2019; NIST, 2023), recognising that risks accumulate over time rather than appearing at a single stage.

Stage	Description
Problem Identification and Design	Defining the problem the AI system is meant to solve, the intended users and context, and the success criteria, while considering feasibility, risks, and ethical implications from the outset.
Data Gathering and Preparation	Collecting, cleaning, labelling, and managing data needed to train the model, ensuring data quality, representativeness, and compliance with legal and ethical standards.
Model Selection and Training	Choosing appropriate algorithms and model architectures, then training and validating the model to optimize performance while addressing issues such as bias, overfitting, and explainability.
Integration and Deployment	Embedding the trained model into existing systems or workflows and releasing it into real-world use, including testing for reliability, security, and user interaction.
Monitoring, Maintenance, and Decommissioning	Continuously tracking performance and impacts over time, updating or retraining the model as conditions change, and safely retiring the system when it is no longer effective or appropriate.

2.2 Implementation Types

How the AI output is produced. This is independent from what the output does (capabilities). The same analytical task can be implemented via classical ML, deep learning, or a foundation model, with distinct implications for validation, compute and governance.

Type	Description	Common SP Applications
Classical ML	Statistical and ML methods including logistic regression, decision trees, random forests, gradient-boosted models. Trained on structured, tabular datasets with explicitly engineered features. Lightweight, high interpretability, strong reproducibility.	PMT/welfare estimation, eligibility scoring, fraud detection, workload forecasting
Deep learning	Neural network architectures including CNNs, RNNs/LSTMs, and task-specific transformers. Requires larger datasets and compute but can learn directly from unstructured inputs.	Satellite imagery analysis, biometric matching, document OCR, speech recognition
Foundation model	Large pre-trained models (LLMs, multimodal models) adapted via fine-tuning, prompting, or RAG. High compute requirements, adaptable to multiple tasks, risk of hallucination, prompt sensitivity.	Chatbots, document summarisation, translation, grievance triage, synthetic data generation
Hybrid	Systems combining multiple types in a pipeline — e.g. a foundation model for extraction feeding a classical ML model for scoring. Governance profile determined by the highest-risk component.	Extract-and-score pipelines, RAG with rule-based guardrails, predictive scoring with generated explanations

2.3 Agentic AI

Agentic AI refers to systems that autonomously plan and execute multi-step workflows, selecting tools, making intermediate decisions, and chaining actions to achieve a goal with limited human intervention. It is not a separate capability category — rather, it is an architectural pattern that orchestrates multiple capabilities within a single automated pipeline.

As agentic systems become more prevalent, they raise distinct governance considerations around human oversight, accountability for intermediate decisions, and the difficulty of auditing multi-step autonomous processes. The tracker flags agentic characteristics as an attribute of documented use cases.

Field	Values
Agentic system	Yes / No / Partial (human checkpoints within pipeline)
Degree of autonomy	Fully autonomous (end-to-end) / Semi-autonomous (human checkpoint at defined stages) / Supervised (human approval at each step)
Human override points	Where in the pipeline human review is triggered

2.4 Decision Criticality

Determines the required level of human oversight based on the rights impact of the decision the system supports.

Level	Description	Required Oversight
High	Errors may directly affect rights, livelihoods, or entitlements. Examples include eligibility recommendation or payment-suspension triggers.	Human-in-the-Loop (HITL)
Moderate	Affects service quality but not entitlements. Errors may delay services but are generally reversible.	Human-on-the-Loop (HOTL)
Low	Supports internal administration with minimal rights impact. Examples include workload forecasting or automated form processing.	Human-out-of-the-Loop (HOOTL)

2.5 Human Oversight Requirements

Three levels of human involvement in AI-assisted decision-making.

Type	What It Means	Description
HITL Human-in-the-Loop	Active human review before outcomes	Officers must confirm or override AI recommendations before rights-affecting decisions are finalised. Essential for eligibility decisions and benefit adjustments.
HOTL Human-on-the-Loop	Human monitoring of performance	Staff observe the tool during operation, intervene when required, and escalate complex cases. Suitable for grievance routing or forecasting tools.
HOOTL Human-out-of-the-Loop	Autonomous with periodic audit	Routine administrative tasks may operate autonomously where the rights impact is low and the task is reversible. Examples: document sorting, OCR processing.

2.6 Development Process

Type	Description
Fully in-house	AI systems designed, built, and maintained entirely by government entities.
Mix of in-house and third-party	AI systems co-developed through collaboration between government teams and external vendors or partners.
Fully third-party developed	AI systems fully designed and delivered by external providers under government contract.

2.7 Deployment Status

Deployment status is distinct from lifecycle stages. Lifecycle stages describe the technical development process; deployment status captures how far the system has progressed into real-world operational use and institutional embedding. Both are recorded independently.

Status	Description
Concept / Exploration Phase	Problem definition, feasibility assessment, stakeholder consultation, or initial scoping underway. No functional system yet.
Design & Development Phase	Model design, data preparation, and system development in progress. Internal testing may be underway but not yet used in live operations.
Pilot / Controlled Trial Phase	Limited deployment in a defined setting (e.g. specific region, department, or use case). Performance, risks, and operational viability being evaluated.
Operational Deployment (Limited Rollout)	Actively used in real workflows but not yet scaled across the full intended scope. May still include monitoring and refinement.
Full Production Deployment	Fully integrated into government operations at intended scale. Embedded in business processes with governance and maintenance structures in place.
Scaled & Institutionalised	System embedded in policy frameworks, budgets, and long-term operational planning. Includes formal oversight, auditing, and lifecycle management.
Suspended / Halted	System was previously deployed but has been suspended due to identified issues, legal challenges, or policy decisions.

2.8 Data Sovereignty & Infrastructure

Classification of data hosting, cross-border transfer, and compute sovereignty.

Data Residency

Domestic	Data stored within the country
Regional	Data stored within a specified regional jurisdiction
International	Data stored in a foreign jurisdiction

Cross-Border Data Transfer

None	No cross-border data transfer identified
With documented safeguards	Transfer occurs with documented legal or technical safeguards
Without documented safeguards	Transfer occurs without documented protections

Sovereignty Quadrant

I — Sovereign AI Zone	Full domestic control over data, compute, and model governance
II — Federated/Hybrid Governance	Shared governance across domestic and regional/international entities
III — Cloud with Safeguards	International compute-intensive cloud deployment with documented governance safeguards
IV — Shared Innovation Zone	International collaboration with limited domestic governance controls

3.1 Structural Risk Categories

Five structural sources of risk that accumulate across the AI lifecycle. Failures in one domain frequently amplify weaknesses in another. Risk assessment must consider all five sources in combination.

Data-related risks

Risks arising when the information used to train, validate or operate an AI system does not accurately or fairly reflect the real world.

Representation bias Data quality failure Cross-dataset inconsistency Data or concept drift Weak provenance or lineage Consent or lawful basis gap

Model-related risks

Risks originating in the architecture, objectives and internal behaviour of the algorithm itself. Particularly consequential where outputs directly influence eligibility, sanctions or payment decisions.

Opacity or limited explainability Objective misalignment Model misspecification Shortcut learning and proxy reliance Subgroup bias Reliability or generalisation failure Hallucination or misinformation Model security vulnerability Behavioural drift

Operational and system integration risks

Risks arising when an AI model is deployed within real infrastructures, workflows and data pipelines. Compounded by staff tendency to defer to algorithmic outputs without critical scrutiny.

Legacy system integration failure Pipeline fragility Threshold or rule misconfiguration Monitoring gap Inadequate real-world validation Scalability or latency constraint Automation complacency

Governance and institutional oversight risks

Risks emerging when institutions lack accountability, capacity, oversight mechanisms, legal safeguards or meaningful control over AI systems.

Unclear accountability Insufficient human oversight Weak monitoring ownership Weak documentation or auditability Inadequate grievance or redress Regulatory non-compliance Purpose limitation failure Insufficient institutional capacity Inadequate resourcing

Market, sovereignty and industry structure risks

Risks arising from external dependencies on vendors, infrastructure providers and global technology ecosystems.

Vendor lock-in Interoperability constraint Jurisdictional hosting risk Restricted audit access Upstream model or API dependency Absent exit or migration pathway Market concentration Opaque supply chain LMIC power asymmetry Systemic or cascading fragility

3.2 Human & Societal Impact Dimensions

How structural risks materialise in people's lives and in institutional outcomes. Organised around four guiding principles established in the Risk Framework.

Privacy and Data Security

Harms arising from excessive data collection, unlawful linkage, breaches of sensitive information, or loss of individual control over personal data.

Privacy violation or data breach Disproportionate surveillance or profiling Loss of individual control over personal data

Equality, Non-Discrimination, Fairness and Inclusion

Discriminatory outcomes, systematic exclusion of marginalised groups, disparate error rates, and reinforcement of structural inequities.

Discriminatory outcome Systematic exclusion from benefits or services Disparate error rates across groups Reinforcement of structural inequity

Autonomy, Human Dignity and Due Process

Loss of agency, opaque or unexplained decisions, inability to contest outcomes, erosion of procedural justice, and psychological stress from intrusive profiling.

Loss of individual agency or autonomy Opaque or unexplained decision Inability to contest or appeal outcome Psychological stress, stigma or dignity harm

Accountability, Transparency and Redress

Absence of named decision owner, missing audit trails, inaccessible grievance mechanisms, inability to trace or correct AI-assisted decisions.

No identifiable decision owner Untraceable decision pathway No accessible or effective remedy

Systemic and Societal

Impacts that erode public trust, reinforce structural inequities, trigger political backlash, increase administrative burden on frontline staff, and deepen digital divides.

Erosion of public trust in SP system Political backlash, litigation or controversy Increased administrative burden on frontline staff Deepened digital divide

3.3 Safeguards

Documented safeguard measures that may be in place for a given AI deployment.

DPIA/AIA conducted	A Data Protection Impact Assessment or AI Impact Assessment has been formally conducted
Human oversight protocol	Documented procedures for human review, intervention, and override of AI outputs
Grievance mechanism	Accessible channels for affected individuals to challenge or appeal AI-influenced decisions
Bias audit	Systematic evaluation of AI outputs for discriminatory patterns across population subgroups
Independent evaluation	External review of system performance, fairness, and compliance by an independent party
Data minimisation controls	Technical and procedural measures limiting data collection and retention to what is necessary
Exit/rollback plan	Documented plan for decommissioning the AI system or reverting to non-AI processes if needed

4.1 Programme Types

Where use cases relate to specific programmes, they are classified under one of three social protection pillars.

Social Assistance

Emergency Cash Transfers
Poverty targeted Cash Transfers (conditional or unconditional)
Child grants/benefits (universal or targeted)
Non-contributory Disability Grants
Non-contributory Social Pensions
In-Kind Transfers
Public Works Programs
School Feeding Programs
Fee waivers and targeted subsidies

Social Insurance

Old age, survivors and disability pensions
Unemployment Insurance
Health Insurance
Work injury and occupational insurance
Maternity and paternity benefits (contributory)

Labour Market Programmes

Skilling and Training Programs
Entrepreneurship Support
Intermediation Services
Employment Incentives / Wage Subsidies
Job search assistance and placement services
Public employment services
Productive inclusion programmes

Where a programme combines elements across pillars (e.g. a conditional cash transfer with a graduation/skills training component), one primary pillar is assigned based on the programme's core function, with the secondary pillar recorded as a linkage.

5. Glossary of Terms

Key terms and definitions used across the AI Hub. These are maintained by the editorial team and can be updated via the admin panel.

AI Techniques

Computer Vision

AI field enabling computers to interpret and understand visual information from images or video.

Computer vision applications in social protection:

Identity verification: Facial recognition, fingerprint matching
Document verification: Reading IDs, scanning forms
Asset verification: Analyzing satellite imagery for housing/land
Biometric enrollment: Capturing and processing biometric data

Raises significant privacy and consent considerations.

Also known as: CV, Image Recognition

Related: deep-learning, facial-recognition, biometric

Deep Learning

A subset of machine learning using artificial neural networks with multiple layers to model complex patterns in data.

Deep learning excels at processing unstructured data like images, text, and speech. In social protection contexts, it powers:

Document verification (reading IDs, forms)
Facial recognition for identity verification
Natural language processing for chatbots
Satellite imagery analysis for poverty mapping

Deep learning typically requires large amounts of training data and computational resources.

Also known as: DL, Deep Neural Networks

Related: machine-learning, neural-network, computer-vision, nlp

Gradient Boosting

An ensemble machine learning technique that builds models sequentially, with each new model correcting errors from previous ones.

Popular implementations include XGBoost, LightGBM, and CatBoost. Used in social protection for:

High-accuracy poverty prediction
Fraud detection
Risk scoring

Often achieves state-of-the-art performance on tabular data like household surveys.

Also known as: GBM, Gradient Boosted Trees

Related: machine-learning, random-forest, xgboost

Machine Learning

A subset of artificial intelligence that enables systems to learn and improve from experience without being explicitly programmed.

Machine learning algorithms build mathematical models based on training data to make predictions or decisions. In social protection, ML is commonly used for poverty prediction, fraud detection, and beneficiary targeting.

Types of machine learning:

Supervised learning: Learning from labeled examples (e.g., predicting eligibility based on historical decisions)
Unsupervised learning: Finding patterns in unlabeled data (e.g., clustering similar beneficiary profiles)
Reinforcement learning: Learning through trial and error (less common in social protection)

Also known as: ML, Statistical Learning

Related: deep-learning, neural-network, random-forest, supervised-learning, gradient-boosting, proxy-means-test

Natural Language Processing

AI techniques for understanding, interpreting, and generating human language.

NLP enables machines to process text and speech. Applications in social protection include:

Chatbots: Answering beneficiary queries 24/7
Document processing: Extracting information from forms and applications
Sentiment analysis: Analyzing feedback and complaints
Translation: Making services accessible in multiple languages

Modern NLP is powered by large language models (LLMs) and transformer architectures.

Also known as: NLP, Computational Linguistics

Related: chatbot, large-language-model, deep-learning

Neural Network

A computing system inspired by biological neural networks, consisting of interconnected nodes that process information.

Neural networks form the foundation of deep learning. They consist of:

Input layer: Receives raw data
Hidden layers: Process and transform data
Output layer: Produces predictions

Common architectures in social protection:

Feedforward networks for tabular data prediction
Convolutional Neural Networks (CNNs) for image processing
Recurrent Neural Networks (RNNs) for sequential data

Also known as: ANN, Artificial Neural Network

Related: deep-learning, machine-learning

Random Forest

An ensemble machine learning method that constructs multiple decision trees and outputs the average prediction.

Random forests are popular in social protection applications because they:

Handle mixed data types well (common in household surveys)
Are relatively interpretable compared to deep learning
Resist overfitting
Provide feature importance rankings

Commonly used for Proxy Means Test (PMT) models and eligibility scoring.

Also known as: RF, Random Decision Forest

Related: machine-learning, decision-tree, gradient-boosting

Frameworks

Delivery Chain

The sequence of functions involved in providing social protection benefits to individuals, from assessment through payment.

The DCI framework identifies four main delivery functions:

1. Assess: Determining eligibility and needs 2. Enrol: Registration and enrollment 3. Provide: Benefit and service delivery 4. Manage: Program oversight and management

Each function can involve AI at different levels of automation and decision-making.

Related: assess, enrol, provide, manage, l0-delivery-function

L0 Delivery Function

The highest level of the DCI classification for AI use cases in social protection.

L0 represents the four core functions in the delivery chain:

Assess: Where AI helps determine who is eligible
Enrol: Where AI supports registration processes
Provide: Where AI optimizes benefit delivery
Manage: Where AI aids program oversight

Each L0 function has associated L1 clusters and L2 specific typologies.

Related: l1-functional-cluster, l2-typology, delivery-chain

L1 Functional Cluster

The second level of the DCI classification, grouping similar AI applications within each delivery function.

L1 provides more specificity within each L0 function. For example, within "Assess":

Eligibility determination
Needs assessment
Risk profiling

L1 clusters help organize the diversity of AI applications within each broad function.

Related: l0-delivery-function, l2-typology

L2 Typology Element

The most specific level of the DCI classification, describing particular AI application types.

L2 provides detailed categorization. Each L2 element includes:

Unique code (e.g., 1.1.1)
Descriptive title
Definition
Examples

L2 elements enable precise comparison of AI applications across countries and programs.

Related: l1-functional-cluster, l0-delivery-function

Governance & Ethics

Data Protection

Legal and technical measures to safeguard personal information collected and processed by AI systems.

Key principles in social protection contexts:

Purpose limitation: Data used only for stated purposes
Data minimization: Collect only necessary information
Storage limitation: Retain only as long as needed
Security: Protect against unauthorized access
Consent: Informed agreement where applicable

Relevant frameworks: GDPR, national data protection laws, sector-specific regulations

Related: privacy, gdpr, data-minimization

Explainability

The ability to understand and communicate how an AI system makes decisions.

Explainability is crucial for:

Accountability: Understanding why a decision was made
Appeals: Beneficiaries challenging decisions
Trust: Building confidence in automated systems
Debugging: Identifying and fixing errors

Levels of explainability:

Global: Overall model behavior
Local: Individual decision explanations

Techniques: SHAP values, LIME, feature importance, decision rules

Related: transparency, interpretability, black-box

Human-in-the-Loop

An AI oversight model where humans review and approve every decision before it takes effect.

In HITL systems, AI provides recommendations or preliminary decisions, but a human must confirm before action. This provides maximum oversight but may reduce efficiency gains from automation.

Best suited for:

High-stakes decisions affecting eligibility
Cases where AI confidence is low
Initial deployment phases

Trade-offs:

Higher accuracy through human judgment
Slower processing
Requires trained human reviewers

Related: human-on-the-loop, human-out-of-the-loop, ai-oversight

Human-on-the-Loop

An AI oversight model where humans monitor AI operations and can intervene when needed, but do not review every decision.

HOTL allows AI to operate autonomously for routine cases while humans focus on exceptions, edge cases, and system monitoring.

Implementation approaches:

Exception-based review (AI flags uncertain cases)
Random sampling for quality assurance
Dashboard monitoring of aggregate outcomes

Best suited for:

High-volume, lower-stakes decisions
Well-tested AI systems with proven accuracy

Related: human-in-the-loop, human-out-of-the-loop, ai-oversight

Human-out-of-the-Loop

An AI oversight model where systems operate autonomously with only periodic human audits.

HOOTL is appropriate only when AI decisions are low-risk, reversible, or have strong safeguards. In social protection, this is rare for eligibility decisions but may apply to:

Automated payment scheduling
Routine data validation
Low-risk process automation

Requirements:

Comprehensive testing before deployment
Regular algorithmic audits
Clear accountability frameworks
Robust appeals processes

Related: human-in-the-loop, human-on-the-loop, ai-oversight

Methodology

Exclusion Error

When eligible individuals or households do not receive benefits they should receive.

Also called "undercoverage" or "Type II error" in targeting.

Causes:

Overly strict eligibility criteria
Barriers to enrollment (documentation, access)
Model bias against certain groups
Lack of awareness

More concerning than inclusion errors as they deny support to those in need. AI systems must be carefully designed and audited to minimize exclusion errors.

Related: inclusion-error, targeting, algorithmic-bias

Inclusion Error

When non-eligible individuals or households receive benefits they should not receive.

Also called "leakage" or "Type I error" in targeting.

Causes:

Inaccurate targeting models
Data manipulation or fraud
Outdated eligibility information
Administrative errors

AI can help reduce inclusion errors through:

Better prediction models
Fraud detection
Real-time data verification
Cross-referencing multiple data sources

Related: exclusion-error, targeting, fraud-detection

Proxy Means Test

A statistical method using observable household characteristics to estimate welfare levels for targeting social programs.

PMT uses easily verifiable indicators (housing quality, assets, education) as proxies for income or consumption, which are harder to measure. AI/ML enhances traditional PMT by:

Handling more variables simultaneously
Finding non-linear relationships
Updating models as new data becomes available

Strengths: Cost-effective, reduces gaming Weaknesses: Exclusion errors, may miss transient poverty

Related concepts: targeting, means testing, poverty prediction

Also known as: PMT, Proxy Means Testing

Related: targeting, machine-learning, poverty-prediction

Targeting

The process of identifying and selecting individuals or households eligible for social protection benefits.

Targeting methods:

1. Means testing: Direct income/asset verification 2. Proxy means testing: Using observable characteristics 3. Categorical: Based on demographics (age, disability) 4. Geographic: Based on location 5. Community-based: Local committees select beneficiaries 6. Self-targeting: Program design attracts only intended beneficiaries

AI primarily enhances means testing and PMT approaches through improved prediction models.

Related: proxy-means-test, eligibility, inclusion-error, exclusion-error, social-protection, social-assistance, beneficiary-registry

Risk

Algorithmic Bias

Systematic errors in AI systems that create unfair outcomes for certain groups.

Bias can enter AI systems through:

1. Training data bias: Historical discrimination encoded in data 2. Selection bias: Non-representative samples 3. Measurement bias: Proxies that correlate with protected characteristics 4. Aggregation bias: Models that work differently for subgroups

In social protection context:

PMT models may underpredict poverty in certain regions
Facial recognition may have higher error rates for some demographics
Language models may not serve minority language speakers equally

Mitigation strategies: Bias audits, fairness constraints, diverse training data, disaggregated testing

Related: fairness, discrimination, ai-ethics, exclusion-error

Social Protection

Beneficiary Registry

A database containing information on individuals and households receiving social protection benefits.

Modern registries (sometimes called Social Registries) serve as:

Single source of truth for beneficiary data
Platform for eligibility determination
Foundation for program coordination

AI applications:

Data quality checks and deduplication
Dynamic eligibility updates
Cross-program coordination
Fraud detection through pattern analysis

Related: social-registry, targeting, enrollment

Source: DCI AI Hub Taxonomy (working version, March 2026). Developed under the Digital Convergence Initiative, integrating the AI Hub Definitional Framework and partner input from the Taxonomy workshop. Risk dimensions from the AI Hub Risk Framework (Byrnes, Dobre & Van der Erve, 2026).