‘Polyanna policy’ – is NZ’s framework for AI use in government overly

Strip away the policy language and New Zealand’s AI governance story says one thing clearly: a framework that names the right values but carries no legal weight is functionally the same as no framework at all.

The press releases say ‘transparency, fairness, human oversight.’ The fine print says ‘non-binding.’ That gap — between aspiration and accountability — is the story. And it’s not just a New Zealand story.

If you’re exploring what agentic AI can do — and thinking about which platforms and governance guardrails you can actually trust — the NZ case is a useful stress test for spotting the difference between real oversight and optimistic paperwork.

What New Zealand’s AI Framework Actually Says

New Zealand’s Public Service AI Framework names three core principles: transparency, fairness, and human oversight. Researchers at a recent International Research Society for Public Management conference named it a ‘Pollyanna policy’ — after the Pollyanna principle, which describes a cognitive bias toward assuming good outcomes.

The framework is explicitly non-binding. Individual government agencies are responsible for interpreting and applying it — agencies that vary enormously in technical capability and resources. The 2025 Public Service Census found that while roughly a third of NZ public servants had used AI for work, only 14% used it regularly. The aspiration and the operational reality sit far apart.

New Zealand’s broader National AI Strategy reinforces the same posture: existing laws are described as ‘technology-neutral’ and therefore sufficient, with no new AI-specific regulation introduced. The regulatory guidance essentially says ‘we won’t regulate.’ That’s a significant bet on optimism over structure.

Some frameworks look solid from a distance — but Beacon knows the difference between a guiding light and a decorative one.

The Robodebt Warning Nobody Wants to Hear

Australia’s Royal Commission into the Robodebt Scheme is the closest regional precedent. An automated debt-recovery system was deployed without clear governance or accountability structures — and produced what the Commission described as catastrophic harm to citizens. The system ran for years before scrutiny caught up with it.

The researchers behind the NZ analysis argue that algorithmic decision-making disrupts traditional accountability chains: information, justification, and consequences can no longer be traced through a single responsible person or process. When something goes wrong, nobody owns it.

This is the structural problem with non-binding frameworks. They don’t assign responsibility. They distribute it so broadly that accountability effectively disappears. A government agency that follows the spirit of the framework and still causes harm has no mandatory review mechanism — just more guidance to add to the pile.

Why This Isn’t Just a Government Problem

Everyone’s reading this as a story about government bureaucracy. We’d push back on that framing.

The same dynamic plays out in every organisation deploying AI agents without formal oversight roles. You adopt the right language — ‘responsible AI,’ ‘human in the loop,’ ‘bias review’ — but without anyone whose actual job is to interrogate AI output for errors and fabrications, those phrases are the same as New Zealand’s non-binding principles. Decorative.

There’s a compounding issue the researchers flag that’s directly relevant to AI agents: the sycophancy problem. The technology behind most AI agents — large language models — has a structural tendency to mirror the user’s existing views rather than challenge them. Ask a biased question, get a biased answer that sounds authoritative. In a policy system, that can reinforce institutional blind spots. In your personal or business workflow, it can quietly confirm bad decisions rather than catching them.

The Five Eyes security agencies addressed this directly in April 2026 joint guidance on AI agents, explicitly calling for incremental deployment, continuous threat assessment, and sustained human oversight. That’s five of the world’s most sophisticated intelligence agencies agreeing that the optimistic adoption posture isn’t enough — even with far greater technical capacity than most organisations.

Legal scholars Woodrow Hartzog and Jessica Silbey put it bluntly in their 2025 paper ‘How AI Destroys Institutions’: AI systems erode expertise, short-circuit decision-making, and isolate people from each other in ways that degrade civic institutions. That’s not an argument against using AI. It’s an argument for building genuine counterweights — not aspirational checklists.

We wrote about a related challenge in our piece on whether your workplace is actually set up for AI agents — the failure rate has almost nothing to do with the technology itself. It’s the governance and accountability layer that breaks first.

What to Do About It

• Do your diagnostic work before deploying. Before adding an AI agent to any workflow, map out who is accountable for the decisions that agent influences. If you can’t draw that map, you’re not ready to deploy.
• Don’t accept speed as a proxy for quality. AI agents are fast. That’s the appeal. But speed without a review layer means errors compound quickly. Build in a human check-point for any output that affects decisions you’d regret.
• Treat ‘responsible AI’ language as a starting point, not a finish line. If your platform, your employer, or your government calls their approach ‘responsible,’ ask what that actually mandates. Principles without enforcement are wishes.
• Watch the agentic AI governance space closely. The Five Eyes guidance from April 2026 signals that incremental deployment and continuous threat assessment are becoming the expected standard — not just best practice. How platforms respond to that signal will tell you a lot about whether their governance posture is real.
• If something goes wrong, trace the accountability chain before it happens. Decide in advance who reviews the agent’s work, who can override it, and who is responsible if it causes harm. Document it. Non-binding frameworks won’t protect you if something goes sideways.

What NZ’s Pollyanna Problem Means for AI Agent Users

• New Zealand’s Public Service AI Framework names the right principles — transparency, fairness, human oversight — but is explicitly non-binding, meaning individual agencies bear accountability without central enforcement.
• Only 14% of NZ public servants use AI regularly, per the 2025 Public Service Census, revealing a significant gap between adoption encouragement and operational readiness.
• Australia’s Robodebt scheme is the warning precedent: algorithmic systems without clear governance accountability caused documented harm and ran for years without adequate scrutiny.
• AI systems have an inherent tendency to mirror user bias, meaning agents can reinforce bad decisions rather than catch them — a risk that non-binding principles don’t address structurally.
• The Five Eyes April 2026 joint guidance on AI agents calls explicitly for incremental deployment and sustained human oversight — the same direction the NZ framework gestures at, but without legislative weight.

The teams that figure this out first — the ones that build genuine accountability structures, not just governance language — will be the ones who can scale AI agents confidently. Everyone else will keep discovering the same failure modes that Robodebt already documented. The technology isn’t the hard part anymore. Deciding to build real oversight rather than optimistic paperwork is.