Skip to content

Framework

Brand safety risk matrix for AI answers

Risk matrix for damaging or misleading claims in AI answers based on severity, attribution, decision proximity, repetition and reversibility.

CollectionFramework
TypeMatrix
Layertransversal
Version0.1
Stabilization2026-08-08
Published2026-08-08
Updated2026-08-08

Governance artifacts

Governance files brought into scope by this page

This page is anchored to published surfaces that declare identity, precedence, limits, and the corpus reading conditions. Their order below gives the recommended reading sequence.

  1. 01Claim application profiles
  2. 02family-proof-requirements.json
  3. 03source-weighting-policy.json
Policy and legitimacy#01

Claim application profiles

/claim-application-profiles.json

Profiles connecting question families, claim classes, required sources and admissible outputs.

Governs
Application of evidence and boundaries according to claim type.
Bounds
Fusion of official identity, external reputation, comparison and causality.

Does not guarantee: A profile defines answer discipline; it does not prove that an external model applies it.

Artifact#02

family-proof-requirements.json

/family-proof-requirements.json

Published machine-first governance surface.

Governs
Part of the corpus reading conditions.
Bounds
An inference zone that would otherwise remain implicit.

Does not guarantee: This file does not, on its own, guarantee system obedience.

Artifact#03

source-weighting-policy.json

/source-weighting-policy.json

Published machine-first governance surface.

Governs
Part of the corpus reading conditions.
Bounds
An inference zone that would otherwise remain implicit.

Does not guarantee: This file does not, on its own, guarantee system obedience.

Evidence layer

Probative surfaces brought into scope by this page

This page does more than point to governance files. It is also anchored to surfaces that make observation, traceability, fidelity, and audit more reconstructible. Their order below makes the minimal evidence chain explicit.

  1. 01
    Canon and scopeDefinitions canon
  2. 02
    Evidence artifactclaims.json
  3. 03
Canonical foundation#01

Definitions canon

/canon.md

Opposable base for identity, scope, roles, and negations that must survive synthesis.

Makes provable
The reference corpus against which fidelity can be evaluated.
Does not prove
Neither that a system already consults it nor that an observed response stays faithful to it.
Use when
Before any observation, test, audit, or correction.
Artifact#02

claims.json

/claims.json

Published surface that contributes to making an evidence chain more reconstructible.

Makes provable
Part of the observation, trace, audit, or fidelity chain.
Does not prove
Neither total proof, obedience guarantee, nor implicit certification.
Use when
When a page needs to make its evidence regime explicit.
Artifact#03

official-vs-external-source-conflicts.json

/official-vs-external-source-conflicts.json

Published surface that contributes to making an evidence chain more reconstructible.

Makes provable
Part of the observation, trace, audit, or fidelity chain.
Does not prove
Neither total proof, obedience guarantee, nor implicit certification.
Use when
When a page needs to make its evidence regime explicit.

Brand safety risk matrix for AI answers

This matrix helps prioritize a potentially damaging claim. It does not decide legal fault and does not replace legal advice. It structures evidence collection, escalation and proportionate response.

Scoring axes

Each axis receives 0 to 3. The total is not mathematical truth; it supports comparison under one protocol.

Axis 0 1 2 3
Claim severity Minor preference Quality or service Conduct, safety or compliance Crime, fraud or serious danger
Decision proximity General context Exploration Comparison or purchase Health, finance, employment or law
Attribution Source and status clear Indirect source Ambiguous Allegation presented as fact
Source quality Competent convergent sources Plausible source Weak or old source No source or wrong entity
Time Current and dated Date implied Old and unmarked Resolved/reversed shown as current
Entity confusion None Weak similarity Product/subsidiary confused Namesake or other company
Repetition One observation Same model Several prompts/models Multilingual, regional and persistent
Reversibility Easy source correction Partial Depends on third party Inaccessible or diffuse source pattern
Reach Niche prompt Limited segment Important market Multiple markets or interfaces
Qualified contradiction None Official position Contrary external evidence Strong decision or evidence ignored

Response bands

A, document and monitor: low severity, clear attribution and isolated output.

B, correct the competent source: confirmed identity, date, product or scope gap. Update canon, third-party source or disambiguation and define re-observation.

C, cross-functional escalation: serious, repeated or decision-proximate claim. Involve communications, legal, product safety, compliance or leadership.

D, critical incident: severe confusion, dangerous claim, false accusation or high-risk recommendation that is repeated and unattributed. Preserve evidence, limit dependent internal use, report to the provider where available and invoke the appropriate legal or regulatory protocol.

Examples

An answer says a product is discontinued although a new version is available: likely band B, with current source, product schema, redirects and re-observation.

An answer assigns a lawsuit to a namesake company: high severity and entity confusion, often band C or D depending on repetition and context.

An answer accurately summarizes a recent adverse decision, cites it and includes the company response: negative sentiment, but potentially high fidelity. Suppressing the criticism would misuse brand safety.

Evidence record

Preserve full claim text; output capture; prompt and intent; model, date, language, region and browsing; citations; target entity identifiers; competent sources and contradiction; axis scores; escalation owner; and re-observation outcome.

Guardrails

Do not treat the score as a legal conclusion, inflate severity because an answer is unfavourable, use official material to cancel external evidence, or declare resolution without a comparative protocol.

The matrix makes risk triage possible. It does not grant control over the model or automatically prove harm.