SC-AI-032 · Security and resilience

Logging and record keeping

Medium effortCovered by ISO 27001: StrongSatisfies 22 obligations in 12 policies

What it means

  • Keep enough records to reconstruct what an AI system did, when, with which version, and why:
    • decision and output logs (inputs or input references, outputs, model and prompt version, explanation, human reviewer),
    • provenance records for generated content (unique identifiers, labels applied),
    • governance records (risk assessments, validation and red team reports, approvals).
  • Protect these records from tampering and keep them for the required period, then delete.
  • Preserve evidence when content is removed or accounts are suspended.
  • ISO 27001 reuse: Partial. ISO 27001 logging, monitoring, clock sync and records protection cover the plumbing. New for AI: what to log (prompts, model versions, retrieved sources, explanations), content provenance identifiers and decision reconstruction.

Why it matters

  • RBI FREE-AI asks for red team findings to be documented for audit and supervisors, and expects audit trails across the model lifecycle.
  • The 2026 IT Rules require embedded identifiers for synthetic content and that accounts be suspended without destroying evidence.
  • MeitY's advisory expects identifiers that can trace synthetic content back to its source.
  • Disputes, complaints and litigation all depend on decision records.

Checklist

  1. Define per AI system what to log (inputs, outputs, model and prompt version, retrieved sources, tool calls, reviewer, explanation) and what not to log (minimise personal data).
  2. Send AI logs to your central logging or SIEM with synchronised timestamps.
  3. Make logs tamper-evident (write-once storage or hashing) and restrict access.
  4. Set retention per log type aligned with legal needs (for example RBI record keeping, DPDP minimisation, IT Rules evidence) (SC-AI-012 Data minimisation and retention for AI).
  5. Record unique identifiers of generated content so it can be traced back (SC-AI-023 Label AI-generated and synthetic content).
  6. Add legal hold: when content is removed or an account suspended, preserve related logs and content (SC-AI-035 Takedown and response timelines).
  7. File governance records (risk assessments, red team reports, approvals) in a controlled repository with version history (SC-AI-018 Robustness and adversarial testing (red teaming)).
  8. Test yearly that you can reconstruct a sample decision end to end.

Evidence examples

  • AI logging standard listing required fields per system.
  • SIEM or log platform screenshots showing AI log sources.
  • Retention schedule covering AI logs and governance records.
  • Legal hold procedure and a sample hold record.
  • Decision reconstruction test report.
  • Document repository showing versioned red team and validation reports.

Common pitfalls

  • Logging full prompts with personal data and keeping them forever.
  • Logging outputs but not the model or prompt version, so decisions cannot be reproduced.
  • Vendor AI tools with no log export.
  • Deleting suspended account data immediately, destroying evidence.

Satisfies

Also relevant to

Standards crosswalk

Clause references only. Check the standard itself for the exact wording.

ISO/IEC 42001

  • 7.5 documented information
  • A.6.2.7 technical documentation
  • A.6.2.8 event logs

ISO/IEC 27001

  • A.5.33 protection of records
  • A.8.15 logging
  • A.8.16 monitoring activities
  • A.8.17 clock synchronisation

SOC 2

  • CC7.2
  • CC7.3

Key takeaways

  • Log enough to reconstruct any AI decision, tied to model and prompt version.
  • Protect logs from tampering, keep them as long as the law needs, then delete.
  • Preserve evidence on takedowns and suspensions.

Find your gaps

Our free compliance gap assessment shows which of these controls you already run.

Start the gap assessment

Have us build it with you

Our AI Governance service implements controls like this one and produces the evidence auditors expect.

Explore AI Governance