How Oiva uses artificial intelligence¶

Oiva uses AI to support educators and operators — for example, to suggest learning observations, summarise documents, and help users navigate the platform. Four commitments bound every AI invocation:

1. Inference in Australia. We use Amazon Bedrock with the AU-prefixed Anthropic Claude Haiku 4.5 model (au.anthropic.claude-haiku-4-5-20251001-v1:0) so that inference occurs in the AWS Sydney region (ap-southeast-2).

2. Personal information gate. Before every AI invocation, a Comprehend PII gate scans the prompt. The gate is fail-closed — if it cannot verify the prompt is safe, the invocation does not proceed.

3. Financial exclusion. Dollar amounts are excluded from AI prompts as a structural rule, not a policy choice. This is enforced by a Lambda assertion with 14 CI contract tests.

4. Human approval for compliance-affecting outputs. No AI output that affects eligibility, payments, or compliance outcomes is acted upon without explicit human approval.

Automated decision-making transparency. In accordance with the Privacy and Other Legislation Amendment Act 2024 (effective 11 December 2026), Oiva will publish the basis for any automated decisions that significantly affect individuals. Oiva does not currently make any fully automated decisions that affect eligibility, payments, or compliance outcomes without human oversight.