Editor’s guide

Version 1, 29 October 2026. Each revision of this guide is listed at its end.

Definition

An AI benefit is a good that a person, an organization or society has received, or is intended to receive, to which an identifiable AI system contributed.

A record enters the database when its sources name the AI system, the beneficiary and the good, and when they state what the system did toward that good. An AI-related topic alone does not qualify. Education about AI, conferences, partnerships, funding announcements and ordinary digitization do not qualify solely because they concern AI.

Each record states whether its benefit is reported or intended. A benefit is reported when a source describes it as having happened. It is intended when a source describes it as planned or expected, and it keeps that label until a source reports a result. Each record also states the stage of use at its date: proposed, simulated, tested in a laboratory, piloted, in routine use, scaled up, or retired. A launch, a simulation or a proposal therefore stays distinct from a benefit that people received.

Kinds of record

A record describes a benefit created or restored, a specific harm prevented, a protective response whose preventive effect is unproven, or a general benefit with no single event, such as a program or an aggregate result. The kind of record does not determine its depth or its evidence grade.

From candidate to record

Candidates come from searches by subject, from the sources that other records cite, from published collections of case studies, and from contributions. Every candidate with a source becomes a record, or an entry in the review log that gives the reason for leaving it out. Several sources about the same benefit belong to one record, and a repeat of the same benefit elsewhere is noted in the record that it repeats.

Sources

Each claim is traced to the work that produced it: the study, the dataset, the official document or the firsthand account. Different claims in one record can have different origins. A further source is kept only when it adds evidence, an independent check, a correction or a limitation. Repetitions of one work, such as a syndicated article or a second format, count as one source.

Each source records its author’s relationship to the system. The makers are its developer and the developer’s staff and funders. The deployer is the organization that used it. An evaluator is commissioned by the makers or the deployer. A source is independent when no author has any of these roles and no author is the beneficiary. A study with one author from the developer or the deployer is not independent. A news report counts as independent support only when the outlet reports something itself, such as its own interviews or observations.

Quotations are copied from the source and checked against it. When an original cannot be read, the record names the intermediary and the limitation.

Evidence grade

The grade comes from the sources that support the benefit. An independent court document or regulatory finding gives the grade found. An independent peer-reviewed study, clinical case report or research report gives studied. An independent descriptive report, news report, preprint or dataset gives reported. Without any of these, the grade is claimed, which means that only the parties to the benefit support it. Contrary evidence, procedural documents and background material stay in the sources and do not raise the grade.

Dates

A record shows the date when its benefit happened whenever a source states it. Without that date, it shows the first available of the following: the start of the deployment whose benefit is described, the date of an official finding or evaluation, the first public report, and the publication date of the cited source. The record says which kind of date it shows and to what precision. A date that rests on a publication is marked ≤.

Depth

Each beneficiary group that benefited is rated from 1 to 5: 1, a convenience, or time or cost saved, with life otherwise unchanged; 2, a measurable improvement in a result or a capability, such as a test score or a crop yield; 3, a lasting change, a capability restored or a risk removed; 4, a serious harm averted or a serious condition resolved in time; 5, a life or a livelihood preserved, or a change documented across a population. A rating describes the outcome that the sources report as realized, with its reason. An intended or undocumented outcome stays gray, and the record says why. Depth says nothing about how well the benefit is verified; the evidence grade does that.

The system’s part

What the AI system contributed is assessed apart from whether the benefit happened. The contribution counts as assessed when a source isolates it, for example through a comparison with and without the system. A system that was present is not thereby responsible for the result.

Comparisons, later outcomes and repeats

A record notes what its benefit is measured against, when a source states it, and whether that baseline was measured or assumed. It records what happened after the first report: whether the benefit persisted, scaled, was challenged, reversed, withdrawn or litigated, or whether nothing later is known. Attempts to repeat the benefit elsewhere appear with their setting and outcome.

Comparable incidents

A record links to an entry in an incident database when that entry concerns the same kind of system harming the same kind of stakeholder. The link helps a reader weigh benefit against risk, and it does not say that the benefit itself caused harm. When no incident meets both conditions, the record has no link.

Corrections

A correction, a further source or a new record can be sent through the contribution form or the public tracker. Peter H. Vartanian reviews each contribution before publication. Every change to a record appears in its revision history with its date and its source.

Revisions of this guide

Version 1, 29 October 2026: the first published version.