Healthcare

Two simultaneous challenges

AI's relationship with healthcare in the NZ context involves two distinct problems that pull in different directions.

The first is the transformation of healthcare delivery itself. AI is already demonstrating clinical capability that rivals specialists in narrow tasks: reading radiology images, detecting early-stage cancers from scans, flagging drug interactions, triaging emergency presentations. NZ faces a chronic shortage of specialist clinicians, particularly in provincial and rural areas. In this context, AI looks less like a threat to the health workforce and more like a desperately needed extension of it.

The second is the health consequences of economic disruption. The evidence on job loss and health is long-established and grim: unemployment is associated with increased rates of cardiovascular disease, mental illness, substance abuse, and premature death. The causal mechanisms run through financial stress, loss of routine and social connection, and reduced access to care. Large-scale AI-driven displacement would generate a wave of health need — arriving precisely when the health system is already under pressure.

Any policy approach needs to address both dimensions.

The governance question for clinical AI

How should AI tools in clinical settings be regulated? This is not a settled question anywhere in the world. Three broad approaches are being discussed:

Voluntary standards: Industry and professional bodies develop their own standards for AI clinical tools, with light-touch regulatory oversight. Fast to implement, preserves innovation, but leaves enforcement to self-interest. Who favours this: technology companies, some clinicians who want faster access to tools. The risk: standards erode under commercial pressure; adverse events go unreported.

Mandatory pre-market approval: AI clinical tools require regulatory approval before deployment, similar to medical devices. Medsafe would need to develop or adopt a framework (the EU's AI Act provides one model; the FDA's approach in the US is another). Who favours this: patient advocates, risk-averse clinicians, those who note that AI tools can fail in systematic and non-obvious ways. The risk: slow to implement, may exclude smaller NZ-specific tools, may not keep pace with rapidly evolving AI capabilities.

Sector-specific regulator: A dedicated health AI regulator with expertise across clinical practice, data science, and ethics — with powers to approve, monitor, and withdraw AI tools. More responsive than Medsafe's existing framework; more accountable than self-regulation. Who favours this: those who see health AI as genuinely novel and believe existing regulatory categories don't fit. The risk: regulatory overhead, possible duplication with existing bodies, and the challenge of building specialised expertise quickly enough.

Māori health equity as a hard constraint

Whichever governance model NZ adopts, Māori health equity cannot be an afterthought. This isn't only a values argument — it's a practical one. AI systems trained predominantly on non-Māori data (as most global health AI systems are) will perform less reliably for Māori patients. Pulse oximeters, dermatology algorithms, and pain assessment tools have all shown racial bias in clinical studies overseas.

Any framework for clinical AI in NZ needs to address: data sovereignty (who controls Māori health data used to train or validate AI systems), performance requirements (disaggregated reporting by ethnicity as a condition of deployment), and Māori governance (meaningful input from iwi and Māori health providers into approval processes, not consultation after the fact).

The Te Whatu Ora restructure provides an opportunity to build these requirements in from the start rather than retrofitting them later. Whether that opportunity is taken is a political choice.

The workforce redistribution question

AI in healthcare is unlikely to reduce total employment in the sector — NZ's workforce shortages are too acute for that. But it will redistribute work: fewer radiologists doing routine reads, more radiologists reviewing AI-flagged cases; fewer GPs doing initial triage, more GPs handling complex cases that AI can't manage.

The question for workforce planning is whether this redistribution benefits patients equitably. If AI extends specialist capability to rural and low-income communities, that's a genuine equity gain. If it concentrates efficiency gains in well-resourced urban hospitals while provincial facilities fall further behind, it compounds existing disparities. The difference lies in procurement and deployment decisions that are largely invisible to the public — but consequential.