A housing authority AI audit cannot obtain training-data samples because the vendor refuses access. What should the auditor pursue next?
Select an answer to reveal the explanation.
Short Explanation
Vendor locked the training data? Push access rights and plan Plan B tests — vendor descriptions don't close the gap. Production logs can't prove training-data representativeness, and scope limits still belong on the record. Asking for a different dataset is not the same as asserting access rights or designing alternative procedures.
Full Explanation
Asserting evidence access rights and designing alternative procedures is the correct next step when a housing-authority AI vendor refuses training-data access. Auditors should invoke contractual, regulatory, or engagement-letter rights and, where full access remains blocked, pivot to feasible tests such as output evaluation, bias probes on production-like inputs, monitoring review, or independent datasets. Silent surrender of access without alternatives or disclosure leaves municipal housing AI risk unevaluated.
Relying on vendor system descriptions to conclude on model fairness risks fails by concept: vendor descriptions are not independent evidence of the training data and cannot substitute for direct access or approved alternative procedures. Reviewing production deployment logs to infer training-data representativeness fails because logs reflect model outputs, not the data used to train the model, and any access limitation must still be disclosed to governance. Requesting a different, less sensitive dataset as a substitute does not establish the auditor's right to the actual training data, nor does it design alternative procedures like output testing or bias probes against the real production model; it just swaps one access request for another without resolving the refusal.
Exam caveat: alternative procedures may not fully replace training-data inspection; residual limitation often remains and must be reflected in conclusions. Operational check: document the access request trail, cite contractual or policy rights, execute approved alternative tests, and report any unresolved training-data access limitation affecting assurance over the housing AI system.