AI Benefits Database / Method

Fields and evidence

What happened, for whom, with which system, and on what evidence.

Each entry has 91 public fields. Five further fields concern editorial work. The answers appear on the entry itself, with source links and a precise note where an inference needs explanation.

Some questions remain open. A missing model version, for example, is not evidence that no version existed. The fields distinguish an unchecked question from information that a reviewed source does not provide.

91public case fields

5internal editorial fields

6collection definitions

Seven ways to state what is known

Recorded value: A value is recorded with source, derivation, or curatorial provenance. This status alone does not certify truth or causality.
Not reported: The identified sources were checked for this question and do not report the answer. Record the source scope and review date.
Unresolved: Available reviewed evidence does not settle the answer. Preserve the competing possibilities and their evidence.
Not assessed: The question remains unreviewed. It remains a gap and cannot satisfy completed curation.
Not applicable: The question has no meaningful referent for the case or related object. Record the scope reason. Source silence is insufficient.
Withheld: A value is known internally but is restricted in the public view. Preserve authorized internal access and disclosure rules.
Source unavailable: A relevant identified source could not be inspected. Record the source, limitation, access-attempt date, and follow-up status.

A value, its evidence, its review status, and the realization of the benefit are separate. A recorded claim does not establish that the outcome occurred or that AI caused it. Unassessed questions cannot count toward completed curation.

The public case fields

Case identity and dates 10

Stable AIBD identifier: Preserve the existing identifier and URL through corrections.
Kind of case: Individual experience, deployment, program/activity, evaluation/study, or another defined unit. Link components of the same project.
Related cases and counting relationship: Link the same deployment, evaluation, replication, duplicate, or incident. Do not count additional articles as additional benefits.
Title: Name the activity and claimed or observed benefit plainly.
Short account: State what happened, for whom, and the main qualification.
Sector and application domain: Support sector and task filtering without forcing one exclusive category.
Intended purpose: Distinguish the deployment's purpose from a benefit that emerged unexpectedly.
Dates and precision: Separate occurrence start/end, first publication, database discovery, added date, and last checked date. Allow unknown dates and day/month/year precision.
Stage of use: Proposed, simulated, laboratory test, pilot, routine use, scaling, retired, or unknown. This is separate from benefit evidence.
Record version: Identify the exact published version that a user cited.

AI system and use 12

AI product or system: Identify the application actually used, including an in-house system. A vendor name alone is insufficient.
Underlying model: Record the model behind the application when known. Allow several components with different roles.
Exact model version: Preserve the AI model version, checkpoint, release, or snapshot that was used in this case. Record hardware models under actual configuration and paper revisions under access and source version. A product name, device generation, publication revision, or current repository commit does not establish the deployed AI model version.
Model or system provider: Distinguish the model provider from the application developer and organization operating it.
Role in the system: Identify generation, retrieval, classification, planning, verification, perception, or another component function.
Actual configuration: Record the hardware model and application or interface version separately from the AI model version. Include wrappers, retrieval, fine-tuning, tools, inference budget, agent coordination, and weight or API access when they materially affect the result. Distinguish a published reproducibility configuration from a verified deployment snapshot.
Inputs and outputs: Describe data types, languages, provenance, and outputs relevant to the task, subject to access restrictions.
Autonomy and permissions: Record what the AI could decide or execute, tool permissions, and escalation or approval boundaries.
Human and organizational workflow: Explain who used the AI, what they did before and after it, and which decisions remained human.
Human review and safeguards: Review steps, responsible roles, override options, training, and controls in the actual deployment.
How the benefit relates to AI: Direct AI use, AI-related education, AI-related convening, institutional activity, another identified relationship, or unresolved. Record the actual route to benefit.
AI contribution and alternative explanation: Distinguish AI's contribution from human work, ordinary digitization, or other simultaneous changes. The contribution may be unassessed.

Organizations and geography 6

Actors and roles: Identify developer, model provider, deployer, operator, evaluator, funder, and beneficiaries where relevant. A single actor can hold several roles.
Places with defined roles: Separate development, compute/hosting, physical site, deployment, beneficiary location, actor headquarters, and applicable jurisdiction. Include place resolution and source.
Ownership or effective control: Record ownership or control only where documented and relevant. Headquarters does not establish nationality, state control, or allegiance.
Physical and digital dependencies: Link hosting/compute, chips, connectivity, energy, water, suppliers, data access, and locations when material to use or replication.
Cross-border relationships: Record sourced flows of data, compute, services, funding, or value, with origin, destination, and relationship type.
Geopolitical interpretation: Specify the country or institution whose interests are being assessed, security/economic domain, assessed capability change, rationale, confidence, sources, and date.

Benefits and beneficiaries 11

Outcome identifier: Give each distinct stakeholder-specific outcome a stable identifier within the case.
Benefit statement: State the particular claimed or observed change. Retain multiple outcomes without duplicating the deployment count.
Benefit category: Use a versioned benefit category and preserve source wording where it differs.
Outcome status: Intended, observed, mixed, not demonstrated, reversed, or unclear. Distinguish status from confidence and deployment stage.
Beneficiary class: Personal, organizational, or societal. Apply the existing shape grammar to the relevant outcome.
Specific beneficiaries: Describe the people or organizations that benefit, rather than stopping at an abstract class.
Beneficiary count and reach: Count actually benefiting, merely exposed, eligible, and intended populations separately. Preserve unit, period, estimation basis, and overlap limits.
Distribution and access: Record relevant differences by subgroup, geography, income, language, disability, or access conditions when evidence supports them.
Duration and persistence: Distinguish a one-time result from recurring or sustained benefit and state the observation horizon.
Expected or unexpected: Expected, unexpected, partly unexpected, or unknown. Do not infer intention from the presence of a positive outcome.
Costs, burdens, and harms to stakeholders: Record costs or harms to the same beneficiaries and to other parties, with timing and evidence. Link relevant incident records where available.

Measurement and comparison 11

Evaluation identifier: Identify an evaluation independently of its reports and the outcomes that it measures.
Evaluation design: Randomized, quasi-experimental, controlled observational, before/after, qualitative evaluation, case study, simulation, testimonial, forecast, or other. Record the actual design.
Measurement method: Instrument, administrative data, observation, interview, survey, model estimate, or another method. Preserve qualitative findings without forcing a number.
Setting and sample: Population, eligibility, sample size, sites, recruitment, allocation, and study period where reported.
Measured outcome: Metric name, exact operational definition, value, unit, numerator, denominator, and period. Retain source wording and distinguish proxy from ultimate outcome.
Comparator and baseline: No AI, prior workflow, human-only work, another AI system, another intervention, or no comparator. Record the exact baseline and comparison period.
Difference from the comparator: Absolute and relative effects with their units, uncertainty interval, calculation, and source. Do not combine incompatible measures.
Increment attributable to AI: Assess changes in cost, time, quality, scale, access, or capability relative to the alternative. Separate estimated contribution from total observed improvement.
Uncertainty and limitations: Record sampling uncertainty, confounding, missing observations, selection, failed comparisons, and bounds on causal interpretation.
Exposure denominator: Population at risk or eligible to benefit, uses, agent-hours, task-hours, transactions, distance, or another matched denominator, with source and period.
Reproduction materials: Protocols, logs, prompts, action traces, code, data, and evaluation artifacts with model/configuration and permitted use.

Monetary value and costs 11

Kind of monetary value: Cash saving, labor capacity freed, revenue, profit, recovered money, avoided loss, recipient transfer, willingness to pay, or another stated valuation. Preserve these distinctions.
Amount: Store the numerical figure with its original quoted amount. Distinguish an outcome level from an incremental gain. Zero differs from unknown.
Currency and price basis: Original currency, price year when relevant, nominal/real basis, and any documented conversion rate/date. Retain the original amount after conversion.
Period and unit: One-time or recurring, period covered, per-user/per-task/per-organization/total basis, and population or volume to which the estimate applies.
Realization and estimation status: Observed cash flow, measured resource saving, estimated monetary value, forecast, or target. A time saving is not automatically a cash saving.
Who benefits and who pays: Identify the recipient of value and each cost bearer. Separate organizational savings from worker or societal effects.
Gross or net: Gross, net, or unspecified. State which cost categories the estimate includes or excludes.
Cost components: Purchase, integration, training, infrastructure, operation, maintenance, human review, compliance, transition, and other documented costs, with time basis.
Economic comparator: Specify the alternative against which costs or gains are measured. Preserve differences in quality, volume, and time horizon.
Calculation and assumptions: Show the source's method, rates, unit conversions, discount rate if applicable, uncertainty, and sensitivity. Label any AIBD calculation explicitly.
Return or payback: Retain a source-reported ROI or payback period, or a transparent calculation with matched inputs. State numerator, denominator, and horizon.

Sources and claim evidence 10

Source identifier: Identify each report or evidence artifact separately from the case.
Citation and exact locator: Title, author/publisher, URL or persistent identifier, publication date, page/paragraph/table, and a short supporting excerpt when permitted.
Source language and translation: Original language, translation status/method, and original wording needed to verify a consequential claim.
Source relationship and interests: Maker, deployer, beneficiary, evaluator, independent reporter, funder, other, or unclear relationship. Unknown does not default to independent. Record disclosed funding/conflicts separately from research design.
Shared evidence lineage: Identify reports that repeat the same press release, study, witness, or dataset. Several articles may constitute one evidentiary origin.
Access and source version: Access date, source version, archive/persistent link when permitted, corrections or retractions, and availability limits.
Precisely supported claim: Tie evidence to the exact factual assertion, outcome, actor attribution, or AI-contribution claim. Publicly identify source statements, analyst inferences, and generated summaries.
Supporting and contradicting evidence: Link the exact passage and distinguish support, contradiction, and contextual evidence.
Evidence assessment: Assess occurrence, outcome, AI contribution, and actor attribution separately where relevant, with reasons and an explicit unassessed state.
Missing information: Distinguish not reported, unknown after review, not assessed, not applicable, withheld, and disputed. Do not substitute zero or false.

Ratings and interpretations 8

Assessment target: Identify the exact benefit and beneficiary to which the rating applies.
Depth of benefit: Keep the current depth scale and color mapping. Money, beneficiary count, confidence, and strategic significance remain separate dimensions.
Rating rationale: Explain why the observed or intended benefit fits the stated depth for that stakeholder.
Observed or intended basis: State whether the depth describes a realized, intended, projected, or unclear outcome. Assess support separately: a claim can concern a realized result. Keep unrealized benefits eligible.
Rubric and taxonomy version: Identify the definitions and rules under which the assessment was made.
Assessor and assessment date: Store who assessed the claim and when. Use a public role or pseudonym where disclosure requires it.
Disagreement and alternatives: Preserve materially different ratings, rationales, and unresolved questions rather than averaging them away.
Confidence shown by fill: Support for the beneficial outcome and its stated relationship to AI. Certainty that someone made a claim or that a benefit failed is separate. Retain unassessed. Do not infer confidence from publisher prestige alone.

Replication, follow-up, and decisions 8

Conditions needed to repeat the benefit: Skills, data, language, infrastructure, workflows, controls, and contextual prerequisites that evidence identifies.
Attempts to reproduce the result: Setting, population, model/version, protocol differences, observed result, source, and independent/same-team relationship. Include failed attempts.
Subsequent outcomes: Track persisted, scaled, mixed, reversed, withdrawn, or unknown outcomes and relevant changes in model or context.
Access and availability: Whether and how the system/activity can be accessed, procurement or license conditions, cost, geographic/language availability, and as-of date.
Institution and decision: Identify who was deciding whether to adopt, fund, scale, regulate, or stop a use, and the task that evidence served.
Decision grounds and alternatives: Relevant mandate, jurisdiction, threshold, evidence considered, alternatives, reasons to act or not act, and constraints.
Action and observed consequence: Record what was done, who owned it, when, and what later evidence showed. Distinguish intended response from tested effectiveness.
What would show success: Preserve the measurable question that would establish whether the adoption or intervention worked, with unanswered follow-up questions.

Interoperability and publication 4

Other databases and identifiers: Link original collection entries, risk/incident records, studies, registries, and standards without implying that they are equivalent records.
Versioned category mappings: Record exact, broader, narrower, partial, or no-match mappings with rationale and versions. Retain disagreement.
Citable case brief: Present the main result, comparison, beneficiary, material limitations, and source links. Preserve qualifications in export.
Field-level changes: Changed field, old/new values or recoverable versions, date, source, editor, and correction reason.

When a question depends on another record

Evaluations, monetary observations, strategic assessments, and institutional decisions can each have multiple records. Their fields are required when those records exist. A documented family assessment can establish that none applies. An unreviewed family remains open.

How the collections are assembled

AI Benefits Database

Collection scope: Reported and intended benefits of AI, including individual experiences, experiments, and deployments. The collection also retains proposals, disputed findings, and uses whose contribution from AI remains unresolved.
How candidates were found: Candidates came from searches by subject, source citations, and reviews of the ITU AI for Good publication and the Local Government Association case-study bank. The retained research files do not provide a complete log of queries, search dates, languages, or rejected search results.
Counting unit and duplicate rules: One entry per AIBD identifier. Several sources about the same benefit remain together. Studies, individual experiences, and aggregate reports all occur in the database; an entry count is therefore not a count of people or deployments.
Coverage limits: The collection is selective. Countries, languages, and sectors were not sampled uniformly, and the total number of eligible benefits is unknown. Intended benefits, contested findings, and unfavorable results may appear. Source availability affects what can be assessed.
Collection and review cadence: Entries are added or revised as research proceeds. Revision histories identify changes to an entry; individual source reviews have their own dates. There is no fixed refresh interval for every entry.
Population and exposure context: The count describes entries in AIBD. It is not a success rate: the database does not enumerate every deployment, attempted use, or person exposed to AI.

Fructus

Collection scope: Monetary statements attached to AIBD entries. Published statements require a checked primary source or primary-source excerpt, a supported connection to AI, and a matching entry. Value created or saved and money paid are shown separately.
How candidates were found: Candidates were drawn from monetary figures in AIBD and reviewed against their sources. Unresolved candidates and contextual amounts remain in separate review lists. Their presence there does not qualify them for the published collection.
Counting unit and duplicate rules: The headline count is the number of published monetary statements, not the number of AIBD entries. One statement may contain several amounts. A converted or source-equivalent amount is another expression of the same amount, not additional value.
Coverage limits: Payments, savings, avoided losses, and modeled values have different boundaries. The collection does not estimate the total economic benefit of AI. A payment connected to an AI-assisted project does not establish how much value AI created.
Collection and review cadence: Publication follows source review. Pending candidates and contextual amounts retain separate statuses. Currency conversions use the exchange-rate dates shown with the amounts; they are not live quotations.
Population and exposure context: Each monetary statement retains its own period, unit, recipients, and accounting boundary. The amounts are not summed across entries. Missing costs or counterfactuals prevent a net-benefit or return calculation unless separately established.

Querist

Collection scope: AIBD entries with court or regulator material, legal proceedings, challenges, reversals, or withdrawals. Inclusion identifies scrutiny; it does not mean that a court or regulator confirmed an AI benefit.
How candidates were found: Selection uses the source types and later developments recorded in AIBD. Identified trial registrations, catalog entries, and reposts that do not establish substantive scrutiny are excluded. A reviewed regulator announcement can document a decision.
Counting unit and duplicate rules: Each AIBD identifier counts once in this collection. Several sources or beneficiary groups do not create additional entries. Collections overlap, so their counts cannot be added.
Coverage limits: This is a view of AIBD, not a comprehensive register of litigation, approvals, or adverse findings. The underlying legal document may concern a drug, organization, or outcome rather than the AI system itself.
Collection and review cadence: Membership is recalculated when the site is published. The evidence is reviewed separately; a new publication of the site does not mean that every source was checked again.
Population and exposure context: The count describes entries in AIBD. It is not a success rate: the database does not enumerate every deployment, attempted use, or person exposed to AI.

Viator

Collection scope: AIBD entries in which at least one recorded case jurisdiction differs from the developer's jurisdiction. Entries can involve several case jurisdictions.
How candidates were found: Selection compares the developer country with the countries identified in the recorded location. Country aliases are normalized. An unknown developer jurisdiction or an unresolved case location cannot establish a cross-border relationship.
Counting unit and duplicate rules: Each AIBD identifier counts once in this collection. Several sources or beneficiary groups do not create additional entries. Collections overlap, so their counts cannot be added.
Coverage limits: A difference between jurisdictions does not establish a cross-border transfer of data, computing, funding, or ownership. The view depends on the geographic information recorded in AIBD and may omit deployments whose locations remain unresolved.
Collection and review cadence: Membership is recalculated when the site is published. The evidence is reviewed separately; a new publication of the site does not mean that every source was checked again.
Population and exposure context: The count describes entries in AIBD. It is not a success rate: the database does not enumerate every deployment, attempted use, or person exposed to AI.

Adyt

Collection scope: AIBD entries classified as a specific harm prevented or a protective response. A protective response can qualify even when prevention has not been demonstrated.
How candidates were found: Selection follows the two protective categories assigned in AIBD. The collection retains the recorded comparison as measured, assumed, or undocumented.
Counting unit and duplicate rules: Each AIBD identifier counts once in this collection. Several sources or beneficiary groups do not create additional entries. Collections overlap, so their counts cannot be added.
Coverage limits: A warning, intervention, or repair does not by itself prove that harm would otherwise have occurred. The records retain that distinction and any reported costs or harms. The collection is not an inventory of every safeguard in use.
Collection and review cadence: Membership is recalculated when the site is published. The evidence is reviewed separately; a new publication of the site does not mean that every source was checked again.
Population and exposure context: The count describes entries in AIBD. It is not a success rate: the database does not enumerate every deployment, attempted use, or person exposed to AI.

Chanticleer

Collection scope: Security, threat detection, and the operation of essential services. Topics include cybersecurity, defense and borders, energy and water, disasters, public and environmental security, financial security, and communications and transport.
How candidates were found: Entries were selected individually from AIBD according to their purpose and the activity described. Membership is explicit; a passing mention of security or a threat does not qualify an entry.
Counting unit and duplicate rules: Each AIBD identifier counts once in this collection. Several sources or beneficiary groups do not create additional entries. Collections overlap, so their counts cannot be added.
Coverage limits: The collection describes reported or intended uses. Inclusion does not endorse an operation or establish that its benefits outweighed its harms. Evidence about prevention, proportionality, and affected people varies by entry.
Collection and review cadence: Membership is recalculated when the site is published. The evidence is reviewed separately; a new publication of the site does not mean that every source was checked again.
Population and exposure context: The count describes entries in AIBD. It is not a success rate: the database does not enumerate every deployment, attempted use, or person exposed to AI.

Collection fields

Collection scope: Define countries, languages, sectors, dates, source classes, and inclusion rules actually searched.
How candidates were found: Queries, languages, channels, sampling or ranking rules, referrals, search dates, and updates. Separate submitted from actively sought material.
Counting unit and duplicate rules: State whether a count describes cases, deployments, outcomes, sources, evaluations, or beneficiaries and explain matching/overlap.
Coverage limits: Explain selective publicity, language/geography gaps, unavailable private data, and source bias beside aggregate comparisons.
Collection and review cadence: Distinguish source collection frequency from review backlog and record freshness.
Population and exposure context: Store matched denominators and their collection rules where available. A benefits collection alone cannot establish a success rate.

The five internal editorial groups cover publication terms, review status, information gaps, inclusion decisions, and capture quality. Their contents are excluded from public downloads.

Download

Cases and public profiles, field definitions, collection profiles, and the field coverage table.

Financial observations also have a dedicated collection: Fructus - AI outcomes, in money.