ps-regulatory-and-ethical-guardrails-for-public-ai

Frames the distinct regulatory and ethical stakes of public-facing AI - higher transparency and accountability bar, disparate-impact risk, due-process concerns when AI touches decisions about individu

SKILL.md

GitHub

Regulatory and Ethical Guardrails for Public-Sector AI

Purpose

ai-strategy-and-governance:responsible-ai-and-governance-check covers the general regulatory, risk, and ethics dimensions of an AI initiative and is explicit that deeper EU AI Act compliance analysis needs separate regulatory expertise. This skill narrows that same posture to what changes specifically when the deployer is a public body: AI touching a citizen's access to a benefit, service, or legal status carries a higher transparency and due-process bar than the equivalent private-sector use case, and often a different regulatory classification. This is a framing and triage skill, not a compliance determination.

Method

1. Ask what kind of decision the AI touches

Sort the use case into one of three bands before anything else:

  • Back-office / internal efficiency — no direct effect on an individual citizen's rights, benefits, or treatment (e.g. internal document search, scheduling optimization). Lowest stakes band.
  • Citizen-facing but advisory — AI informs a decision a human still makes and can override (e.g. flagging a case for human review, drafting a response a caseworker edits and sends).
  • Citizen-facing and decision-affecting — AI materially shapes an outcome that affects a specific person's access to a service, benefit, or legal standing (e.g. eligibility scoring, risk flagging that triggers enforcement action, automated triage that limits which service track someone is offered).

The third band is where public-sector AI regulation and administrative due-process expectations bite hardest — flag it explicitly and treat it as requiring specialist regulatory review, not just this skill's checklist.

2. Check the transparency and explainability bar

Public bodies are generally held to a higher explainability standard than private companies for decisions affecting individuals — an opaque "the model said so" is much harder to defend when the decision affects someone's access to a public service and is subject to appeal or freedom-of-information requests. Ask: can this system's individual decisions be explained to the affected person and to an appeals body in terms a non-technical person can follow? If not, that's a design requirement to solve before launch, not a caveat to disclose after.

3. Screen for disparate impact before launch

Ask specifically who might be systematically disadvantaged by this system relative to the status quo — by language, disability, digital access, age, or any protected characteristic relevant in the jurisdiction — not just whether the system works well "on average." This overlaps with, and should be run together with, ps-community-and-equity-impact-assessment for the full picture.

4. Check for a human-in-the-loop requirement

For the decision-affecting band (step 1), default to assuming a human review/override point is required unless a specialist confirms otherwise — many public-sector AI regulatory regimes require this explicitly for higher-stakes automated decisions. If the design doesn't currently have one, flag it as a likely gap rather than a stylistic choice. For the mechanics of designing that human checkpoint well, use ../../../../human-ai-collaboration-design/ (this repo's core pack on HITL design).

5. Name the regulatory regime, don't guess its content

If the initiative is inside the EU, note explicitly that the EU AI Act's risk classification (and its stricter obligations for systems used by public authorities, particularly in areas like eligibility for public benefits, law enforcement, and migration) is likely relevant and needs confirmation from a qualified source — don't state a specific risk classification or compliance obligation from general knowledge. The same applies to national data-protection and administrative-law requirements. If a dedicated AI-regulation or legal skill resource is available separately, hand this question to it explicitly rather than answering it inside this skill.

What this skill does NOT do

  • Doesn't determine actual legal or regulatory classification — flags when that determination is needed and by whom.
  • Doesn't replace ../../../../ai-strategy-and-governance/skills/responsible-ai-and-governance-check/SKILL.md — narrows and sharpens it for the public-sector case.
  • Doesn't design the human-oversight mechanism itself — see the human-ai-collaboration-design pack for that.

Refinement notes

Written deliberately without citing specific article numbers or national statutes, since regulatory specifics change and vary by jurisdiction — the goal is durable triage logic (what to ask, when to escalate), not a compliance summary that goes stale.

Continue from here

  • General governance check: ../../../../ai-strategy-and-governance/skills/responsible-ai-and-governance-check/SKILL.md
  • Human-oversight design: ../../../../human-ai-collaboration-design/
  • Equity/community impact: ../ps-community-and-equity-impact-assessment/SKILL.md
  • Procurement/funding triage: ../ps-procurement-and-public-funding-navigation/SKILL.md

References

  • ../../references/source-notes.md

Reminder: frontmatter has only name and description. Everything else goes into skills_index.json (run scripts/generate_index.py).