How AIBlindspot works

The AI Blindspot Framework

The framework comprises 8 categories and 48 blindspots, mapped to the R3AI parent framework (Reliable, Resilient, Responsible). Each blindspot is defined by:

• An executive question a board can ask
• A precise definition of the failure mode
• A representative case study
• Considerations and stakeholders
• Regulatory framework references (NIST AI RMF, ISO/IEC 42001, EU AI Act, and others)
• Detection signals used by the classifier

Categories are stable. Individual blindspots are dynamic: they can be added, merged, split, or deprecated as new failure patterns emerge from the live case base. The current taxonomy is v1.0 (48 blindspots, May 2026). Every published case study records the taxonomy version it was classified under.

Data sources

We ingest from credible OSINT sources, tiered by structure and reliability:

• Tier 1: AIID, AIAAIC, OECD AI Policy Observatory
• Tier 2: NewsAPI, GDELT
• Tier 3: ICO, FCA, EU AI Act tracker, FTC, SEC EDGAR, NHS Digital
• Tier 4: MIT Technology Review, Wired, The Guardian, arXiv, AlgorithmWatch

Every approved case carries source attribution and a working URL. PII redaction is applied before publication. No case is auto-published; every case is editorially reviewed against the standards below.

The Risk Index

The AIBlindspot Risk Index is the published share of severity-weighted incidents in each of the 8 categories. We model the share vector as a Dirichlet-Multinomial posterior with a symmetric Dirichlet(1) prior.

For category d, the posterior mean share is:

share_d = (1 + c_d) / (8 + Σ c_i)

where c_d is the severity-confidence-weighted count of incidents classified to category d. Each incident contributes (severity × confidence), not a unit count, so high-severity and high-confidence cases move the posterior more than minor or uncertain ones.

95% credible intervals are derived from the marginal Beta distribution Beta(1 + c_d, 7 + Σ_{i≠d} c_i). Trends are classified over a rolling 30-day window of stored snapshots.

Per-blindspot Risk Score

For individual blindspots, we publish a Beta(α=1, β=9) sceptical-prior Bayesian risk score. The prior is intentionally pessimistic so that a single incident does not lift an empty blindspot's score to a misleading value.

Editorial standards

Every approved case study must:

• Be attributable to a named organisation or identified system
• Have a credible, resolving source URL
• Have a published incident date after January 2015
• Map to at least one specific blindspot with confidence ≥ 0.35
• Pass the PII redaction step (no personal data of private individuals)
• Be readable by a FTSE 100 director in under two minutes

No case study auto-publishes. Every publish requires human approval.

Dynamic taxonomy promotion

New blindspots can be added once Bayesian evidence supports them. Cases that the classifier cannot match with confidence are routed to a misfit pool. Weekly clustering identifies recurring patterns. A candidate cluster is eligible for promotion to a named blindspot when:

• Posterior mean against Beta(1, 9) prior ≥ 0.65
• Weighted evidence count ≥ 10
• Time span ≥ 90 days
• Source diversity ≥ 4 distinct sources
• Mean intra-cluster cosine similarity ≥ 0.78

Eligibility is statistical. Promotion is editorial. The active taxonomy is soft-capped at 60 active blindspots; promotion of #61 requires retiring one.

Versioning

Patch versions (1.0.1): definition edits only. Minor versions (1.1): new blindspots, merges, or splits. Major versions (2.0): structural changes that require a published methodology note. Every classified case records the taxonomy version it was classified under, so historical analytics remain reproducible.