South Australia opened an AI Royal Commission to the public. Make every submission traceable

THE SIGNAL IN ONE SENTENCE

A public inquiry can collect an impressive pile of documents and still lose the plot. One submission describes a nurse correcting an AI transcript after every shift. Another predicts billions in productivity. A vendor promises safer public services. A teacher describes students outsourcing their thinking. An Aboriginal community asks who controls the data used to model its people and Country. All of those voices matter. They are not the same kind of evidence. South Australia's Royal Commission into Artificial Intelligence opened its first call for submissions on October 29. The accompanying scoping paper asks the public to help shape an inquiry across six themes: AI in society, better services, innovation, economic transition, education and skills, and safe and trusted AI. The Commission is chaired by former Federal Court judge and economist Iain Ross. It must provide advice and recommendations to the South Australian Government by July 1, 2027. The inquiry is future-facing. It is not principally investigating past wrongdoing. That distinction matters. This is a chance to decide what evidence South Australia will require before AI becomes ordinary infrastructure in hospitals, schools, emergency warnings, transport, workplaces and government services. The scoping paper already names practical tests under way in the state. It describes exploration of clinical documentation tools, AI-assisted drafting of bushfire warnings, and computer vision applied to existing traffic cameras. Those are not abstract chatbot debates. A bad clinical summary can move into a patient record. A confusing fire warning can cost precious minutes. A traffic model can direct attention toward one incident and away from another. The Commission needs more than a comment box. It needs a public evidence register. Start by assigning every submission a stable public identifier. Record who submitted it, what perspective they represent, what question they answered, what evidence they supplied, which interests could shape the claim, and whether confidential material was withheld. Then classify the claim. Lived experience is evidence of what happened to a person or community. It may reveal a failure that no benchmark was designed to notice. A controlled study is evidence under specified methods and conditions. An operational record can show what happened in a real service, but its missing fields may hide who was excluded. A forecast is a model of what could happen. A commercial claim is a statement from an interested party until independent evidence supports it. A policy preference is an argument about what government should choose. Do not flatten those into one bucket called feedback. Next, connect every claim to the authority that can act. South Australia can make decisions about state services, schools, hospitals, workplace policy, procurement, local regulation and state investment. The terms of reference explicitly place data centres outside the inquiry. They also exclude matters primarily belonging to the Commonwealth, including national security, financial market stability, critical national infrastructure, intellectual property and copyright. That boundary is useful if the register makes it visible. A submission about copyright should not vanish because it sits outside scope. Mark it outside state authority, explain where it was referred, and preserve the link. A concern about a hospital procurement should not be buried inside a national policy debate. Name the state agency with power to answer it. Public participation becomes frustrating when people describe a real harm and receive a polite sentence about jurisdiction. The register should show the route from concern to responsible institution. Conflict disclosure belongs beside the evidence, not inside a footnote. A vendor may have valuable deployment experience and a direct financial interest in rapid adoption. A union may hold detailed knowledge of workplace change and a duty to protect members. A university may supply independent research while also seeking public funding. An advocacy group may surface harms ignored by industry while representing a defined constituency. None of that disqualifies the submission. It tells readers how to interpret it. The same discipline should apply to commissioned expert work. Publish the question, selection method, funding, declared conflicts, underlying data where lawful, methodology, limitations and response from affected groups. The Commission's six themes are broad enough to invite almost everything. That is useful at the start. It also creates the risk that powerful sectors produce polished submissions while people living with the consequences submit a short story, if they submit at all. The scoping paper says the Commission plans public submissions, stakeholder consultations, industry roundtables and community forums. It identifies businesses, unions, developers, researchers, teachers, students, advocacy groups, creative industries and cultural groups among the people it wants to hear. The participation register should measure who actually arrived. Publish representation by region, age group, disability, employment sector and community role when privacy permits. Show which regional communities were reached and what support was available for people without reliable internet, technical language or time to write a formal document. Count is not enough. Ten submissions from a mining company and one conversation with workers do not create balance because both sides appear on a list. Record the depth, format and accessibility of engagement. Indigenous data sovereignty needs its own visible path. The scoping paper says the inquiry will consider Aboriginal knowledge, Indigenous data sovereignty and cultural governance of AI systems. That cannot become a paragraph pasted into the final report. Record whether Aboriginal communities defined the questions, chose the evidence conditions, controlled sensitive material, approved interpretations and can challenge reuse. Consultation is not sovereignty if the institution still decides what can be taken. Public services need service-level evidence. For every AI use case, publish the existing baseline before describing the future benefit. How long does clinical documentation take now? How often do bushfire warnings require correction? How accurately do current systems detect traffic incidents across daylight, weather, camera quality and different road users? Then publish the proposed measure, affected population, failure threshold, human override, complaint path and conditions that would stop the trial. Otherwise efficiency can become a magic word that means fewer workers, more correction and no public record of the trade. Workers deserve a task ledger, not a jobs headline. The inquiry should separate tasks removed, tasks accelerated, new review work created, skills required, bargaining impacts, surveillance risk and responsibility when the model is wrong. An assistant that saves ten minutes and adds twenty minutes of verification is not productivity. Automation that moves risk from an agency to a contractor is not a public saving. A model that improves average performance while making expert judgment harder to maintain creates a long-term capability cost. Education evidence needs the same specificity. Ask what students learn, what teachers can inspect, which ages are affected, how assessment changes, what happens when families cannot pay, and whether local and Aboriginal knowledge is represented with permission and context. Do not treat access to a chatbot as AI literacy. Literacy includes knowing when the system is guessing, who benefits from the answer, what data moved, how to verify a source, and when to refuse the tool. The Commission should publish a living contradiction map. When one submission says AI reduces workload and another says it creates invisible correction work, put them together. Name the difference in sector, task, model, training, measurement period and worker role. Contradiction is not a problem to smooth away. It is where the inquiry earns its keep. Every draft recommendation should point backward to the register. Readers should be able to see which evidence supports it, which evidence challenges it, what authority can implement it, what it will cost, who owns the next step and how success will be measured. Add a minority or unresolved-evidence note when the record does not justify consensus. Then keep the register alive after the final report. The Commission ends in 2027. The systems will keep changing. Each accepted recommendation needs an implementation date, responsible agency, budget, public metric, review point and stop condition. Each rejected recommendation needs a reason. Each deferred concern needs a destination. South Australia has opened a serious door. The public can now say what AI is doing to their work, services, classrooms, culture and communities. The next responsibility belongs to the inquiry. Do not turn those voices into a decorative consultation summary. Give each claim an address. Show the evidence behind it. Name the institution with power to respond. Then let the public follow the decision all the way home.