We should require Canberra to publish redacted agent interaction logs before any national school AI pilot scales, because classroom chatbots are already advising teenagers on homework, wellbeing, and career choices without a public record parents can inspect when something goes wrong.
Why logs come first
Agent systems do not just retrieve facts; they chain tools, call APIs, and rewrite prompts mid-session. A log—timestamped intents, tool calls, policy blocks, and human escalations—lets auditors see whether a “math helper” quietly routed students to unvetted web sources or stored sensitive disclosures. Vendors pitch personalization; without logs, personalization is a black box funded by term fees.
State trials in Victoria and New South Wales have moved faster than federal privacy guidance. That is not inherently bad, but pilots without shared logging standards force each principal to negotiate data terms alone—a recipe for inequity between wealthy PTAs and regional schools that accept default vendor contracts.
The strongest objection
Industry lobbyists say publication risks exposing trade secrets or student identities. Both concerns are manageable: logs can redact names, mask proprietary model weights, and still reveal whether an agent attempted a barred action. Open publication is not the only model; a federal repository with researcher and parent access under OAIC rules would beat the status quo of “trust the dashboard.”
Defenders also claim logs will chill innovation. We answer that public money already chills innovation when districts fear litigation after an agent hallucinates a medical directive. Transparency reduces that fear by showing where guardrails fired and where humans intervened.
What agencies should mandate
The Department of Education and the eSafety Commissioner should require any federally funded pilot to ship logs in a common schema before expanding beyond term two. Procurement should score vendors on exportable audit trails, not slide-deck accuracy claims. ACARA-aligned assessments can wait; accountability cannot.
Parliament should fund independent red-teamers—not the vendor’s pen testers—to replay logged sessions quarterly and report to ministers. Schools already manage duty-of-care for excursions; agent logs are the digital equivalent of the permission slip.
Limits of this argument
We are not calling for a ban on classroom AI or for publishing children’s names. We are calling for a public mechanism of record before scale. If Canberra will not publish logs, it should not publish press releases claiming pilots are “safe by design.”
Parents deserve to know what the agent did in their child’s session last Tuesday. Without logs, “personalized learning” is just marketing—and the bill still arrives on the school fees statement.
Teachers unions added that staff cannot intervene in agent sessions they cannot replay; logging is a workload issue as much as a privacy one. A federal schema would let educators flag harmful outputs without learning each vendor’s export format.
Canberra has funded digital literacy for a decade; agent literacy without logs is incomplete. Publish the record, then scale the pilot.
Ministers should treat withheld logs the way they treat withheld NAPLAN item banks: acceptable for security, unacceptable as a permanent shield against scrutiny. Parents are not asking for model weights; they are asking whether an agent tried to book a counselling referral without a counsellor in the loop.
