Public AI That Can Be Appealed
An operating standard for automated essential services. An operating standard for public-sector AI built around evidence, notice, human responsibility, independent audit, and a meaningful right to challenge consequential decisions.
Research Question
Public institutions are already using artificial intelligence to tailor services, detect anomalies, support decisions, and forecast demand. The same systems can propagate error, obscure responsibility, discriminate, exclude people who cannot navigate digital processes, or scale an untested policy judgment faster than an institution can detect harm.
A rights-preserving operating standard must make consequential systems contestable. That means a person can know that a system materially influenced a decision, understand the decision in terms relevant to them, reach a responsible human with authority to correct it, and obtain a timely remedy. Technical transparency without institutional responsibility is not enough.
Findings
Source markers link findings to the evidence below.
OECD's 2026 outlook reports AI use in at least one government function in 35 of 36 surveyed countries and a public-sector AI governance institution in 30 of 36. Formal transparency standards, public algorithm registers, enforceable oversight, and robust impact measurement remain much less common. [1]
Across 200 government use cases reviewed by OECD, 57% streamlined or tailored services, 45% supported decisions or forecasting, and 30% pursued accountability or anomaly detection. The same review identifies rights infringement, error propagation, cyber risk, exclusion, and weakened trust as material risks. Efficiency evidence should never be treated as rights evidence. [2]
The OECD AI Principles call for meaningful information, human oversight, traceability, and the ability to challenge outputs. The UN Global Digital Compact adds human-rights due diligence, effective oversight, remedy, and human supervision. Together they point to a socio-technical duty spanning the model, data, policy rule, workflow, vendor, and accountable public authority. [3] [5]
NIST's AI Risk Management Framework organizes work as Govern, Map, Measure, and Manage. This supports continuous monitoring, incident response, change control, and retirement decisions rather than a one-time approval certificate before deployment. [4]
Institutional Implications
These are design implications drawn from the synthesis, not claims that one institutional model will fit every jurisdiction.
People should receive accessible notice when AI materially influences a consequential decision, the principal reasons and evidence used, a non-digital review route, a decision deadline, and correction of downstream records when an error is found.
Contracts should require documented purpose, validation limits, subgroup testing, audit access, incident disclosure, data and model change logs, portability of public records, and a safe termination path that does not trap the institution in an unreviewable vendor system.
A public register should name the responsible authority, decision context, affected groups, legal basis, vendor role, validation status, monitoring metrics, known limitations, appeal channel, incidents, and current deployment state.
Method & Limits
This rapid evidence brief synthesizes cross-government use evidence and governance guidance from OECD, NIST, and the United Nations. It draws operational implications from overlapping public standards, but it does not certify any system, product, or legal regime as compliant or safe.
Public-sector AI evidence is still emerging, use-case inventories are incomplete, and OECD country findings are not globally representative. Error and fairness metrics depend on the policy goal, legal context, population, and available ground truth. A meaningful appeal standard must be adapted to local law and developed with people directly affected by the service.
Source Record
OECD · 15 June 2026
Cross-country adoption, governance, transparency, and measurement findings.
OECD · 18 September 2025
Review of government use cases, benefits, risks, and governance capabilities.
OECD · 3 May 2024
Updated principles on rights, transparency, challenge, traceability, and safety.
National Institute of Standards and Technology · 26 January 2023
Operational risk-management functions and trustworthy-AI characteristics.
United Nations · Adopted 22 September 2024
Intergovernmental commitments on human rights, oversight, remedy, and AI governance.
Suggested citation: True Tomorrow Foundation (2026). Public AI That Can Be Appealed. Evidence Brief. Reviewed July 17, 2026.
