Generative IP Liability — building an underwriting model for AI content risk
22 pages · Insurance and underwriting · CIP Underwriting Working Group · Updated March 2026
The CARI risk scoring framework, GIPL policy architecture, and the actuarial considerations specific to AI content pipelines — including Training Data Dividend exposure, NILP Downstream Obligation tail risk, and platform compliance correlation.
CARI risk scoring framework
The Creative AI Risk Index (CARI) is a per-operator risk score computed across four layers:
- Layer 1 — Rights Payload: The percentage of an operator's training corpus that has documented rights coverage (CDR records, licences, or explicit consent). Higher coverage = lower risk.
- Layer 2 — Declaration Coverage: The completeness of the operator's cip.md declarations. Are all required fields populated? Are TDM opt-outs honoured?
- Layer 3 — Output Provenance: Does the operator maintain a complete provenance chain from input to output? Are Output-Provenance Blocks generated for all AI outputs?
- Layer 4 — Certification Status: What level of CIP Platform Certification has the operator achieved? Higher certification = lower risk.
GIPL policy architecture
The Generative IP Liability (GIPL) policy is a seven-section insurance product designed specifically for AI content risk:
- Section 1: Insuring clauses with per-layer trigger structure
- Section 2: Standard exclusions with deliberate-violation distinction
- Section 3: Conditions precedent at certification-track-specific levels
- Section 4: Claims notification accommodating discovery latency
- Section 5: Subrogation provisions cross-referencing the CIP Subrogation Framework
- Section 6: Definitions referencing the published CIP Glossary
- Section 7: Schedule with per-policyholder customisation
Actuarial considerations
AI content pipelines present actuarial challenges absent from traditional IP insurance: output volume at scale (hundreds of thousands of outputs per day), delayed discovery of infringement, correlated exposure across operators using the same training data, and regulatory uncertainty across jurisdictions.
The CARI score provides the primary rating factor. Empirical data from early-adopter operators shows a strong correlation between CARI score and claims frequency: operators with CARI above 70 have 3.2x fewer claims than operators below 40.
The Training Data Dividend creates a predictable, quantifiable exposure that can be reserved for actuarially. The NILP Downstream Obligation creates a tail-risk exposure that requires explicit loading in the policy premium — particularly for operators in the voice-clone and deepfake categories.