Rio wrote AI into public services. The ethics committee needs a public docket

THE SIGNAL IN ONE SENTENCE

A resident should not need to become an algorithm detective to understand a government decision. If a computer system helps route a health request, flag a tax record, rank an inspection, translate a form or suggest that an application needs another document, the person on the other side deserves a basic receipt. What did the system do? Which information did it use? Who checked the result? What happens if it was wrong? Rio de Janeiro has now created a statewide framework that could make those questions easier to answer. Decree 50.523, dated October 13 and published in the state gazette on October 14, establishes the Política de Inteligência Artificial Fluminense, or PIAF. It also creates the Fórum de Inteligência Artificial Fluminense and the Comitê de Ética em Inteligência Artificial do Estado do Rio de Janeiro, known as CEIA-RJ. The state science and technology secretariat announced the policy on October 15 as a way to make service faster, simplify processes and improve public administration. The official Rio Digital page goes further. It describes a governance structure, agency responsibilities, data-protection duties, monitoring and an ethical impact self-assessment for public-sector AI. That is more substantial than a press release saying the future has arrived wearing a government badge. It is still a framework, not evidence that a particular system has improved a particular service. The plain signal is this: Rio has built the rooms where AI governance can happen. Residents now need a public docket that shows what enters those rooms and what comes out. The policy applies across the state executive branch. Its definition of an AI solution is broad enough to include applications that use AI in at least one feature, operation or stage. The decree covers development, contracting and use, which matters because public agencies often encounter an algorithm as a vendor product rather than a model they trained themselves. The governance map has several parts. The Secretaria de Estado de Transformação Digital acts as the central body. Proderj provides technical support. CEIA-RJ is a consultative technical body. The existing data-governance committee has a role, and each state agency remains responsible for implementation in its own domain. The forum has a different job. It is intended to bring together government, universities, civil society and industry. Its final reports and documents are supposed to be published on the Portal RJ Digital. Seats for universities, civil-society organizations and productive-sector institutions are to be filled through a public selection process that may consider diversity and representation. The ethics committee includes state digital, science, planning, control, legal, technical and research institutions, plus universities or scientific institutions. Its members are expected to make a public declaration that they have no conflict of interest and to observe confidentiality rules. The official policy page also says agencies should designate focal points, maintain a sectoral data and AI plan, organize the data they control, use anonymization and other safeguards, notify the government when AI systems enter production when institutional registration and monitoring are required, monitor performance and effects, answer committee inquiries and recommendations, and apply a public-sector ethical impact self-assessment where appropriate. When data processing may create a relevant risk to people's rights and freedoms, the policy calls for a personal-data protection impact report. A serious personal-data security incident triggers containment, mitigation and the legally required communications. Those are useful ingredients. They create named owners and recurring duties instead of treating responsibility as a mood. The missing bridge is public visibility at the level where a resident actually meets the system. The policy page describes institutional registration and monitoring. It does not point readers to a public, case-by-case register of deployed AI systems. The announcement does not list systems already in production, publish impact assessments, identify contracts or explain an appeal route for a person affected by an automated recommendation. That does not mean none of those records exist. It means the public-facing materials reviewed for this article do not yet make them easy to find. A public docket would fix that without requiring the government to publish source code, expose personal data or hand attackers a security manual. Each system could have one durable page. It should name the agency, service, purpose, owner, vendor and model or product version. It should say whether the system produces a draft, a recommendation, a risk flag, a ranking or a final action. It should describe the data categories used and the legal basis for processing them, while withholding personal records and genuinely sensitive security details. The page should also show the decision boundary. Can the tool translate a citizen's message but not alter a benefit? Can it prioritize a maintenance ticket but not close it? Can it suggest a fraud lead while a trained employee must independently review the evidence? These boundaries are more informative than the word assist, which can stretch from spell-checking to quietly deciding who waits longest. Testing belongs on the same page. Publish the groups and conditions used to check error rates, the metric that matters to the service, the date of the most recent evaluation and what changed after a failure. Where demographic testing is lawful and methodologically sound, show whether error rates differ across relevant populations. Where it is not possible, explain why and identify another way to test unequal impact. Then show the human route back. An affected resident needs a contact point, a deadline, a clear request for human review and a way to correct bad data. The employee reviewing the case needs authority to disagree with the system, enough time to do it and a record of the reason. An appeal button that sends the same information through the same model is not an appeal. It is a loading animation with ambitions. Public-sector AI is not one risk category. Using a model to summarize an internal meeting creates different stakes from using one to rank inspection targets. A translation aid creates different harm from a system that influences health access, policing, schooling, taxes or social support. Rio's docket should label those differences and require more evidence as potential harm rises. Low-risk administrative uses can still go wrong, but they can be tested quickly. A form classifier can be compared with trained staff. A scheduling forecast can be evaluated against actual demand. A translation assistant can be reviewed by speakers of the relevant language. A document search tool can be checked for missed records and unsupported answers. Higher-stakes uses need more than accuracy. They need a clear legal basis, documented alternatives, independent review, audit logs, security testing, bias analysis, staff training, public notice and an appeal process that works offline as well as online. Procurement is part of the ethics story. If a vendor will not disclose how versions change, what data leave the state environment, how incidents are reported, what logs are retained or how the government can export its records at the end of a contract, the public agency has bought uncertainty. The docket should link to the procurement record, contract value, renewal dates, relevant subcontractors and exit plan. This also protects competent vendors. Clear expectations let suppliers compete on evidence rather than on the thickness of a slide deck. A small local team with a narrow, testable tool may be a better public-service partner than a giant platform sold as a universal answer. Rio's geography and inequality make local testing especially important. A system trained on clean digital records can perform neatly for residents already well represented in government databases while stumbling over inconsistent addresses, shared phones, missing documents, unstable connectivity or language that does not resemble a template. Efficiency for the database can become friction for the person. The policy's forum could help surface those problems before deployment. Civil-society organizations, universities, public servants and residents can identify failure modes that a procurement team may not see. But participation needs receipts too: meeting calendars, agendas, attendance, submitted evidence, recusals, final recommendations and the government's response. CEIA-RJ is consultative. That makes publication even more important. If the committee cannot itself halt a system, the public should be able to see what it reviewed, what it recommended, which agency accepted or rejected the recommendation and why. Confidentiality may be necessary for personal data, active security work or protected procurement information. It should be an exception with a stated reason, not the default texture of the institution. A useful incident register can also be public without identifying residents. It can record the system, date, type of failure, services affected, severity, containment, corrective action and verification that the fix worked. Publishing near misses would be unusually valuable. They show whether monitoring finds problems before a resident pays the price. The first version does not need to be grand. Start with systems entering production after the decree. Add existing systems in stages. Give every entry a stable identifier. Let people search by agency, service, risk level and status. Keep retired systems visible so the history does not disappear when a contract ends. Publish machine-readable data for researchers and a plain-language page for everyone else. Then measure the policy itself. How many systems are registered? How many received an ethical impact assessment? How many personal-data impact reports were required? How often did human reviewers overturn a recommendation? How many residents appealed, how long did review take and how many cases changed? Which systems were paused, redesigned or retired? Those numbers will not prove fairness. They will show whether the governance structure is doing work. Rio has made an important move by putting data governance, monitoring, ethical review and cross-sector participation into one state policy. The decree recognizes that AI in government is not merely a software purchase. It is a change in how public power gathers information, recommends action and explains itself. Now the policy needs a front door. For a resident, responsible AI should not be an institutional diagram buried behind a service counter. It should be a public receipt attached to the system, a named person who can answer a question and a human route for fixing the result. The ethics committee has a room. Give the public the docket on the table.