Explainability

Explaining what an AI system decided and why.

The EU Court of Justice ruled in February 2025 that people scored by automated credit systems are owed meaningful information about the logic, and trade secrets are no blanket excuse. The US went the other way: the CFPB withdrew its AI adverse-action guidance in May 2025.

Features extracted by interpretability research: Anthropic, Claude 3 Sonnet, 2024 34; Anthropic, Claude 3.5 Haiku, 2025 30; OpenAI, GPT-4, 2024 16. Sonnet figure not confirmed this edition FEATURES EXTRACTED BY INTERPRETABILITY RESEARCH Anthropic, Claude 3 Sonnet, 2024 34 Anthropic, Claude 3.5 Haiku, 2025 30 OpenAI, GPT-4, 2024 16 MILLIONS OF FEATURES SONNET FIGURE NOT CONFIRMED THIS EDITION
Prompts where attribution graphs explain well
~25%
CFPB guidance documents withdrawn
67
EU Annex III start date
Dec 2027
Registry incidents
0

Evidence

Courts strengthened the right to an explanation as US guidance was withdrawn

Documented cases: explainability
DateCaseCountryWhat happenedSource
Feb 27, 2025Dun & Bradstreet Austria, C-203/22EUCourt can review trade-secret material itself; complexity is no excuseSource
May 12, 2025CFPB withdraws AI circularsUnited StatesAdverse-action guidance for complex algorithms removedSource
Mar 27, 2025Anthropic attribution graphsUnited StatesTraced Claude 3.5 Haiku's internal stepsSource
Mar 1, 2022Algorithm recommendation provisionsChinaProviders disclose basic principles; users can turn off personalizationSource

Registry

No registry incident turns on explainability yet

The cases above are documented, but none yet meets the registry bar of a named developer, a failed control and a known finder.

Rules

The EU and China require explanations; US credit law requires reasons but not AI-specific ones

Rules and frameworks: explainability
RuleJurisdictionWhat it requiresStatus
EU AI Act Article 86EUExplanation of AI's role in high-risk decisionsWith high-risk timeline, Dec 2027
GDPR Articles 15 and 22EUMeaningful information about automated decision logicIn force
ECOA and Regulation BUSSpecific reasons for credit denialIn force; AI circulars withdrawn
Algorithm recommendation provisionsChinaDisclose principles; opt out of personalizationSince Mar 1, 2022
Colorado SB 26-189ColoradoPlain-language explanation within 30 days of an adverse outcomeEffective Jan 1, 2027

Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.

Record

Page Record

Datasets
ai_trust_evidence.csv (4 rows); registry incidents (0 linked)
Key sources
Dun & Bradstreet Austria, C-203/22
CFPB withdraws AI circulars
Anthropic attribution graphs
Algorithm recommendation provisions
Last verified
October 2026