Responsible AI in the public sector: principles and practice

Reference article · 2 min read

Public agencies are adopting artificial intelligence for document processing, service request routing, forecasting, and resident-facing assistants. Because agencies exercise public authority, their use of AI attracts accountability that private use does not.

Guiding frameworks

The NIST AI Risk Management Framework provides a structure for governing, mapping, measuring, and managing AI risk. Federal guidance for agency AI use, state executive orders, and emerging legislation add requirements for inventory, impact assessment, and transparency. Requirements are evolving, and agencies should track their state's current guidance.

Principles that recur

• Human accountability. A person remains responsible for consequential decisions.
• Transparency. The public can learn where AI is used and how.
• Fairness. Systems are tested for disparate impact and monitored after deployment.
• Privacy. Data used to train and run systems is handled under existing privacy law.
• Security. AI systems are subject to the same security controls as any other system, plus AI-specific risks such as prompt injection.
• Reliability. Systems are validated before use and monitored for degradation.

Practices

Inventory. Know where AI is used, including features embedded in purchased software.

Use case review. Classify use cases by risk. Higher-risk uses, such as those affecting benefits, enforcement, or safety, get formal impact assessment.

Documentation. Model cards or equivalent documentation for each system, in plain language.

Procurement. Contract language requiring vendors to disclose AI use, provide testing evidence, and support audits.

Staff policy. Clear rules for staff use of generative AI tools, including what data may be entered.

Public disclosure. An AI use register or equivalent.

Where to start

Inventory and staff policy are quick and address immediate risk. Use case review and documentation follow. Agencies do not need to solve everything before benefiting from AI; they need to be able to answer basic questions about what they are doing and why.