Benefits

Admission

Cases concern reported or intended benefits with an identifiable AI contribution. These include AI-assisted learning, research, accessibility, and operational improvements. An AI-related topic alone is insufficient. Reported outcomes and intended benefits have distinct labels. Every case has a beneficiary shape, a depth color, and a confidence fill. Gray means that depth is unassessed.

Kinds of event

A record describes one of four kinds of case: a benefit created or restored, a specific harm prevented, a protective response whose prevention is unproven, or a general benefit case. General cases can include launches, simulations, accuracy studies, aggregate claims, cohorts, AI-assisted learning, and proposed applications. Their category does not determine their depth or confidence.

Depth

Each beneficiary group that benefited is rated from 1 to 5, the depth of the benefit for that group: 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. Every rating has a reason. Depth describes the magnitude of the benefit that the source reports; it does not establish that the claim is independently verified. Aggregate outcomes and gains in capability can be rated. Intended benefits remain explicit, and a group stays gray when the extent of its benefit cannot be established. Corrections appear in the revision history.

The system's part

What the system contributed to the benefit is assessed separately from whether the benefit occurred, and may be left unrated. A system that was present is not thereby responsible.

A system’s name, base model, adapted variant, interface, hosting, and settings are assessed separately. A supported inference appears with its basis in an editor’s note. Later documentation may identify a plausible model, but does not establish that it was available or used on the event date. The importance of a task does not establish its reasoning settings.

Evidence and confidence

Source review follows citations, notes, appendices, and supplementary material to the study, dataset, document, or firsthand account that produced each finding. Different claims in one case may have different primary sources. An additional source must contribute evidence, verification, a correction, or a material limitation. Repetition does not constitute corroboration. A primary source may have an interest in the result, so origin and independence are assessed separately. When the original cannot be inspected, the record identifies the intermediary and the limitation.

Each source identifies its kind and its author's relationship to the system: maker, deployer, beneficiary, evaluator, independent author, or unassessed. An evaluator's role alone establishes neither independence nor a conflict of interest.

Only sources that support the benefit determine its evidence grade. Among independent sources, a court or regulatory finding gives the grade found. A peer-reviewed study, clinical case report, or research report gives studied. A descriptive report, news report, preprint, or dataset gives reported. The strongest qualifying source determines the grade. Without one of these sources, the grade is claimed. This includes accounts from interested parties, even when published as studies, and sources whose independence is unassessed.

The fill is solid for found and studied, hatched for reported, and outlined for claimed. These correspond to high, medium, and low confidence. Contrary evidence, procedural documents, and background material remain in the sources without raising the benefit's grade.

Source notes describe authorship, independence, and any documented conflicts of interest. They identify gaps in the cited evidence. Inferences are labeled.

Briefs

Briefs by Peter H. Vartanian examine the evidence and its gaps. Their dates appear in the revision history.

Replication

Each entry includes fields for the conditions that produced the benefit and the conditions that would have to hold again. Subsequent attempts appear with their setting, model version when reported, and outcome. An outcome that the cited sources do not establish remains unresolved.

Comparable incidents

Where an incident database holds an entry about the same kind of system harming the same kind of stakeholder, the record links to it, so that a reader weighing benefit against risk can open both. No severity is shown for the harm.

Exploring the cases

Cases can be filtered and viewed as a list, table, map, timeline, or table of benefit types. The view and the filters are kept in the address, so that a view can be shared.

Candidate reviews

The review log gives the reasons for excluding a candidate. Unfavorable results remain visible when they bear on a reported benefit. Multiple sources about the same benefit belong to one entry. Intended benefits, aggregate results, and weak evidence can qualify.

Dates

A case distinguishes an observed event date from the publication date, the date through which a source says that its information is current, or the date on which the source was checked. Month-only dates display the month without an invented day. The timeline uses that date only when an event or source supports it. Dates of access and revision do not determine the order.

Newest follows event or source dates. When a case's date basis has not been assessed, a dated source must support the same date before the case appears in Newest. That date is labeled as a source publication date. Source-check and revision dates do not determine the order. Cases without a supported date remain in the full database.

Coverage

Counts on this site describe the database's coverage. They do not measure how beneficial AI is, and a trend in records is a trend in reporting.

The dataset is one file, records.json, released under CC BY 4.0. Three flat tables are derived from it: records.csv, one row per record; ratings.csv, one row per beneficiary group rated; and echoes.csv, one row per attempt to make a benefit happen again. The schema describes every field. Each record is also available on its own page as JSON.

Feed of new and revised cases

Required coverage

Each case has 96 required field groups: 91 public and five internal editorial groups. Six collection-wide definitions apply through versioned collection profiles. Every field carries a value or an explicit assessment status. Unassessed questions remain visible and incomplete. Explore the fields and their meanings.

Download the public case profiles, field dictionary, and collection profiles. The coverage table lists each public field and its status for every case.

Fructus

AI outcomes, in money. A separately curated collection of financial claims, with original currencies, periods, recipients, comparisons, and supporting sources. JSON and CSV.

Counting

One system, one setting, one benefit is one record. A benefit reported in several sources is one record with several sources. A benefit that happened again elsewhere is an echo on the original record, not a new record.

Fields

id

A stable identifier.

title, summary, date_observed, account

What happened, in one line and in one paragraph, and when.

brief

A brief from Peter H. Vartanian: his reading of the evidence and its gaps, where one has been written, with its sidenotes in brief_notes.

headline, metric

A short title that adds no fact (asterisks around a name set it in italics), and the AI's quantified effect, copied verbatim from the summary or account (null when the record states no number).

date_basis, date_precision

Whether date_observed identifies an event, source publication, source information cutoff, or source access. Precision is day, month, year, or unassessed. A partial date uses the first day of its month or year only as a sorting anchor.

benefit_status

Reported, intended, mixed, or unclear. This is separate from evidence confidence.

category

The kind of event: benefit created, harm prevented, protective response (prevention unproven), or general benefit (including aggregate results and proposed applications).

unexpected, purpose

Whether the benefit was what the system was built to do, and what it was built to do.

system, model_version, developer, deployer, location, domain, segment

The system, its version where known, who made and ran it, and where it was used.

benefit_type, taxonomy_note

The benefit type from the taxonomy, with a note where the fit is loose.

unrated_because

For an event with no depth, the reason in a clause: the good is denied, or the AI's part is undocumented.

ratings

For each beneficiary group that benefited, a depth from 1 to 5 or null for unrated, and the reason.

ai_role

What the system contributed, assessed separately, or marked not assessed.

evidence, confidence

The sources include their kind, citation, link, date, a short quote, and author relationship: maker, deployer, beneficiary, evaluator, independent author, or a relationship that remains unassessed. Only sources that support the benefit determine confidence (found, studied, reported, or claimed). Shade (high, medium, or low) corresponds to a solid, hatched, or outlined fill. The method explains these grades. An evaluator's role alone establishes neither independence nor a conflict of interest.

comparator

Better than what: the baseline the benefit is measured against (against), the figure without and with the AI (before, after), whether the source measured or assumed the difference (basis), and the passage that states it (quote, url). Null when no comparison is recorded. The case profile distinguishes an unassessed question from a reviewed absence.

later

What happened after the first report: persisted, scaled, challenged, reversed, withdrawn, litigated or unknown, with the date, a note, and the passage that states it. A challenge disputes a claim without establishing that it was reversed. Unknown means that no later outcome is established in the record.

value

A financial benefit stated by a source: the amount verbatim with its currency, what it measures, who gains or pays, whether it is an estimate, and the supporting passage.

claim_status, claim_basis

The strongest source's authority, the cited evidence by kind and author, and its limitations. Both are computed from the evidence.

replication

The conditions that produced the benefit, the conditions that must hold again, and the echoes (replications): each attempt to repeat it, with setting, model version, and outcome.

paired_incident

Harm on record: an entry in an incident database about the same kind of system harming the same kind of stakeholder, with the database's name and a one-sentence basis for the link.

source_collection, source_entry

The imported collection and the stable entry key that AIBD assigns within it. These entry keys are not publisher-assigned identifiers.

registry_ids

Identifiers of the same event in other registries.

revisions

Every change, with its date, source, and editor.

AI Benefits Database is made by Peter H. Vartanian. Write to peter.vartanian@icloud.com.

The dataset and this site are released under CC BY 4.0. Benefit types and beneficiary groups follow Trustible's benefits taxonomy.

Set in Newsreader, IBM Plex Sans and IBM Plex Mono, all used under the SIL Open Font License. Sector icons are Bootstrap Icons, under the MIT License. Organization logos are the organizations' own marks, shown to identify them; each one's source is listed in logos.json. Built 25 October 2026.