Governance & Policy

Establishes lifecycle management standards for all data assets used to train, validate, and operate OIG AI systems, including lineage tracking, retention schedules, and access controls. Aligns with OMB M-25-21 data documentation requirements and NIST AI RMF Data function.

• All training datasets must have a documented Data Card before model registration.
• Data lineage must be traceable from ingestion source to model training artifact.
• PII-bearing datasets require Privacy Impact Assessment (PIA) approval prior to use in model development.
• Annual data quality audits are mandatory for all Tier 1 (high-stakes) AI systems.

Defines rules for the identification, classification, masking, and disposal of Personally Identifiable Information encountered in AI training pipelines and model outputs, consistent with the Privacy Act of 1974 and OMB Circular A-130. Prohibits use of unmasked PII in non-production environments.

• PII fields must be tokenized or pseudonymized before inclusion in any model training dataset.
• AI-generated outputs that re-identify individuals require mandatory human review before dissemination.
• Access to PII-containing model artifacts is restricted to personnel with current Privacy Act training certification.
• Breach response procedures must be initiated within 1 hour of suspected PII exposure in an AI pipeline.

Requires that all AI-assisted decisions presented to investigators or external stakeholders be clearly labeled as AI-generated, accompanied by confidence levels and explainability outputs, and subject to documented human review before any enforcement action is taken.

• All AI-generated risk scores must display the contributing model version and inference date.
• SHAP or LIME feature attribution is mandatory for any output classified as high-stakes.
• Public-facing reports derived from AI analytics must include a standardized AI disclosure statement.
• Investigators must complete an attestation log entry before acting on any AI recommendation above the high-risk threshold.

Extends the CISA Traffic Light Protocol with AI-lifecycle overlays covering output provenance, PII/CUI sensitivity layering, role-based access mapping, and temporal reclassification triggers, ensuring consistent information-sharing controls across the OIG's internal and interagency AI workflows.

• TLP:RED designation applies to raw inference outputs from models trained on CUI datasets.
• All TLP:AMBER AI outputs must include provenance metadata (model ID, version, inference timestamp).
• Temporal reclassification review is required at 90-day intervals for any AI artifact tagged TLP:RED or TLP:AMBER.
• Role-based access matrices are maintained in the AI Inventory and enforced via RBAC controls in the model-serving infrastructure.

Establishes mandatory human-in-the-loop checkpoints for all consequential AI-assisted decisions, defines override and escalation authorities, and requires auditable sign-off logs tied to individual investigator credentials for any enforcement referral originating from an AI recommendation.

• No enforcement referral may be submitted based solely on an AI recommendation without documented human review.
• Override decisions must be logged with a rationale code within 24 hours of the review action.
• Senior investigator countersign is required for any high-risk (≥0.80 score) fraud flag before case opening.
• Quarterly human-review audits assess calibration between model recommendations and investigator override patterns.

Defines permitted and prohibited use cases for large language models and generative AI tools within OIG operations, establishes output-validation requirements for any generative AI content entering official records, and prohibits the submission of CUI or PII to external generative AI services without an approved data-processing agreement.

• Generative AI tools may be used for draft summarization and report structuring only; final products require human authorship attestation.
• Submission of CUI, PII, or law-enforcement sensitive data to non-FedRAMP-authorized generative AI services is prohibited.
• All generative AI outputs used in official reports must be flagged with a standard disclosure tag and retained in the case record.
• An approved tool registry will be maintained listing generative AI services that meet OIG data-handling standards.