Healthcare AI in the Pacific: when missing data creates the inequality it promised to eliminate

By Momentum IA · June 28, 2026.

At the Te Poutoko Ora a Kiwa symposium held on June 24 in Auckland, Professor Robyn Whittaker, co-director of the University of Auckland's TRANSFORM Research Centre, delivered a plenary lecture on artificial intelligence in health aimed at researchers, clinicians, health professionals and members of Pacific communities. The central message was as clear as it was uncomfortable: AI in healthcare is reaching these communities without having been designed or validated for them.

Whittaker organized her analysis around the three major vectors of AI in health: predictive analytics, computer vision and generative AI. For each she set out both the potential and the concrete limits. On the predictive front, she gave the example of algorithms used during the COVID-19 pandemic to identify Pacific patients at higher risk of hospitalization, which allowed community centers to prioritize care. But she warned that such usefulness depends on high-quality, locally relevant data, something that is scarce today. In computer vision, she pointed to retinal-detection tools deployable in community settings as a promising advance in access to care, although held back by obsolete systems, implementation difficulties and a lack of community trust. And on generative AI —ChatGPT and its equivalents— she used two metaphors that do not contradict each other: 'a paradigm shift' and 'the wild west'. Systems trained on massive internet data, unregulated, probabilistic by nature and capable of producing convincing but incorrect outputs.

The common thread running through the entire intervention was data representation. 'It is highly unlikely that these tools have been properly tested in Māori and Pacific communities', Whittaker stated. This is not rhetorical denunciation: it is a precise technical description. Diagnostic and predictive models learn from the data they are trained on. If that data comes overwhelmingly from European or North American populations —as is the case with most large biomedical datasets—, the model's performance for underrepresented populations is statistically inferior. Not out of ill will, but because of poor design.

This is not an issue exclusive to the Pacific. It is the general case, and the Pacific illustrates it especially clearly because the gap between technological promise and local reality is particularly visible there. Whittaker also underscored the absence of specific regulatory frameworks: neither New Zealand nor most of the Pacific island states have legislation governing the use of AI in healthcare. Her team leads a multidisciplinary national advisory group to evaluate AI tools against criteria of safety, ethics, equity and a Pacific perspective. It is, on a small scale, exactly the kind of governance that is missing on a global scale.

Our read is this: the problem Whittaker describes is not peripheral. It is central. Medical AI will only fulfill its transformative promise —early diagnosis, universal access, reduced burden on clinical staff— if the data that feeds it represents all of humanity, not only the part of it that has historically dominated healthcare systems and scientific databases. The underlying trap is structural: systems are built with the available data, and the available data reflect decades of unequal investment in public health. An AI trained on that inequality does not correct it; it encodes it.

In the short term, the real risk is that tools meant to improve collective health become mechanisms that amplify existing inequities, with the added legitimacy conferred by the technological halo. An algorithm that underdiagnoses in a community is harder to question than a professional who underdiagnoses: the bias is hidden in the model's weights. Whittaker and Professor Sir Collin Tukuitonga, co-director of Te Poutoko Ora a Kiwa, frame it in terms of Indigenous data sovereignty and community control, which is exactly the right lever: before a tool is deployed, the community it will operate on must have a voice in its design, access to its results and the ability to reject it.

In the long term, there are reasons for optimism that should not be abandoned. The computer vision that detects diabetic retinopathy in rural Pacific settings without the need for an ophthalmologist present is, potentially, one of the great stories of healthcare democratization of the century. Well-calibrated predictive models can save lives precisely in the systems with the fewest resources. Generative AI can ease the administrative burden that consumes hours of clinicians who could be devoting them to patients. All of that is real. But the condition is that those tools be built well, with representative data, with local governance and with accountability mechanisms.

The Auckland symposium is a small but significant sign that there are communities that are not passively waiting for Silicon Valley to hand them solutions. They are articulating their own epistemic demands: that the evidence be local, that the data be theirs, that the design be participatory. That is not an obstacle to innovation; it is the condition for innovation to be legitimate and lasting.