From principles to production: closing the AI assurance gap for government.

Most AI safety effort goes into models before they ship, but the International AI Safety Report is clear that general-purpose systems drift, degrade, and surface new failure modes after release, so assurance cannot stop at deployment. Drawing on conversations with Australian agencies working through the Policy for the Responsible Use of AI in Government and the National Framework for the Assurance of AI in Government, this talk argues that the hardest part of public-sector AI safety is not writing the policy, it is sustaining the evidence that a deployed system still behaves as intended: monitoring behaviour, drift, cost, and human oversight across the lifecycle, re-validating on material change, and being able to show it to a board, the DTA, the ANAO, or the public on demand. The takeaway is a technology-agnostic way to frame the operational assurance and reporting gap: the shift from "is the model safe?" to "can we prove, continuously, that it still is?".