Accountable by Design: AI Infrastructure in the Public Interest

When the system that decides cannot be inspected

Artificial intelligence is moving from novelty to infrastructure, taking on roles in how public services are delivered, how decisions are made, and how resources are directed. When an AI system helps decide who gets a benefit, how a case is prioritized, or what a clinician sees, it has become infrastructure, and infrastructure that makes consequential decisions about people carries a special obligation: it must be inspectable, accountable, and governed in the public interest. This primer is about what it takes to build AI that meets that bar, and why the properties that make AI trustworthy in public service are the same ones that make any public infrastructure sound.

The argument is that public-interest AI is not defined by the sophistication of its models but by whether it can be understood, contested, and held to account by the people it affects, and that achieving this depends far more on data foundations, transparency, and governance than on the algorithms themselves.

Why public-interest AI is a different problem

A commercial recommendation engine can be a black box and still be useful; if it suggests a poor film, little is lost. An AI system that informs whether someone receives care, housing, or a public benefit cannot be a black box, because the stakes are different and the people affected have a right to understand and challenge what is decided about them. The shift from low-stakes commercial AI to high-stakes public AI changes the requirements fundamentally.

This is why public-interest AI has to be accountable by design rather than explainable after the fact. The properties that matter are familiar from the rest of this series: transparency, so the system can be inspected and trusted rather than taken on faith; interoperability and open standards, so it connects to other systems and is not a proprietary trap; and clear governance, so someone is answerable for what it does and the public can contest it. An AI system that is accurate but unaccountable fails the public-interest test, because accuracy without the ability to inspect, explain, and challenge is just a more confident black box.

The foundation is data, not algorithms

It is tempting to treat AI as fundamentally about models, but in public service the harder and more decisive part is the data foundation underneath. An AI system is shaped by the data it learns from and operates on, which means the quality, standardization, and governance of that data determine whether the system is trustworthy, fair, and even functional. A good data strategy, one that lets information move across an organization without being trapped in silos, is the foundation any serious AI strategy depends on.

This connects public-interest AI directly to the standardization and interoperability work elsewhere in this series and in health informatics specifically. Where data is standardized, governed, and interoperable, AI can be built on a sound and inspectable foundation. Where data is fragmented, proprietary, or poorly governed, AI built on top inherits those flaws and amplifies them, producing decisions that cannot be audited because the inputs themselves cannot be traced. The unglamorous data layer, not the model, is where public-interest AI is mostly won or lost, which is a familiar lesson for anyone who has worked on standardized health data: the analysis is only as trustworthy as the data model beneath it.

Governance and the avoidance of capture

The same lock-in risks that threaten other public infrastructure threaten AI, and arguably more so, because AI systems are complex and their inner workings are opaque even to many of the institutions that deploy them. An agency that adopts a proprietary AI system it cannot inspect, built on data formats it cannot port, governed by terms it did not shape, has created a particularly deep form of dependency: it has outsourced consequential public decisions to a system it does not understand and cannot leave.

Avoiding this requires the governance disciplines this series keeps returning to, applied with extra care. Open standards and interoperability, so AI systems are not proprietary traps. Transparency requirements, so the basis of decisions can be examined. And chartered, accountable governance, so a public body, answerable to the people affected, retains real authority over how the system is used and can change or retire it. The aim is to capture the genuine value AI offers public service while refusing the capture that opaque, proprietary AI invites. Accountability is not a constraint added after the model works; it is a design requirement present from the first decision about data and architecture.

What this means for public-interest infrastructure

Treating AI as accountable public infrastructure changes the standard it is held to. The measure is not the model's sophistication or even its accuracy alone but whether the system can be inspected, explained, contested, and governed by the public it affects, and whether the data beneath it is sound enough to make those things possible. That favors investment in data foundations, transparency, and governance, the unglamorous layers, over the model-centric excitement that treats AI as magic rather than infrastructure.

This is the Foundation's stewardship concern applied to the most consequential new layer of public infrastructure. AI that serves the public must be accountable by design, built on governed and interoperable data, transparent in its operation, and held by custodians answerable to the people it affects. Build it that way, with the data and governance foundations that accountability requires, and AI can genuinely serve public ends. Build it as an opaque, proprietary system that decides about people without recourse, and it becomes the least accountable infrastructure of all.

References

1. Nextgov/FCW. Interoperability and modernization: Competition drives progress (2025). A nimble, interoperable data strategy across an organization as the foundational step for any AI strategy; open standards reduce lock-in. https://www.nextgov.com/sponsors/2025/11/interoperability-and-modernization-competition-drives-progress/408499/

2. Linux Foundation Europe. Building Digital Public Infrastructure Through Open Source (2025). OSPOs advancing AI policy, digital governance, and innovation; avoiding proprietary dependency. https://linuxfoundation.eu/newsroom/building-digital-public-infrastructure-through-open-source-key-insights-from-un-open-source-week-2025

3. Public Digital. Digital public goods. Open AI systems among digital public goods; the need for chartered custodians empowered to govern in the public interest. https://public.digital/pd-insights/signals/signals-5/digital-public-goods

4. Chatham House. The case for expanding digital public infrastructure (2025). Open standards, APIs, and safeguards create a flexible, accountable foundation and reduce dependency. https://www.chathamhouse.org/2025/10/case-expanding-digital-public-infrastructure/08-conclusion-and-recommendations

5. Digital Public Goods Alliance. Our policies to unlock the promise of digital public goods (2025). Security by default, governance indicators, and accountable stewardship of open digital systems. https://www.digitalpublicgoods.net/blog/our-policies-to-unlock-the-promise-of-digital-public-goods