A housing authority rejects opaque AI advice for benefit eligibility without human review. Which responsible AI stance is that?

When a housing authority rejects opaque AI advice for benefit eligibility without human review, the responsible AI stance is that high-impact eligibility guidance needs human oversight rather than opaque, unreviewed AI conclusions. Eligibility decisions affect housing access and resident livelihoods. Copilot or a chatbot may summarize policy text, but conclusions that steer who qualifies should be reviewable, explainable to staff, and owned by a human decision process—not delivered as an unexamined black-box result.

Disclosing AI use and citing policy sources for applicant transparency reviews fails oversight: disclosure shows what was referenced, not that a human verified the conclusion applies correctly to this resident's case. Evaluating eligibility models for fairness across demographic groups before deployment fails accountability for this decision: pre-deployment fairness review does not replace human sign-off on an individual case, and group-level fairness can still miss an incorrect determination for this applicant. Requiring that sensitive eligibility data remain protected with access controls during model processing fails the oversight concept: access controls protect data privacy, but they do not replace human accountability for high-impact eligibility outcomes.

Exam caveat: AB-900 focuses on responsible AI principles and when human oversight is required for high-impact scenarios. It does not require configuring access controls during model processing or configuring a full eligibility rules engine. “High-impact + human oversight” is the durable pattern for benefit, enforcement, and similar civic decisions assisted by Copilot.

Operational check: Classify housing-eligibility prompts as high-impact, require staff confirmation before communicating eligibility outcomes to residents, keep an audit trail of who approved AI-assisted guidance, and prohibit publishing unreviewed model conclusions as final determinations.