Stanford Graduate School of Business · Published 2026-10-06

World Development Report 2026: How Can AI Improve the Delivery of Public Services?

privateReview: unreviewedConfidence: high
artificial intelligencepublic servicesdeveloping economiessocial protectiongovernancepublic-sector-aiai-targetingwelfare-optimizationscreening-costObservabilityevaluation-gapdata-infrastructurelast-mile
Open on YouTube ↗

Summary

Overview

  • Speaker: Anja Sautmann and Joshua Blumenstock
  • Channel: Stanford Graduate School of Business
  • Main topic: How Artificial Intelligence can improve the delivery of public services in developing economies, highlighting use cases, front-end vs. back-end deployments, and governance challenges.
  • Purpose: To provide an overview of World Development Report 2026 findings on AI in developing economies, focusing on public service delivery, implementation challenges, and policy governance. This session explores how artificial intelligence can address pervasive human capacity and service delivery gaps in low- and middle-income countries (LMICs). Presenters Anja Sautmann and Joshua Blumenstock examine public sector AI applications across front-end (citizen-facing and frontline provider tools) and back-end (system-level forecasting, resource allocation, and oversight) operations. They discuss empirical evidence on health and education applications, data infrastructure bottlenecks, governance frameworks, and critical trade-offs between predictive accuracy, privacy, interpretability, and cost-effectiveness in targeted social programs.

Topic Map

Introduction to AI in Public Service Delivery

  • Explanation: Pervasive capacity gaps in low- and middle-income countries across health, education, and public administration, and how AI can potentially improve access to critical services for billions.
  • Key claims:
    • LMICs suffer from severe capacity gaps, such as having fewer than 50,000 radiologists for 3.7 billion people.
    • AI deployment divides into front-end (citizen/provider apps) and back-end (centralized agency and big data systems).
  • Examples:
    • Global total of radiologists by country income group showing severe shortages in LICs and LMICs.
  • Terminology:
    • LMIC
    • Front-end AI
    • Back-end AI
    • Capacity gaps
    • Last mile
  • Why it matters: Contextualizes the scale of resource scarcity in developing nations and establishes the framework for where and how AI can be deployed.

AI in the Front End of Government

  • Explanation: Examining front-end AI applications like chatbots, tutoring systems, and clinical decision support tools aimed at expanding access and increasing service quality.
  • Key claims:
    • Front-end use types fall into augmenting existing quality and expanding access through automation and task shifting.
    • While promising, many high-value applications lack rigorous evidence on cost-effectiveness and long-term human development outcomes.
    • Front-end deployments face severe last-mile constraints: connectivity, literacy barriers, device access, and language support.
  • Examples:
    • ConsiliumBots in Chile increasing college enrollment rates by up to 44% and higher-ranked program enrollment by 20%.
    • AI-assisted tuberculosis screening from chest X-rays in Nigeria and India.
  • Terminology:
    • Task shifting
    • LLM
    • Clinical decision support
    • High-intensity tutoring
    • Last-mile challenge
  • Why it matters: Highlights both the transformative potential and the hidden barriers (costs, risks, and equity concerns) of deploying user-facing AI tools in low-resource settings.

AI in the Back End of Government

  • Explanation: Analyzing back-end AI implementations in centralized agencies for forecasting, resource allocation, targeting, and monitoring.
  • Key claims:
    • Back-end AI is dominated by prediction and optimization tasks, especially in procurement contract analysis, tax administration, and environmental monitoring.
    • Improves forecasting without requiring extensive local sensing infrastructure (e.g., Google Floodhub).
    • Requires strong data infrastructure, secure access mechanisms, and clear standards, while carrying risks of misuse, manipulation, and cybercrime.
  • Examples:
    • Google Floodhub flood forecasts using past data rather than extensive in-river sensing infrastructure.
    • Supply chain management tools in Sierra Leone improving essential medicine distribution.
  • Terminology:
    • Predictive analytics
    • Resource management
    • Interoperability
    • Data infrastructure
    • Optimization
  • Why it matters: Demonstrates how back-end systems can scale solutions efficiently across populations but emphasizes the critical need for institutional capacity and data governance.

Governing AI in Government and Social Protection

  • Explanation: Best practices for governing AI in government, focusing on monitoring and evaluation, smart procurement, accountability frameworks, and navigating trade-offs in social protection targeting.
  • Key claims:
    • 81% of lower-income economies report basic or no performance monitoring for AI investments.
    • Better prediction does not equal better policy; trade-offs exist between model accuracy, interpretability, privacy, and community legitimacy.
    • When screening costs dominate at a national scale, simpler or alternative targeting methods can outperform complex ML models in welfare maximization.
  • Examples:
    • Cina Lawson and Togo's digital cash transfer program during COVID-19 using mobile metadata for geographic and individual targeting.
    • Comparing machine learning phone-based targeting against poverty maps and PMT surveys in Togo, Malawi, and Bangladesh.
  • Terminology:
    • Proxy means test (PMT)
    • Differential privacy
    • Interpretability
    • Exclusion errors
    • Targeting efficiency
  • Why it matters: Provides vital policy guidance on how governments can retain control over AI systems, avoid costly targeting pitfalls, and measure true welfare impacts.

Key Points

Front-End vs. Back-End AI Division

  • Explanation: Public sector AI applications divide into front-end tools used by citizens or frontline providers (facing last-mile constraints) and back-end tools used by centralized agencies for system-level analysis.
  • Evidence: WDR 2026 survey data on government AI usage across income groups and procurement contracts.
  • Practical implication: Governments must match deployment choices to their digital maturity and internal institutional capacities.

Evidence Gaps on Development Outcomes

  • Explanation: While many AI applications sound promising and technically mature, rigorous evidence on cost-effectiveness and human development outcomes remains scarce.
  • Evidence: Systematic literature reviews on AI in frontline healthcare and tutoring showing that only a tiny fraction of papers evaluate patient outcomes or cost-effectiveness in low-resource settings.
  • Practical implication: Policymakers should target investments in domains with consolidated evidence rather than adopting unproven technologies.

Better Prediction Does Not Equal Better Policy

  • Explanation: Advanced machine learning models often yield superior predictive accuracy, but when screening costs, administrative friction, and political constraints are factored in, simpler policies may be more effective.
  • Evidence: Case studies from Togo, Malawi, and Bangladesh demonstrating trade-offs between ML accuracy, proxy means test surveys, and community-driven targeting.
  • Practical implication: Evaluation must encompass the entire delivery chain, including implementation costs, exclusion errors, and welfare outcomes.

Frameworks, Models & Processes

Front-End vs. Back-End AI Framework

  • How it works: Categorizes public sector AI implementations based on user group and scale of analysis.
  • Components:
    • AI in the Front End (apps used by citizens or local providers for individual prediction, task shifting, and augmentation)
    • AI in the Back End (centralized agency tools using big data and machine learning for system-level forecasting, targeting, and oversight)
  • When to use: When designing digital transformation strategies for government agencies and assessing infrastructure requirements.

AI Governance and Evaluation Lifecycle

  • How it works: A structured approach to managing AI deployment in government from definition to outcome evaluation.
  • Components:
    • Defining objectives in context (setting goals and benchmarks)
    • Model validation against benchmarks
    • Managing intermediate outcomes and human-machine interactions (adoption, bias mitigation)
    • Evaluating broader development outcomes (welfare impacts, cost-effectiveness)
  • When to use: When procuring, auditing, and monitoring government AI projects.

Examples & Case Studies

Togo's Ministry of Digital Economy deployed mobile phone metadata and machine learning during the COVID-19 pandemic to identify poor households for cash transfers.

  • Illustrates: How auxiliary data can fill information gaps and improve geographic and individual targeting under severe constraints.
  • Lesson: Better predictive models surface critical trade-offs regarding privacy, interpretability, and administrative exclusion errors.

ConsiliumBots in Chile used AI-driven chatbots to provide custom guidance to high school students regarding college applications.

  • Illustrates: Front-end AI augmentation expanding access and improving service delivery.
  • Lesson: Increased college attendance rates by up to 44% and higher-ranked program enrollment by 20%.

Actionable Takeaways

  • Immediate:
    • Validate AI model performance against clear benchmarks before deployment.
    • Maintain data rights, interoperability, and exit options when procuring private sector AI solutions.
  • Strategic:
    • Invest heavily in foundational data infrastructure and internal government capacity.
    • Evaluate AI solutions based on overall welfare impacts and cost-effectiveness across beneficiaries and non-beneficiaries rather than pure predictive accuracy.
  • Questions to investigate:
    • How can governments in low-income settings build sustainable monitoring and evaluation systems for AI?
    • What are the optimal policy levers to balance algorithmic accuracy with privacy guarantees and community legitimacy?

Claims Worth Verifying

  • In all low- and middle-income countries together, there are less than 50,000 radiologists for a population of 3.7 billion people. (statistical)
  • Among lower-income economies, 81% report basic or no performance monitoring for AI investments. (survey finding)
  • ConsiliumBots in Chile increased college attendance rates by up to 44%. (empirical evaluation)

Notable Quotes

"How can AI improve the delivery of public services in developing countries?" (at 1:40) "Better prediction does not equal better policy." (at 57:14) "The most accurate model is not always the most effective." (at 62:11)

Compressed Summary

  • AI offers immense potential to bridge human capacity gaps in LMIC public service delivery.
  • Front-end applications focus on augmentation and task shifting but face last-mile barriers.
  • Back-end applications excel in forecasting and optimization but require robust data infrastructure.
  • Governance must address data quality, procurement lock-in, privacy, and continuous performance evaluation.
  • Keywords: artificial intelligence, public services, developing economies, social protection, governance
  • Core insight: While artificial intelligence can dramatically improve public service delivery and fill critical data gaps in developing economies, successful deployment requires rigorous governance, robust data infrastructure, and careful balancing of predictive accuracy against equity, privacy, and cost-effectiveness.

Core insights

6
Tradeoffhigh noveltystrong evidence

Optimizing predictive accuracy is a different objective from optimizing welfare: when screening/administrative costs dominate at national scale, simpler or alternative targeting methods can beat complex ML models in welfare maximization.

Why it matters

It breaks the default assumption that a better model means a better system. The cost structure of the surrounding pipeline (verification, screening, enrollment friction) can dominate model error, so the correct optimization target is end-to-end welfare, not a model metric like AUC or accuracy.

Generalization

Any agent or ML system whose output must be acted upon by a human/institutional process should be evaluated against a full cost-inclusive objective; the scoring function should price the downstream verification and action costs, not just prediction quality.

Better prediction does not equal better policy; trade-offs exist between model accuracy, interpretability, privacy, and community legitimacy.
Open source video
When screening costs dominate at a national scale, simpler or alternative targeting methods can outperform complex ML models in welfare maximization.
Open source video
Architecturemedium noveltystrong evidence

Public-sector AI decomposes into two architecturally distinct tiers: front-end tools used by citizens/frontline providers (constrained by last-mile conditions) and back-end tools used by centralized agencies for system-level prediction, allocation, forecasting, and oversight.

Why it matters

The two tiers have different failure modes, different data dependencies, and different latency/availability requirements. Front-end is bottlenecked by connectivity, literacy, devices, and language; back-end is bottlenecked by data infrastructure, secure access, interoperability, and standards. Treating them as one system hides which constraints actually bind.

Generalization

Agent systems can be partitioned into user-facing interaction layers with hostile/unreliable environments versus centralized orchestration/analytics layers with strict data-governance requirements; each tier warrants its own reliability and evaluation strategy.

AI deployment divides into front-end (citizen/provider apps) and back-end (centralized agency and big data systems).
Open source video
Front-end deployments face severe last-mile constraints: connectivity, literacy barriers, device access, and language support.
Open source video
Mechanismhigh noveltymoderate evidence

Back-end AI can substitute for physical sensing/measurement infrastructure by learning from historical data rather than requiring dense new instrumentation (e.g., flood forecasting without extensive in-river sensing).

Why it matters

It reframes AI as an infrastructure-compression technology: instead of building out expensive physical sensor networks, existing historical data can be leveraged to produce system-level forecasts. This changes capital allocation decisions and the data-collection roadmap.

Generalization

Where instrumentation is expensive, models trained on existing historical/observational data can act as virtual sensors; the engineering question becomes whether the data regime is rich enough to support the substitution, and how to handle distribution shift.

Improves forecasting without requiring extensive local sensing infrastructure (e.g., Google Floodhub).
Open source video
Google Floodhub flood forecasts using past data rather than extensive in-river sensing infrastructure.
Open source video
Empirical Resultmedium noveltystrong evidence

There is a structural evidence gap in AI-for-development deployment: many applications are technically mature and promising, but only a tiny fraction of published work evaluates real-world outcomes or cost-effectiveness — evaluation is measuring capability demos, not effects.

Why it matters

It identifies a mismatch between what is published and what a deployment decision needs. It predicts that the binding constraint on adoption is not model capability but the absence of outcome and cost evidence in the target environment.

Generalization

For agentic systems, publishing benchmark performance on capability tasks without measuring task-level cost and end-outcome effects yields investment decisions with no calibration; the evaluation harness must measure outcomes and cost, not just capability.

While promising, many high-value applications lack rigorous evidence on cost-effectiveness and long-term human development outcomes.
Open source video
Systematic literature reviews on AI in frontline healthcare and tutoring showing that only a tiny fraction of papers evaluate patient outcomes or cost-effectiveness in low-resource settings.
Open source video
Policymakers should target investments in domains with consolidated evidence rather than adopting unproven technologies.
Open source video
Failure Modemedium noveltystrong evidence

Production AI deployments in this sector operate with almost no observability: the large majority of lower-income economies report only basic or no performance monitoring for their AI investments.

Why it matters

Without monitoring there is no feedback loop, no drift detection, and no ability to attribute welfare effects or detect harm — meaning systems stay in production on unreviewed assumptions. Observability is a prerequisite, not an optimization.

Generalization

Any long-running AI or agent deployment needs instrumentation as a first-class component; absence of M&E turns silent regressions into unmeasurable policy risk and prevents post-hoc audit.

81% of lower-income economies report basic or no performance monitoring for AI investments.
Open source video
Mental Modelmedium noveltymoderate evidence

Front-end AI splits into two distinct modes — augmenting the quality of existing services vs. expanding access through automation and task shifting — and each implies a different reliability contract.

Why it matters

Augmentation assumes a human is in the loop to catch errors; task-shifting replaces a scarce expert (e.g., clinical decision support where radiologists are near-absent), so errors have no backstop. The same model quality has very different risk profiles depending on which mode it serves.

Generalization

Classify agent deployments by whether a competent human verifies each output; systems that remove the verifier need far stronger validation, guardrails, and escalation paths than systems that assist one.

Front-end use types fall into augmenting existing quality and expanding access through automation and task shifting.
Open source video
AI-assisted tuberculosis screening from chest X-rays in Nigeria and India.
Open source video

Deep dives

5

Cost-inclusive welfare objective for targeting and triage systems

Research question

At what screening/administrative cost ratio does a simple targeting rule overtake a complex ML model in welfare maximization, and can the crossover be predicted ex ante?

Why

It determines whether model selection should optimize accuracy or end-to-end welfare, which changes architecture instrumentation and can make simpler, more interpretable systems optimal.

Better prediction does not equal better policy; trade-offs exist between model accuracy, interpretability, privacy, and community legitimacy.
Open source video
When screening costs dominate at a national scale, simpler or alternative targeting methods can outperform complex ML models in welfare maximization.
Open source video
Source video

Minimal observability primitives for low-capacity public AI deployments

Research question

What minimal instrumentation can detect drift, harm, and silent regressions in institutions that currently report only basic or no performance monitoring?

Why

Without observability there is no feedback loop, drift detection, or welfare attribution; monitoring is a prerequisite for safe production deployment.

81% of lower-income economies report basic or no performance monitoring for AI investments.
Open source video
Source video

Outcome-and-cost evaluation harnesses for frontline AI

Research question

What evaluation design produces credible human-development and cost-effectiveness evidence for AI deployments in low-resource settings?

Why

The adoption bottleneck is not model capability but the absence of outcome and cost evidence in the target environment; evaluation must measure effects, not demos.

While promising, many high-value applications lack rigorous evidence on cost-effectiveness and long-term human development outcomes.
Open source video
Systematic literature reviews on AI in frontline healthcare and tutoring showing that only a tiny fraction of papers evaluate patient outcomes or cost-effectiveness in low-resource settings.
Open source video
Source video

Tier-specific reliability contracts for front-end vs back-end public-sector AI

Research question

How should evaluation, reliability, and data-governance requirements differ between citizen/frontline-facing AI and centralized agency back-end AI?

Why

The two tiers have different failure modes and bottlenecks; treating them as one system hides which constraints actually bind and leads to mismatched investment.

AI deployment divides into front-end (citizen/provider apps) and back-end (centralized agency and big data systems).
Open source video
Front-end deployments face severe last-mile constraints: connectivity, literacy barriers, device access, and language support.
Open source video
Source video

Virtual sensors and historical-data substitution for physical sensing infrastructure

Research question

Under what data regimes can learned models reliably substitute for physical sensing infrastructure, and where does historical-data-only forecasting degrade?

Why

It reframes AI as infrastructure compression, changing capital allocation and data-collection roadmaps, but requires conditions on data richness and distribution shift.

Improves forecasting without requiring extensive local sensing infrastructure (e.g., Google Floodhub).
Open source video
Google Floodhub flood forecasts using past data rather than extensive in-river sensing infrastructure.
Open source video
Source video

Article ideas

4

Better Prediction Is Not Better Policy: Targeting Systems Need a Welfare Objective

Public-sector targeting systems should be selected on end-to-end welfare including screening and administrative costs, not on predictive accuracy alone, because at national scale simpler and more interpretable rules can dominate complex ML.

Angle

Use the accuracy-vs-welfare tradeoff to argue for cost-parameterized benchmarking and instrumentation of every downstream verification step.

Source video

The Observability Gap Is the Real AI Governance Crisis

Deploying public-sector AI without performance monitoring is not a gap to fix later; it is an active policy risk that prevents drift detection, harm attribution, and audit, and must be designed as a first-class component.

Angle

Start from the 81% monitoring figure and derive minimal viable observability patterns for low-capacity agencies.

Source video

Front-End vs Back-End AI Is an Architecture Decision, Not a Use-Case Decision

Choosing where to deploy AI in government should be driven by institutional digital maturity and tier-specific constraints, not by the attractiveness of the use case; otherwise front-end tools fail on last-mile realities and back-end tools fail on data governance.

Angle

Compare failure modes, data dependencies, and latency/availability requirements across the two tiers.

Source video

Capability Demos Do Not Justify Deployment: Outcome Evidence Is the Binding Constraint

The bottleneck for AI in development is not model capability but the absence of rigorous outcome and cost-effectiveness evidence in low-resource settings; evaluation harnesses must measure task-level cost and end outcomes, not benchmark scores.

Angle

Contrast published capability research with what a deployment decision actually requires.

Source video

Project ideas

4

Welfare-Targeting Crossover Benchmark

beyond-evals

For national-scale targeting, as the screening/administrative cost ratio rises above a threshold, a simple PMT or poverty-map rule achieves higher welfare than a complex ML model despite lower AUC.

Proof of concept

Build a cost-parameterized reanalysis benchmark over public targeting datasets and sweep screening costs to compare ML, proxy-means tests, poverty maps, and community targeting on welfare rather than accuracy.

Measurement

Welfare curves across screening-cost levels, crossover threshold location, and divergence between accuracy rankings and welfare rankings.

Source video

Minimal Observability Kit for Public AI

gatehouse

Lightweight event logging, outcome proxies, and drift alarms can detect silent regressions and likely harms in low-capacity public agencies at low operational overhead and with usable false-alarm rates.

Proof of concept

Instrument a prototype public-service triage or chatbot with only basic logs, proxy outcome labels, and weekly drift alarms; test with simulated and field data.

Measurement

Detection latency, false alarm rate, cost per deployment, and staff usability in low-capacity settings.

Source video

Front-End Reliability Contract Simulator

movement-lab

The same model accuracy produces materially different harm rates when deployed in augmentation mode versus task-shifting mode, and task shifting requires escalation and calibration thresholds beyond those needed for human-in-the-loop augmentation.

Proof of concept

Simulate clinical decision support and tutoring deployments with and without a competent human verifier, varying model accuracy, prevalence, and escalation policy.

Measurement

Harm rate per decision, escalation coverage, expert time required, and cost per correct decision across modes.

Source video

Virtual Sensor Substitution Tester

new

Historical-data-only forecasting models match instrumented ground truth in regimes with long stable histories but degrade predictably after distribution shift or regime change.

Proof of concept

Compare a historical-data-only flood-forecasting model against instrumented basin baselines across regions and time periods.

Measurement

Forecast skill versus instrumented truth, degradation after regime shifts, and data-sufficiency thresholds for substitution.

Source video

Architectural implications

3

The optimization objective of a targeting/triage model should include the cost of the surrounding screening and administrative process, not only predictive accuracy.

Before

Model selection is driven by predictive accuracy metrics (best AUC / lowest prediction error).

After

Model selection is driven by an end-to-end welfare objective that explicitly prices screening cost, administrative friction, and exclusion errors.

Consequence

Simpler, cheaper, more interpretable models can win; architecture must expose and instrument the cost of each downstream verification step.

Source video

Back-end agency AI sits on top of a data-governance substrate: interoperable data infrastructure, secure access mechanisms, and clear standards are prerequisites, not add-ons.

Before

Deploy the model; treat data plumbing and access control as an implementation detail.

After

Build data interoperability, secure access, and standards as a platform layer that back-end AI depends on.

Consequence

The scarce resource shifts from model capability to institutional data infrastructure and governance capacity.

Source video

Deployment tier (front-end vs back-end) should be chosen based on an institution's digital maturity and internal capacity rather than on the attractiveness of the use case.

Before

Adopt the highest-value AI application available.

After

Match deployment tier and complexity to existing connectivity, device, literacy, and institutional capacity.

Consequence

Prevents investing in front-end last-mile-constrained tools where the environment cannot support them.

Source video

Tradeoffs and failure modes

4

Model accuracy vs interpretability, privacy, and community legitimacy

Benefit

Higher predictive accuracy from complex ML models.

Cost or risk

Loss of interpretability and privacy, and reduced legitimacy with the affected community; accuracy gains may not translate to better policy.

Better prediction does not equal better policy; trade-offs exist between model accuracy, interpretability, privacy, and community legitimacy.
Open source video
Source video

ML targeting vs simpler/alternative targeting at national scale

Benefit

Complex ML can select beneficiaries more precisely.

Cost or risk

When screening costs dominate at national scale, the complex model can underperform simpler methods on welfare maximization.

When screening costs dominate at a national scale, simpler or alternative targeting methods can outperform complex ML models in welfare maximization.
Open source video
Source video

Back-end AI scale vs manipulation and abuse risk

Benefit

Scalable forecasting, resource allocation, and oversight across populations.

Cost or risk

Requires strong data infrastructure and secure access; carries risks of misuse, manipulation, and cybercrime.

Requires strong data infrastructure, secure access mechanisms, and clear standards, while carrying risks of misuse, manipulation, and cybercrime.
Open source video
Source video

Front-end AI access expansion vs last-mile feasibility and equity

Benefit

Expands access to services and shifts tasks away from scarce experts.

Cost or risk

Connectivity, literacy, device access, and language barriers can leave the intended beneficiaries unreachable, and cost-effectiveness is largely unmeasured.

Front-end deployments face severe last-mile constraints: connectivity, literacy barriers, device access, and language support.
Open source video
Source video

Open questions

4

At what level of screening/administrative cost does a simple targeting rule overtake a complex ML model in welfare maximization, and can that crossover be predicted ex ante?

Why unresolved

The summary reports that simpler methods can outperform complex ML when screening costs dominate, but does not give the threshold or a method for estimating it before deployment.

Research direction

Build a cost-parameterized benchmark that sweeps screening cost and compares ML, PMT, poverty maps, and community targeting on welfare rather than accuracy.

Source video

What evaluation design would produce credible human-development and cost-effectiveness evidence for AI deployments in low-resource settings?

Why unresolved

The summary states that only a tiny fraction of papers evaluate patient outcomes or cost-effectiveness, but does not specify what valid evaluation looks like.

Research direction

Prototype outcome-and-cost evaluation harnesses (field trials or quasi-experimental designs) for frontline AI tools such as tutoring and clinical decision support.

Source video

Under what data regimes can learned models reliably substitute for physical sensing infrastructure?

Why unresolved

The summary offers flood forecasting from historical data as an example but does not define the data sufficiency or drift conditions under which such substitution holds.

Research direction

Test virtual-sensor models against instrumented ground truth across regions to characterize where historical-data-only forecasting degrades.

Source video

How can monitoring and evaluation be made cheap and feasible enough for institutions that currently report no or only basic performance monitoring?

Why unresolved

81% of lower-income economies have basic or no monitoring, but the summary does not describe a workable low-cost observability design.

Research direction

Design minimal-friction observability primitives (lightweight logging, outcome proxies, drift alarms) suitable for low-capacity public agencies.

Source video

Key claims

7
factualVerification needed

LMICs have fewer than 50,000 radiologists serving 3.7 billion people.

Evidence

LMICs suffer from severe capacity gaps, such as having fewer than 50,000 radiologists for 3.7 billion people.

Question

What are the source and year of the radiologist-per-population figures by income group?

Source video
causalVerification needed

ConsiliumBots in Chile increased college enrollment rates by up to 44% and higher-ranked program enrollment by 20%.

Evidence

ConsiliumBots in Chile increasing college enrollment rates by up to 44% and higher-ranked program enrollment by 20%.

Question

What study design and effect sizes underlie the 44% and 20% figures, and do they persist at scale?

Source video
factualVerification needed

81% of lower-income economies report basic or no performance monitoring for AI investments.

Evidence

81% of lower-income economies report basic or no performance monitoring for AI investments.

Question

What survey instrument, sample, and definition of 'basic performance monitoring' produced the 81% figure?

Source video
causalVerification needed

Better prediction does not equal better policy.

Evidence

Better prediction does not equal better policy; trade-offs exist between model accuracy, interpretability, privacy, and community legitimacy.

Question

In which documented cases did higher predictive accuracy fail to improve welfare outcomes?

Source video
comparativeVerification needed

When screening costs dominate at national scale, simpler or alternative targeting methods can outperform complex ML models in welfare maximization.

Evidence

When screening costs dominate at a national scale, simpler or alternative targeting methods can outperform complex ML models in welfare maximization.

Question

What is the threshold of screening cost at which simple targeting dominates ML across Togo, Malawi, and Bangladesh?

Source video
causalVerification needed

Flood forecasting can improve without extensive local sensing infrastructure by using past data.

Evidence

Improves forecasting without requiring extensive local sensing infrastructure (e.g., Google Floodhub).

Question

How accurate is historical-data-only flood forecasting relative to instrumented baselines, and where does it fail?

Source video
factualVerification needed

Only a tiny fraction of AI healthcare and tutoring research evaluates patient outcomes or cost-effectiveness in low-resource settings.

Evidence

Systematic literature reviews on AI in frontline healthcare and tutoring showing that only a tiny fraction of papers evaluate patient outcomes or cost-effectiveness in low-resource settings.

Question

What proportion of reviewed papers measured outcomes and cost, and were the reviews systematically sampled?

Source video

Connections

5