Verifying International Agreements on AI: Six Layers of Verification
RAND working paper analyzing six largely independent technical and personnel-based approaches that could eventually help verify possible international rules on large-scale AI development and deployment.
Automated source check: failed. The last direct retrieval failed; the curated facts have not been automatically changed. Supplemental audited sources may be available below.
- Issuer
- RAND Corporation
- Jurisdiction / parties
- International research
- Instrument
- working paper
- Status
- published · as of 2025-07-24
- Published
- 2025-07-24
- Last source check
- 2026-10-31T07:30:18.124533+00:00
What the instrument says
Original-author research; not a current compliance regime.
No legal actor
The working paper creates no legal duty; it analyzes verification options for possible future rules.
RAND WR-A4077-1 summary and working-paper notice
What would need to be proven?
The following mappings are analysis. Candidate evidence is not proof that a vendor satisfies an obligation, and hardware attestation alone does not establish legal compliance.
- Claim to examine
- A specified rule is supported by sufficiently independent verification layers.
- Candidate evidence
- Rule specification, layer coverage analysis, control evidence, and adversarial tests.
- Possible approach
- Assess the stated rule and threat model rather than infer assurance from a control inventory.
- Assumptions
- States would grant access and cooperation required by the proposed layers.
- Limits and open questions
- A working paper is not a treaty or evidence of operational verification.
Original sources
- RAND original working-paper page
summary and document details
Published: 2025-07-24 · Retrieved: 2026-10-31T03:16:21Z
Manual source audit: 2026-10-31T03:24:41Z
Retrieval method: manual source audit
- RAND identifies the five authors, document number WR-A4077-1, and publication date 24 July 2025.
- The working-paper summary describes six largely independent verification approaches for possible interstate AI guardrails and notes technical research needs.
Audit limit: RAND labels this a working paper; it is not a treaty, legal obligation, or proof that the analyzed approaches are operational.
Observation history
This is Bill Commons’ retained observation history, not a complete history of the instrument. A changed source body requires review before its policy meaning is updated.
- 2026-10-31T03:24:41Z · registry record created
Curated registry entry created; this is not a source retrieval.
- 2026-10-31T07:30:18.124533+00:00 · source check failed
Source fetch outcome; body changes require human review and do not assert a policy change.