Advancing Responsible AI in Public Sector Application (Abhishek Singh and Balaraman Ravindran, eds.)
In this Book Essential
INTRODUCTION
Governments now buy artificial intelligence faster than they understand it. This edited volume, produced from a Ministry of Electronics and Information Technology and IIT Madras symposium, documents what happens when public agencies procure, deploy, and audit systems that decide who receives welfare, who is screened for disease, and who is flagged for investigation.
The book's central argument is that procurement documents, not legislation, are the operative policy instrument for public sector AI. Gisele Waters and Cari Miller put it directly in Chapter 1: government agencies rarely understand the systems they buy, and the resulting contracts silently become policy. Chapters 2 through 14 then work through the consequences across data architecture, market structure, health screening, citizen participation, risk scoring, agriculture, global equity, higher education, French public ethics, APEC standards, Indian algorithmic auditing, and Rwanda's Mbaza chatbot.
This Book Essential is for public officials, procurement staff, policy analysts, engineers, and researchers who need to move from AI principles to operational instruments. It is also for anyone assessing whether a government AI system should exist at all. The book argues that question belongs at the start of the acquisition, not in the post-incident review.
The volume is unusual in two respects. First, it is genuinely interdisciplinary: contributors include medical professionals, government officials, civil society actors, lawyers, economists, and computer scientists. Second, several chapters report negative or incomplete findings, including a health validation where vendors refused to disclose training data provenance, a Z-Inspection assessment still in progress at publication, and a Rwandan chatbot whose headline user numbers belong to a rules-based channel rather than its AI components. Those admissions raise the book's credibility.
U365'S VALUE PROPOSITION
WHO THIS IS FOR
Public sector procurement and policy staff, who need verifiable evidence requirements they can write into tenders rather than accepting vendor claims.
AI engineers and data scientists building for government, who need to understand why technical accuracy alone does not survive an audit.
Fellows working in health, agriculture, education, or public administration, who will meet these systems as operators rather than as authors.
Researchers and students in AI governance, who need the specific case evidence (IEEE P3119, DEPA, Z-Inspection, the Mbaza project) rather than abstract principle lists.
Anyone preparing to argue for or against a public AI deployment, who needs a defensible framework for risk, liability, and remedy.
KEY TENSIONS
Evidence versus marketing: Governments rely on vendor claims because independent validation data is unavailable. Chapter 5 records a health study where training data provenance, annotation method, and prior validation were all withheld, leaving procurement officers to compare algorithms they cannot inspect.
Voluntary standards versus binding effect: IEEE P3119 is a voluntary consensus standard, yet once a regulator or procuring agency calls it up in a contract it becomes mandatory for bidders. Chapter 12 describes this as a strength. The same mechanism transfers compliance cost to the smallest suppliers.
Transparency versus gaming: Publishing how an algorithm works lets citizens understand it and lets bad-faith actors exploit it. Chapter 4 names this tension directly and proposes mediated and direct register variants rather than pretending it resolves.
Local harm versus imported metrics: Fairness metrics developed in the Global North misfire on caste, religion, and language patterns specific to India. Chapter 13 shows the Foundation Model Transparency Index cannot capture these because it measures developer disclosure, not outcome disparity.
Participation versus delivery speed: Chapter 6 documents voluntary community annotators producing low and inconsistent agreement rates and dropping out, while Chapter 14 shows Rwanda building a Kinyarwanda speech model from scratch with a partner that later withdrew.
Risk scoring versus weight selection: Chapter 7 builds a weighted risk score for AI, then concedes the weights are set by a regulator with no published meta-rule. Its own worked example moves the scaled score from 2 to 4.5 for the same technology.
WHY IT MATTERS NOW
Public sector AI deployment has moved past the pilot stage in several countries while oversight has not. Chapter 12 reports that only 55 percent of institutions that adopted AI consider their automation programme successful, and that three out of five surveyed people are wary about trusting AI systems. Chapter 9 quantifies the asymmetry: of 91.9 billion US dollars in 2022 private AI investment, 47.36 billion went to the United States and 3.24 billion to India.
The regulatory landscape is also consolidating around instruments this book examines. The EU AI Act, NIST's AI Risk Management Framework, ISO/IEC 42001, and IEEE 3119 now coexist, and Chapter 12 maps how standards get incorporated into the European framework through the Vienna Agreement. Any public agency buying AI today is already inside one of these regimes.
OVERVIEW
The book's core message is that responsible AI in the public sector is an institutional design problem, not a technical one. Each chapter examines a different institution: procurement offices, data intermediaries, algorithm registers, hospital validation committees, participatory research teams, insurers, standards bodies, audit functions, and government delivery units. The recurring finding is that the failure point is usually the interface between organisations, not inside any single model.
Its approach is deliberately mixed. Chapter 1 is a standards-construction account of the IEEE P3119 working group, 48 members from 12 countries building consensus over two years. Chapter 2 is a technical architecture description with concrete parameters, including a training contract fixing an epsilon threshold of 1.5 and a noise multiplier of 2.0. Chapter 5 is a prospective validation study with 1,099 retinal images from 250 patients. Chapter 6 is a reflexive account of building a gendered-abuse detection model on roughly 1.3 million tweets. Chapter 14 is a qualitative single case study built on nine semi-structured interviews.
Structure follows the governance lifecycle rather than the technology stack. Chapters 1 through 3 address acquisition and infrastructure. Chapters 4 through 6 address citizen voice, health validation, and participation. Chapter 7 supplies risk and liability machinery. Chapters 8 and 9 address agriculture and global inequity. Chapters 10 through 12 address trust assessment, applied ethics, and regional standards. Chapters 13 and 14 close with auditing and delivery experience.
KEY IDEAS
RESPONSIBLE AI IN THE PUBLIC SECTOR
Procurement is policy: The acquisition document defines what a system must do, what evidence a vendor must supply, and what happens when it fails. Waters and Miller argue that this document does more to determine public outcomes than any published AI principle.
Automation without remedy: Chapter 1 documents systems deployed without redress mechanisms. It cites the Internal Revenue Service contract worth 86 million US dollars with ID.me for biometric identity verification in 2021, followed by nearly half a million unresolved identity-theft assistance cases.
A fiduciary duty to the public: The authors frame high-risk AI procurement as a duty of loyalty and care owed by government to citizens, not a commercial transaction between equal parties.
Data empowerment with protection: Chapter 2 argues the binding constraint on useful public sector machine learning is shareable data, and that the barrier is structural. Benefits of sharing are asymmetric while risk concentrates on the data holder.
Digital contracts and clean rooms: DEPA rests on three mechanisms: self-enforcing electronic contracts recorded in a verifiable registry, hardware-isolated confidential clean rooms, and differential privacy with an explicit privacy-loss budget.
Differential privacy in practice: An algorithm is differentially private when the output cannot reveal whether any individual's data was included. DP-SGD is named as the most common training method.
Technological monoculture: Chapter 3 shows that costly training data and compute concentrate the market in a few predominantly closed-source foundation models, so downstream applications inherit shared weaknesses.
Correlated failure: Fine-tuning eight firm-specific LLaMA-2 models produced area under the curve comparable to a logistic regression baseline, yet with significant correlated false negatives between firms.
Public algorithm registers: Chapter 4 assesses the first wave of registers in Amsterdam, Helsinki, Nantes, Antibes, New York City, Rotterdam, and Brussels, and finds their implementation undermines their stated purpose of building trust.
Validation before adoption: Chapter 5 masks vendor identities as A1, A2, and A3 and reports sensitivity from 59.69 to 97.74 percent and specificity from 14.25 to 96.01 percent, showing that real-world performance diverges sharply from vendor claims.
Created and invited spaces: Chapter 6 uses Cornwall's distinction between participation people originate and participation they are invited into, and shows why the second kind breeds fatigue.
Quantified AI risk: Chapter 7 scores each criterion 1 or 2 across five OECD dimensions, applies regulator-set weights, and publishes a QR-coded risk profile for citizens, insurers, and authorities.
Algorithmic auditing as layers: Chapter 13 treats impact assessment, technical audit, accountability, and redress as connected layers, where each level's output becomes another level's input.
ULM ALIGNMENT
At University 365, the CI-First (Co-Intelligence First) doctrine teaches you to always invite AI into your reflection and work while remaining the orchestrator. Human Intelligence leads, AI amplifies. This Book Essential connects the book's ideas to U365's proprietary methods: ULM+EVA (University 365 Life Management powered by the Explore-Visualize-Action Plan cycle) helps you map goals across six life domains; LIPS+CARE (your digital second brain with the Collect-Action Plan-Review-Execute cycle) captures and organizes what you learn; and SL-OS (Successful Life Operating System) integrates all of these with UP-Context (context engineering for AI) into a unified life and learning system.
SCHEMA PRIME: ULM Domain: Career and Finance
DOMAIN MAPPING: Primary: Career and Finance (professional practice in AI governance). Secondary: Quality of Life (public services shaping daily conditions).
EVA PARAGRAPH: You want to work with AI systems, or make decisions about them, in a way you could defend publicly. The obstacle is that vendor claims, dashboards, and accuracy figures arrive before you have any method for testing them, so you accept them and move on. If a tool or a proposal reaches you this month, then before responding you will write down one question about training data provenance, one about who is accountable when the output harms someone, and one about what remedy exists. If the answers are unavailable, record that as a finding rather than a gap in your knowledge. That three-question habit is the whole method. Apply it first to the lowest-stakes system in your environment so the wording is ready when the stakes rise.
LIPS+CARE CAPTURE CARD:
Book: Advancing Responsible AI in Public Sector Application (GPAI Edition)
Primary ULM Domain: Career and Finance
3 Key Takeaways:
1. Procurement documents set public policy, so evidence requirements must be written into tenders rather than assumed.
2. Technical accuracy is only one of four layers; accountability and redress must be designed alongside it.
3. Data sharing fails because benefit is asymmetric and risk is concentrated, so the fix is contractual and architectural, not rhetorical.
Apply It Action: In the Career and Finance domain, take one AI system you depend on at work and write a three-line card covering its training data, its accountable owner, and its remedy path.
Next CARE Step: Collect the answers you can obtain, then review which questions the supplier refuses to answer.
EXPLAINER LINK: For more on ULM, EVA, LIPS, and CARE methods, see the [ULM page](https://www.university-365.com/ulm) and the [LIPS page](https://www.university-365.com/lips). For the CI-First doctrine, see the [CI-First page](https://www.university-365.com/ci-first). For the full SL-OS, see the [SL-OS page](https://www.university-365.com/slos).
APPLY IT (domain-tagged): In the Career and Finance domain, draft a one-page evidence checklist for any AI tool you or your organisation is assessing, listing the training data source, the accountable owner, the monitoring frequency, and the remedy path, then use it on the next proposal you receive.
SUMMARY
MINDMAP SKELETON: Advancing Responsible AI in Public Sector Application
Center: Advancing Responsible AI in Public Sector Application
Branch 1: AI Procurement Standards
IEEE P3119 processes
Vendor claims versus evidence
Market-shaping power
Sandbox benchmarking
Branch 2: Data and Governance
DEPA architecture
Digital contracts
Confidential clean rooms
Technological monoculture
Correlated failures
Branch 3: Citizen Voice
Public algorithm registers
First and second order transparency
Accessibility for non-experts
Accountability channels
Branch 4: Health and Participation
Diabetic retinopathy validation
Masked vendor evaluation
Created and invited spaces
Trust despite opacity
Branch 5: Risk and Liability
Risk quantification
Cybersecurity precedent
Insurance application
Liability determination
Branch 6: Agriculture and Equity
Bhashini for farmers
Four chatbot deployments
Global South extraction
Regional inequity
Branch 7: Trust and Standards
Z-Inspection assessment
Actionable ethics
APEC standards landscape
AI at scale
Branch 8: Audit and Practice
Third-party auditing
Impact assessments
Rwanda Mbaza chatbot
Government delivery
Reconstruction prompt: "Draw a mindmap with this structure. Center node at top, branches arranged vertically below, leaves extending outward from each branch. Use a soft modern color palette with clean lines on a white background."
Chapter 1: Raising a Global Standard in AI Procurement
Waters and Miller establish procurement as the decisive governance stage. They document that governments contract up to 11 trillion US dollars in goods and services, that roughly 12 percent of OECD spending is public procurement, and that government IT spending reached 600 billion US dollars, an 8.9 percent rise over 2022. Their method is the IEEE P3119 working group: 48 members from 12 countries over two years, briefed to more than eight national or transnational agencies, compared against the World Economic Forum's AI Procurement in a Box, the United States Government Accountability Office accountability framework, and the Ford Foundation's guiding framework. The standard supplies five processes across four procurement stages, each with purpose, scope, outcomes, activities, tasks, inputs, outputs, and rubric appendices, aligned with NIST's AI Risk Management Framework.
The chapter's strongest counter-evidence undercuts its own remedy. The New York City Automated Decision Task Force, created by Local Law 49 in 2018 in a city of nearly nine million people, was blocked from identifying a single automated decision system by the city itself. The task force was already cross-disciplinary, legally mandated, and well intentioned. What it lacked was an inventory of installed systems, not collaboration design or standard content. P3119's problem definition step governs new acquisitions, and nothing in its five processes compels an agency to disclose its existing base. The standard is aimed at the stage where this chapter's best evidence shows the gap was not located.
A second limitation is legal rather than procedural. Chapter 1 cites Article 47 of the EU Charter of Fundamental Rights on the right to an effective remedy, and the Australian Robodebt Royal Commission as evidence of what proceeding without caution costs. Yet the chapter never specifies who pays for remedy when a system is withdrawn after deployment. Its fiduciary framing implies a public duty, while the procurement mechanism it proposes binds suppliers through contract terms. Those are different obligations with different enforcement paths, and the chapter treats them as one.
Chapter 2: Data Empowerment and Protection Architecture
Shyam Sundaram and colleagues argue that the binding constraint on useful public machine learning is not compute or talent but shareable data, and that the barrier is structural: benefits of data sharing are asymmetric, while risk concentrates on the data holder. They call the result a frozen market. A hypothetical Bangalore startup training chest X-ray COVID detectors on fewer than 50,000 publicly available images demonstrates the ceiling: local hospital deployment works, national and global scaling require tagged clinical data that hospitals will not hand over. DEPA for training is presented as the pivot that permits scaling without surrendering control.
The technical content is unusually specific. The framework rests on electronic contracts recorded in a verifiable registry, confidential clean rooms built on Intel SGX or TDX, AMD SEV-SNP, ARM CCA, and NVIDIA confidential GPUs with remote attestation, and differential privacy where a privacy-loss budget controls added noise. A sample training contract fixes an epsilon threshold of 1.5, a noise multiplier of 2.0, a delta of 0.01, and two epochs per report. The ADEPTS principles name accountability, democratized data, transparency, privacy by design, and security. Governance sits beneath GDPR, CCPA, and India's DPDP Act 2023, with the Account Aggregator cited as proof the pattern works.
The chapter's blind spot is what happens after training. Its three pillars secure the pipeline, and differential privacy is a guarantee about the training dataset rather than about what the finished model later infers about people who never contributed data. The sample contract constrains epochs and epsilon thresholds during training only, and the techno-legal framework limits model data usage without any post-deployment monitor. Chapter 1 supplies exactly that missing function as its fifth process, contract monitoring with tolerance thresholds that trigger intervention. A monitoring clause enforceable through the registry, covering live epsilon budget and drift, would close the gap that DEPA's own architecture leaves open.
Chapter 3: Generative AI Governance and Correlated Failures
Ramayya Krishnan and colleagues show why responsible use cannot be judged one model at a time. Costly training data and compute concentrate the market in a few predominantly closed-source foundation models, so downstream applications share upstream weights and overlapping training data. Their empirical work uses applicant tracking data from eight United States firms spanning 2014 to 2018, covering 1.17 million applications for 6,600 job postings, reduced to 25,000 individuals with 65,000 applications across 3,600 jobs. Off-the-shelf LLaMA-2-7B and 13B models are prompted with resume and job description.
Two results anchor the argument. Off-the-shelf models have almost no predictive power for screening, producing uncorrelated errors. Parameter-efficient fine-tuning then creates eight firm-specific LLaMA-2-7B models that match a logistic regression baseline on area under the curve, yet show significant correlated false negatives between firms. Correlation is therefore a property of adaptation rather than of shared base models alone. The authors place this in a lineage running from agricultural monoculture to the Windows security monoculture, citing Braess's paradox results on homogeneous hiring algorithms.
The design excludes the case that would make the title claim strongest. Correlation is demonstrated only for open-weight models that firms can fine-tune, while the closed-source models whose market dominance the introduction invokes cannot be weight-adapted and appear nowhere in the experiment. The natural policy inference, that permitting API-level access rather than weight transfer breaks the correlation channel, remains untested. A second gap is legal: correlated false negatives harm applicants without obviously violating any single firm's obligations, so the chapter names a collective-action problem without proposing a collective instrument such as shared audit, base-model disclosure, or joint liability.
Chapter 4: Public Algorithm Registers and Citizen Empowerment
Jens Meijen and Niharika Gujela assess the first wave of public algorithm registers and find a strategic misalignment: registers are intended to build trust and democratic accountability, yet their implementation undermines both. The review synthesises five design questions covering legal context, disclosure volume, intended audience, information content, and disclosure method, alongside principles from the Ada Lovelace Institute, AI Now, and the Open Government Partnership. Case material covers registers in Amsterdam, Helsinki, Nantes, Antibes, New York City, Rotterdam, and Brussels, plus the Dutch national register.
The method is literature synthesis plus interviews with administrators in Rotterdam, Brussels, and the national register. Four shortcoming categories emerge: quality and quantity of information, citizen experience, accessibility for non-experts, and accountability. New York City's directory is criticised for technical language and PDF format; Amsterdam's register limits technical detail to third-party auditors. The authors distinguish mediated registers reached through experts from direct citizen-facing registers, and concede their recommendations presume highly digitised administrations.
The chapter recommends binding legal obligations to complete registers without any outcome validity test. An evaluative move the authors do not make is to demand evidence that register content changes administrative conduct rather than merely documenting it. A falsifiable design would correlate published redress pathways and complaint disclosures with actual overturned decisions, which the accountability section recommends publishing but never measures. Chapter 7 offers a partial answer, since its QR-coded risk portal carries a structured dimension profile that would resolve the expert-versus-citizen trade-off this chapter calls intractable, yet neither chapter cites the other.
Chapter 5: Validating Cloud AI for Diabetic Retinopathy Screening
Mona Duggal and colleagues report prospective validation of commercial cloud AI for diabetic retinopathy screening in Indian public health settings, arguing that real-world performance rather than vendor claims must govern adoption. Governance is operationalised through a seven-principle matrix mapping safety, equality, inclusivity and non-discrimination, privacy and security, transparency, accountability, and protection of human values onto specific study activities including ethics approval, advisory group formation, site selection, staff training, validation, data management, and patient feedback.
The detail is concrete and auditable. Ethics approval was granted under the Declaration of Helsinki with trial registration in India. Sites were one tertiary centre and one primary health centre in North India, with recruitment through community health workers. The sample was 250 diabetic patients, calculated from a 17 percent prevalence, an 18.4 percent non-gradable image rate, 10 percent non-response, 95 percent precision, and a 5 percent margin of error. Two optometrists trained for two weeks; images came from a portable fundus camera; data were deidentified onto a government cloud server. Five Indian cloud AI companies were invited, four agreed, one was dropped after interim analysis showed extremely low specificity, and three completed as masked entries A1, A2, and A3.
Results from 1,099 retinal images across 500 eyes show sensitivity from 59.69 to 97.74 percent and specificity from 14.25 to 96.01 percent, with one algorithm clearly best on specificity, positive predictive value, accuracy, and agreement. The governance failure is sharper than the technical one: vendors refused to disclose training data provenance, annotation method, or handling of ungradable images, and the authors' response is a recommendation for an oversight body with no mechanism named. The obvious remedy is to relocate that requirement into procurement law, making provenance, annotation protocol, and prior external validation mandatory disclosure conditions of public tenders with disqualification for non-compliance. The chapter also asserts responsible adoption without any cost-per-screen or throughput figures, so an authority cannot weigh a 37.19 percent accuracy algorithm against an 88.43 percent one, and it does not examine the ethics of silently dropping the single United States-approved algorithm after it underperformed.
Chapter 6: Participation in AI from the Trenches
Tarunima Prabhakar, Cheshta Arora, and Arnav Arora give a reflexive account of building a gendered-abuse detection model in India, motivated by feminist principles of reflexivity, participation, intersectionality, and structural change. Interviews with more than thirty researchers and activists showed most problems did not need an AI solution, so the tool was scoped narrowly: easing the fatigue of activists, researchers, and journalists, with detection and redaction the only machine-learning-dependent features. India's twenty-one constitutionally recognised languages and several hundred non-official ones forced a choice of Hindi, Tamil, and Indian English, the last because transliteration and code-mixing evaded platform moderation.
The numbers are reported honestly. Close to 1.3 million tweets from 2018 to 2021 were collected through crowdsourced slurs, known perpetrators, and replies to influential women. Sampling targeted roughly 24,000 posts, with about 20 percent triple-annotated, budgeted against local researcher rates. Annotators drawn from affected communities labelled posts between March and July 2022. Agreement was low and language-dependent, with Krippendorff alpha values from 0.258 to 0.402 in English, 0.314 to 0.501 in Hindi, and 0.411 to 0.721 in Tamil. Voluntary engagement and dropout meant the ideal dataset was never reached. The chapter's contribution is a translation framework in which needs become technical terms and constraints become experiential terms, horizontally across communities and vertically across abstraction levels, producing only partial connections.
The translation grid depicts funders, project management, data science, design, and research coordination, but no commercial model vendor and no legal or procurement intermediary. That omission matters because Chapter 5 of the same volume shows vendors withholding training data provenance from public health researchers, and Chapter 3 shows shared foundation models propagating correlated failures. Community participation is bounded by contracts the community never sees, which argues for adding the vendor and the procurement officer as mandatory translation nodes. A second unexamined point is methodological: the chapter reports agreement statistics but not what lost triple-annotation coverage did to measured model performance, so it cannot say whether retaining disagreement improved or degraded the shipped artefact.
Chapter 7: Risk Assessment and Liability Determination
Aditya Mohan and Karthik Satishkumar ask how AI risk should be quantified so that regulation, insurance, and liability determination operate predictably across borders. Their answer is an explicit analogy to the cybersecurity industry's decades-long construction of standards and compliance regimes built on a trust-but-verify posture, which they argue enabled mass IT adoption and should now be repeated for AI through GPAI-level consensus. They note definitional contestation across the OECD, the EU AI Act, and NIST before diagnosing structural problems: multi-party product chains make root cause attribution cumbersome, black-box components resist intermediate inspection, explainability gaps weaken liability determination, and cross-border deployment demands agreement.
The method is a scoring framework built on the OECD classification of AI systems, organised into five dimensions covering people and planet, economic context, data and input, model, and task and output. Each criterion receives 1 for lower risk or 2 for higher risk, with one optionality criterion scored 0 or 1 because opting out nullifies AI risk entirely. A worked example compares two facial recognition deployments: access control to private buildings scores 9.4 net and 2 scaled with equal weights, while law enforcement monitoring of public spaces scores 11.9 net and 4.5 scaled against a maximum of 14.8. Applications extend to insurance pricing, a QR-coded public portal, and liability, where a dimension score can falsify a complaint about personal data exposure.
The chapter concedes that weight selection belongs to a regulatory authority but proposes no meta-rule for setting or reviewing weights, and its own example shows how decisive that is: identical criteria move the scaled score from 2 to 4.5 when the people dimension rises from 0.2 to 0.3 and economic context falls from 0.2 to 0.1. The same framework therefore yields opposite outcomes for the same technology under two administrations, which argues for weights fixed by an independent body, published with justification, and revised only through an auditable process. Two schema entries also conflict and are never reconciled: generative and discriminative models receive the same value, while a large dataset scores as lower risk on one criterion and a sampled set scores higher on another, so a vast personal data collection can be scored safest depending on which criterion is emphasised. No inter-rater data, confidence intervals, or sensitivity analysis accompany the scoring, so two assessors could land on opposite sides of the inoperability threshold the chapter grants regulators.
Chapter 8: AI for Indian Agriculture and Bhashini
Abhishek Raj, Harsh Singh, and Anshul Pachouri argue that India's agriculture extension and advisory services cannot absorb demand for personalised advice, and that generative chatbots integrated with Bhashini can fill the gap if adoption barriers are handled responsibly. The method combines secondary research with interviews of farmers, agricultural experts, solution architects, and government officials, grounded in four named deployments. Ama KrushAI, launched by the Government of Odisha in February 2023, complements an existing service with roughly 690,000 farmer enrolments, uses a knowledge base built over five years with a state agricultural university, and ran its first pilot with 10,000 farmers. Jugalbandi converts WhatsApp text or audio through speech recognition, Bhashini translation, and a large language model, covering 10 of India's 22 official languages and 171 of approximately 20,000 government programmes. Kisan e-Mitra handles grievance resolution in five languages.
The barrier analysis quantifies a difficult context: 86 percent of farmers are small or marginal, India has 121 languages and 127 agro-climatic zones, only 11 percent of Indians speak English against 57 percent for Hindi, and only 41 percent of the rural population uses the internet. Recommendations cover regional training data and an agricultural glossary, human-in-the-loop moderation, add-on service packaging, partnerships with state agricultural universities, needs assessment, and privacy safeguards against commercial misuse of farmer data.
The chapter presents Bhashini as a neutral linguistic utility but never audits what routing the entire advisory chain through a single government API layer plus a foreign model fallback does to accountability. When the fallback model answers a query absent from the curated database, no entity named in the chapter owns the error, and the democratic knowledge system the authors invoke carries no governance design for who validates contested agronomic advice. Read against Chapter 14, the contrast is instructive: India buys language capability as shared national infrastructure, whereas Rwanda had to build speech corpora before any service existed. Bhashini therefore masks rather than solves the underlying scarcity of speech data for low-resource Indian languages, which the chapter's own barrier analysis concedes.
Chapter 9: Regional Inequities and the Global South
Saikat Datta, Shachi Solanki, and Anand Venkatanaryanan argue that AI is recapitulating the extractive logic of colonisation, and that Indian and Kenyan experience shows what the Global South must do and refuse. Their method is comparative policy and secondary data analysis. The asymmetry is quantified: of 91.9 billion US dollars in 2022 private AI investment, 47.36 billion went to the United States and 3.24 billion to India, while United States private investment reached 248.9 billion between 2013 and 2022, more than the next fourteen nations combined. North America averages 81.56 on the AI readiness index against India's 63.67 and a global average of 44.61. A study of six major technology firms shows four of them running 62 AI laboratories, 68 percent outside the United States, yet with 68 percent of staff inside it.
The labour case is the chapter's sharpest evidence. Kenyan workers labelling training data for a major AI provider earned as little as 1.3 US dollars per hour against a contract worth over 150,000 US dollars. The authors trace technology denial as an inherited pattern from the 1987 refusal to sell India a Cray supercomputer, modelled on nuclear and missile technology control regimes. On bias they show that fairness concepts developed in the North do not transfer, citing text-to-image and facial recognition failures on Southern contexts, and internet access asymmetries by gender, geography, religion, and caste, with 57.1 percent of Indian men connected against 33.3 percent of Indian women. Remedies proposed include South-South cooperation, regional research networks, geographic indication tags, and profit sharing for extracted datasets.
The chapter's flagship remedy arrives without an enforcement mechanism. Geographic indication tags are territorial statutory instruments enforceable at borders, whereas training corpora are scraped, mixed, and absorbed into model weights that carry no provenance chain, so a tag would require a value-attribution formula the chapter does not attempt. Staged against Chapter 12, a sharper contrast appears: APEC's standards harmonisation presumes interoperability benefits distribute broadly, while this chapter's investment data implies harmonised standards would entrench the North's market power. The testable extension is to run Chapter 12's survey instrument on Global South economies and check whether lack of access to guidance is genuinely a deficit or, as this chapter would predict, a rational withholding of consent.
Chapter 10: Assessing Trustworthiness in Higher Education
Adarsh Srivastava and colleagues report a pilot applying the Z-Inspection process to generative AI in higher education, asking whether tools used for course and programme outcome planning in an Indian engineering college can be assessed as trustworthy. The process has three phases, set-up, assess, and resolve, and runs socio-technical, socio-legal, and socio-economic scenarios with a protocol log. The assessment framework is the EU trustworthy AI stack: a lawful, ethical, and robust triad, four ethical principles covering human autonomy, prevention of harm, fairness, and explicability, and seven requirements with sub-requirements. The team adds two requirements of its own, assessing whether systems respect the values of modern democracy and avoiding concentration of power, and aligns with UNESCO guidance on pilot testing and monitoring.
The use case is an engineering college in Pune where faculty in three departments use generative tools to draft course outcomes and map them to accreditation-mandated programme outcomes, with manual validation by senior professors. Seven working groups operate in parallel, covering technical, ethics, ethics and education, teaching staff, law, a student representative, and a lead, each producing independent free-text reports to mitigate cognitive bias, which are then mapped onto the closed vocabulary of the EU framework. Three scenarios are assessed: current three-department use, prospective all-department use with a prototype, and large-scale adoption by other universities under governance guidelines. The assessment was still in the assess phase at publication, and legal issues are explicitly excluded.
The process design contains an internal tension the authors do not examine. To avoid cognitive bias, working groups are grouped by homogeneous expertise and run in parallel, yet the student stakeholder is represented by a single-person group, so the structure that maximises disciplinary rigour simultaneously gives the actor most exposed to the system's consequences the least analytic capacity against an expert majority. Combined with the decision to set legal questions aside, the pilot systematically under-weights the two things that most constrain real deployment: the affected party's voice and the data protection statute that would bind it. Applying the same rubric to admissions screening or student assessment would produce materially different findings than course outcome drafting, because the tertiary actor logic and liability allocation change.
Chapter 11: Actionable Ethics in the French Public Sector
Anthéa Serafin, Lisa Fériol, and Bertrand Monthubert argue that France's emerging doctrine for trusted public AI is necessary but inert without local intermediaries that convert principle into instrument, and they document one such intermediary in Occitanie. The method is institutional analysis plus two documented support engagements. The doctrine rests on seven principles from the Conseil d'État covering human primacy, performance, equity and non-discrimination, transparency, safety, environmental sustainability, and strategic autonomy. The diagnosis is structural: the public sector spans an archipelago of actors with unequal resources, ethics skills are rare, and the resulting dependence on outsourcing undermines the very autonomy the doctrine requires.
A survey of roughly 3,700 responses found that 70 percent wanted their opinion taken into account on AI projects in their territory, which grounds the participatory approach, including citizen workshops using design fiction scenarios and a charter revised at least every three years. Two cases show the tooling. A decision-support tool mapping regional pandemic resurgence risk drew recommendations on stigmatisation and re-identification through spatio-temporal resolution and minimum population thresholds. A regional employment and training system comprising two tools drew risks including erosion of atypical training provision, reinforcement of the digital divide, displacement of guidance counsellors, and foreign access to personal data. Five of six project owners who requested support obtained the ethical label.
The case for flexible, close-to-the-ground normativity rests on a neutral third party, yet the body in question is funded and populated by the very actors it assesses, and the reported outcome, five of six supported project owners passing, is a self-selected sample with no failure case, no scoring distribution, and no disclosed refusal. A label that no one fails functions as a market signal rather than an accountability instrument, and the chapter's own conclusion concedes that economic pressure pushes ethics to the background. The unexamined remedy is to make the label consequential by tying it to procurement eligibility, which would convert a voluntary grid into the operational condition the authors say public sector ethics lacks.
Chapter 12: AI at Scale and International Standards in APEC
Aurelie Jacquet, Karen Batt, and Jesse Riddell argue that the APEC region's AI opportunity is throttled by fragmentation, and that internationally agreed standards supply the missing common architecture. The chapter combines secondary evidence with original survey data from more than 70 responses across 14 member economies, grouping challenges into five themes: awareness and trust, data quality and availability, privacy and security, safety and legal and ethical concerns, and lack of skilled personnel. The opportunity is large, with AI estimated to add 15 trillion US dollars to the global economy by 2030 and lift global GDP by 14 percent. The drag is equally quantified: 61 percent of surveyed people are wary about trusting AI, only 55 percent of institutions that adopted AI consider their automation programme successful, and 85 percent of failed data-driven projects are attributed largely to unsuitable data.
Three standard types are distinguished: governance standards aimed at boards and executives, management system standards carrying risk frameworks and controls, and technical standards defining design, interoperability, performance, and security. The landscape is then mapped in detail, covering the ISO and IEC joint committee created in 2017 with 36 participating and 22 observing members, 20 published standards and 27 in development, the data and model standards series, the organisational standards including ISO/IEC 42001, the IEEE P7000 series which includes the procurement standard from Chapter 1, NIST's framework crosswalks, and the Vienna Agreement that incorporates ISO and IEC standards into the EU AI Act. Trade influence closes the case: up to 80 percent of global trade is affected by standards.
The harmonisation thesis is left under-specified precisely where it bites. ISO/IEC 42001 is described as auditable and expected to become a certification pathway, yet the chapter never asks who pays for conformance or how the small and medium enterprises it identifies as least aware of standards could afford certification that becomes de facto market entry once called up in regulation or procurement. The framing of standards as simultaneously voluntary and mandatory is treated as a strength rather than as a mechanism for transferring compliance cost from regulators to the smallest suppliers. Chapter 9's evidence on investment asymmetry predicts that harmonised APEC standards would concentrate advantage in the economies already holding technical committee seats, a distributional question this chapter's survey, which asks about awareness rather than bargaining power, is not designed to detect.
Chapter 13: A Framework for Algorithmic Auditing in India
Harsh Lailer, Gadamsetti Srija, Aseem Saxena, and Agrima Lailer assemble the conceptual vocabulary, available methods, and India-specific harm evidence needed for a dedicated algorithmic auditing framework. Algorithmic auditing is defined as systematic and rigorous examination of algorithms in automated decision-making systems for fairness, accountability, transparency, and legal and ethical compliance. Impact assessments are framed not as impartial measuring instruments but as co-constructed accountability relationships aligned with actual harms, and algorithmic accountability is treated as a connected explanation across lifecycle stages with formal and informal forums capable of judgment.
The methodological review is the chapter's practical contribution. It examines the six-step auditing protocol developed for a hiring technology company, covering contracting, client survey, data collection, model development and evaluation against employment selection guidelines, deployment, and longitudinal back-testing. It sets out five audit designs including code audit, non-invasive user audit, scraping audit, sock puppet audit, and crowdsourced audit, each with documented legal and validity limits, noting that proprietary claims block code audits and that platform terms of service constrain several other methods. System-level disclosure is measured through an index of 100 indicators across upstream, model-level, and downstream subdomains, scoring ten named developers. India-specific harms are catalogued across six risk areas and mapped to axes of unfairness covering caste, gender, religion, ability, class, gender identity and sexual orientation, and ethnicity, with proxies such as surname, skin tone, and neighbourhood.
The framework is labelled suggested but never identifies the institution that would conduct or compel an audit, and its governance section defers entirely to unspecified national institutions with authoritative powers, leaving unresolved the access problem its own methodological review exposes. Two extensions follow. First, ISO/IEC 42001 from Chapter 12 offers precisely the missing institutional carrier, yet this chapter never maps its three-layer audit blueprint onto standards, so an India-only instrument could not be recognised or reused transnationally. Second, the chapter's strongest evidence is that India's harms are caste, religion, and ethnicity specific, using proxies that a developer disclosure index cannot capture because it measures publication rather than outcome disparity. A credible Indian framework would therefore need locally validated proxy suites and caste-disaggregated evaluation data, an infrastructure demand the chapter identifies but does not cost.
Chapter 14: Rwanda's Mbaza AI-Chatbot Project
Lea Gimpel and Keegan McBride argue that Western challenge typologies miss the contextual factors determining whether public sector AI works, and offer Rwanda's Mbaza chatbot as evidence from sub-Saharan Africa. Mbaza originated in a 2020 development hackathon inspired by a German counterpart, hosted with European Commission support, and was selected from more than 1,000 proposals as one of 20 ideas. The government's three stated problems were limited access to reliable information in rural areas without smartphones, an overburdened call centre, and widespread misinformation. Distribution language was decisive: roughly 95 percent of Rwandans speak Kinyarwanda, so English-only delivery would have excluded most of the population. The method is a qualitative exploratory single case study using nine semi-structured interviews and inductive coding.
The deployment data reward attention. The rules-based USSD version launched at the start of the third pandemic wave in July 2021; by September 2021 more than 580,000 people had used it with around 15,000 interactions daily, and by April 2022 there were more than 2,200,000 unique users. The AI components lagged. The Kinyarwanda speech recognition model, trained on 2,300 hours of voice data and 3 million validated text lines, started at a 60.1 percent word error rate and 23.5 percent character error rate, later reported down to 39 percent and still far from adequate, with its model card restricting use to keyword spotting and simple transcription. The discussion maps findings onto seven categories and adds problems absent from the literature, including turf wars between agencies, slow government ownership, data-sharing resistance, legal grey areas over voice data and personally identifiable information, localisation requirements conflicting with foreign cloud servers, and procurement inflexibility after an emergency-style start.
The chapter's headline success metrics measure the delivery layer rather than the AI. The 580,000 users and 2,200,000 unique users belong to the rules-based channel that launched first, while the authors' own evidence shows the Kinyarwanda semantic chatbot still had loopholes, speech synthesis needed to sound more natural, and the speech recognition model was explicitly not fit for use as a voice assistant at a 39 percent word error rate. Mbaza therefore demonstrates that a rules-based public service can scale in Rwanda, not that public sector AI did. Read against Chapter 8, the point sharpens: India's approach buys language capability as a shared national API layer, whereas Rwanda had to build corpora first and lost its technical partner mid-project when a volunteer speech programme was discontinued, which is exactly the South-South corpus pooling remedy Chapter 9 proposes and neither chapter operationalises.
IN PRACTICE
1. Write the evidence requirement before you write the requirement: Most public AI failures begin with a tender that asks for capability instead of evidence. Waters and Miller show that IEEE P3119's value lies in its process steps, particularly problem definition before solicitation. Action: draft the problem definition section of one active or planned AI acquisition, stating what public problem exists and what evidence would show an AI system solves it.
2. Make training data provenance a contractual condition: Chapter 5's most damaging finding is not a performance number but a refusal. Vendors withheld training data provenance, annotation method, and ungradable image handling from a public health validation, and the authors had no mechanism to compel disclosure. Action: add a tender clause requiring dataset provenance, annotation protocol, and prior external validation, with disqualification for non-compliance.
3. Separate the four audit layers and staff each one: Chapter 13 treats impact assessment, technical audit, accountability, and redress as connected layers where each output feeds the next. Auditing that stops at technical testing leaves harmed citizens without recourse. Action: for one system in your environment, name the owner of each of the four layers and record which layer is currently unfunded.
4. Score risk by layer and publish the weights: Chapter 7's scoring framework is only as legitimate as its weights, and its own example moves the scaled score from 2 to 4.5 for identical criteria. Action: if you use any weighted AI risk score, publish the weights, the justification, and the review process, and record who has authority to change them.
5. Test whether your transparency instrument changes behaviour: Chapter 4 recommends publishing registers and redress pathways without measuring whether they alter decisions. Action: pick one published transparency artefact in your organisation and define one metric that would show it changed an outcome rather than documented a process.
6. Ask what happens after training ends: Chapter 2's data architecture secures the training pipeline but says nothing about post-deployment inference or drift. Action: require one monitoring clause per AI contract covering live performance, data drift, and the tolerance threshold that triggers intervention, with a named responsible party.
7. Treat language coverage as infrastructure, not a feature: Chapters 8 and 14 both show that language capability determines who can use a public service at all, and that building it after the service exists is costlier. Action: before approving any citizen-facing AI service, list the languages it supports and compare that against the languages your population actually speaks.
QUIZ: TEST YOUR UNDERSTANDING
1. Waters and Miller argue that procurement is the decisive governance stage for public AI. Which single artefact do they treat as the operative policy instrument?
Answer: The acquisition or solicitation document, including the request for proposals and the resulting contract. The chapter's claim is that these documents define required evidence, performance obligations, and failure consequences more concretely than any published AI principle.
2. Chapter 5 reports sensitivity from 59.69 to 97.74 percent across three masked algorithms. What does that spread imply for a procurement officer?
Answer: That vendor accuracy claims cannot be compared without validation in the intended setting. The same class of commercial system produced radically different real-world performance, and one invited algorithm was dropped after interim analysis showed extremely low specificity.
3. Chapter 3 fine-tuned eight firm-specific models on the same base model and found correlated false negatives between firms. Why does this matter more than average accuracy?
Answer: Because correlation means many people are wrongly rejected by every firm at once. Individually accurate systems can produce a worse collective outcome than the pre-automation baseline, which is a systemic harm no single firm's accuracy figure reveals.
4. Chapter 2 secures the training pipeline with digital contracts, clean rooms, and differential privacy. Apply that architecture to a different context than the book presents, such as a municipal benefits agency sharing applicant data with a research partner, and identify what remains unprotected.
Answer: What remains unprotected is everything after model issuance. Differential privacy is a guarantee about the training dataset, not about what the finished model later infers about people who never contributed data, so the agency would still need post-deployment monitoring, drift detection, and a remedy path for harms arising from inference.
5. Chapter 14 reports more than 2,200,000 unique users for the Mbaza chatbot. Why do the authors' own findings qualify that number?
Answer: Because those users belong to the rules-based USSD channel, not the AI components. The speech recognition model was still at a 39 percent word error rate and explicitly unfit for voice assistant use, so the project demonstrates that a rules-based service can scale, not that public sector AI delivered at scale.
How many did you get right? Which ones surprised you?
CAN THIS BOOK REPLACE THE ORIGINAL?
No summary replaces this book. The Essentials cannot substitute for reading the original: each chapter's value lies in its specific method, its numbers, and its candid reporting of what did not work, including vendor refusals, incomplete assessments, and speech models that were not yet usable. Read the full volume for the case detail and the chapter-level argument; use this Book Essential to decide which chapters matter to your work.
QUOTES
"certainly greater consideration should be given to adapting procurement processes when AI-enabled solutions or ADS are making critical decisions for the populace"
"a right to an effective remedy (Article 47 in the European Union's Charter for Fundamental Human Rights) can potentially be violated"
"The P3119 standard includes five processes that help users identify, map, treat, and monitor risks commonly associated with high-risk AI systems"
"Trained on limited publicly available datasets of X-ray images (less than 50,000)"
"By far, the most common is using Differentially Private Stochastic Gradient Descent (DP-SGD)"
"a small number of primarily closed-source foundation models are establishing leadership in the generative AI market"
"the off-the-shelf model has almost no predictive power for this screening task, leading to uncorrelated errors across firms"
"the data in the registers is either sparse or non-existent"
"The commercial AI algorithms' validation steps are unavailable in the public domain."
"efforts to obtain similar transparency from AI algorithm companies were futile"
"The conversations made it clear that most problems did not need an AI solution."
"the total score for a risk profile should be scaled within the range of maximum and minimum possible score"
"it covers 10 of India's 22 official languages and 171 of approximately 20,000 government programs"
"In 2022 private investments in AI totaled to $91.9 billion, out of which $47.36 billion was invested in the US alone"
"Algorithmic auditing is a systematic and rigorous examination of algorithms, usually implemented in automated decision-making systems"
AUTHOR'S EXPERTISE
Abhishek Singh is a distinguished officer of the 1995 batch of the Indian Administrative Service and serves as Additional Secretary in the Ministry of Electronics and Information Technology, Government of India, with additional charge as Chief Executive Officer of the IndiaAI Mission. He holds a Bachelor of Technology in Mechanical Engineering from IIT Kanpur and a Master's in Public Administration from the Harvard Kennedy School as a Mason Fellow. He has served as Chief Executive Officer of Karmayogi Bharat, MyGov, and the National e-Governance Division. His leadership contributed to CoWIN, DigiLocker, and DIKSHA, and he previously served in the Governments of Nagaland and Uttar Pradesh on grassroots development schemes, law and order, and state revenue collection.
Balaraman Ravindran heads the Robert Bosch Centre for Data Science and Artificial Intelligence and the Centre for Responsible AI at IIT Madras. He is the Mindtree Faculty Fellow and Professor in the Department of Computer Science and Engineering, and he heads the Department of Data Science and Artificial Intelligence and the Wadhwani School of Data Science and Artificial Intelligence. He received his doctorate from the University of Massachusetts Amherst and a Master's degree from the Indian Institute of Science, Bangalore. He is an elected fellow of the Association for the Advancement of Artificial Intelligence and the Indian National Academy of Engineering, and an ACM Distinguished Member. The Centre for Responsible AI he founded at IIT Madras is India's leading interdisciplinary centre for responsible AI research and has launched capacity-building programmes for government officials in partnership with industry and civil society.
The volume's contributors span more than ten countries and include medical professionals, government officials, civil society actors, private enterprises, and academics. Notable contributors include Roberto V. Zicari of the Trustworthy AI Lab and Z-Inspection, whose assessment process is applied in Chapter 10, and researchers from the Post Graduate Institute of Medical Education and Research Chandigarh, who led the diabetic retinopathy validation in Chapter 5. The book's style is deliberately institutional rather than promotional: chapters report ethics approval numbers, trial registrations, survey response counts, error rates, and refusals of cooperation. The editors' preface notes that the call for papers received more than 150 submissions from 38 countries, of which 11 were shortlisted after review by a committee drawn from academia, industry, and government.
RESOURCES
The book itself: https://www.taylorfrancis.com/books/oa-edit/10.1201/9781003663577/advancing-responsible-ai-public-sector-application-gpai-balaraman-ravindran-abhishek-singh (CRC Press, 2026, edited by Abhishek Singh and Balaraman Ravindran).
The standard examined in Chapter 1: https://standards.ieee.org/ieee/3119/10729/ .
The research centre led by the co-editor: https://cerai.iitm.ac.in/ .
Government framework referenced in Chapter 12: https://www.nist.gov/itl/ai-risk-management-framework .
Open volume DOI record: https://doi.org/10.1201/9781003663577 .
NEXT STEPS
Demand provenance before performance: Ask where the training data came from, how it was annotated, and who validated it externally. A refusal is itself a finding, and Chapter 5 shows what a refusal costs a public health programme.
Put the remedy in the contract: Redress does not appear after deployment unless it was written into the agreement. Name the mechanism, the responsible party, and the response time before signature.
Weight the score in public: Any risk score is a policy choice expressed as arithmetic. Publish the weights and the review process, or accept that the same technology will be scored differently by successive administrations.
Fund all four audit layers: Impact assessment, technical testing, accountability, and redress each need an owner and a budget. Auditing that stops at the second layer produces reports without consequences.
Check the language before the model: Language coverage determines who can use a public service at all, and it is far cheaper to settle before procurement than after deployment.
Record the negative result: The most useful chapters in this book report what failed, what was withheld, and what remained unfinished. Write your own failures down where colleagues can find them.
U365'S RECOMMENDATIONS TO LEARN MORE
This section curates the most useful external resources on responsible AI in the public sector: standards bodies, primary research centres, recorded lectures, and serious analysis. Every link below was verified live as of 2026-09-22.
Official learning resources
Video tutorials and channels
University lecture on responsible AI governance as a core public policy challenge, covering how AI systems shape access to public services and risk assessment, by IE University, Jan 26, 2026, 32:29
Introduction to the centre founded by the book's co-editor, covering its interdisciplinary research mission for responsible AI in the Indian research system, by WSAI IITM, Oct 22, 2024, 5:58
Explainer on responsible AI principles and governance frameworks, useful as a primer before reading the standards chapters, by IBM Technology, Jan 8, 2025, 15:10
Written tutorials and deep-dive articles
Community and social
Centre for Responsible AI, IIT Madras on X : the centre's own channel, running the CeRAI Conversations series
IndiaAI Mission on X : official account of the national AI mission led by the book's co-editor as CEO
Community discussion: the Z-Inspection initiative maintains open documentation of its trustworthy AI assessment process at z-inspection.org
Resources on X
Dedicated X channels:
X posts with video content:
This list was assembled by an exhaustive search across standards bodies, primary research centres, peer-reviewed and institutional publications, recorded lectures, and practitioner communities. Community and social sources are labelled so you know their provenance. Promotional and affiliate material was excluded. Every link returned HTTP 200 at the time of review; 403 responses from bot-blocking are retained.
IMPORTANT NOTICE
This Book Essential is an original summary and critical analysis of Advancing Responsible AI in Public Sector Application, GPAI Edition, edited by Abhishek Singh and Balaraman Ravindran (first edition, CRC Press, 2026). Short quotations from the book are attributed and cited for purposes of criticism, review, and education. All rights in the original work belong to its editors and contributors and to the publisher; this Essential is not a substitute for the book: read the original at [Taylor and Francis](https://www.taylorfrancis.com/books/oa-edit/10.1201/9781003663577/advancing-responsible-ai-public-sector-application-gpai-balaraman-ravindran-abhishek-singh).
This book is part of University 365's learning library. Explore INSIDE, our publications, and our programs. The best summary is not a substitute for the book. Read the original. Discuss this book with a U.Coach.
Note: verify current rates before relying on this comparison. Information verified as of 2026-09-22.
Comments