The 2024 European Parliament elections brought AI regulation to the forefront of public discourse. Citizens and policymakers alike are asking: when a public institution deploys an AI system—whether for benefits processing, border management, or public safety—what guarantees exist that the system is fair, transparent, and accountable?
Policy is necessary but insufficient. Every responsible AI framework—from the EU AI Act to the OECD AI Principles to the NIST AI RMF—establishes principles: transparency, fairness, accountability, safety. These principles are correct. But principles do not deploy systems; engineers do. And engineers interpret principles through the lens of architecture.
Fairness by construction. Algorithmic fairness is not a property that can be tested into a system after deployment. It must be designed into the data pipeline, the feature engineering process, and the model evaluation criteria from the start. This means: auditing training data for representational gaps before model training, not after deployment complaints. It means establishing fairness metrics as acceptance criteria in the CI/CD pipeline, not as post-hoc monitoring thresholds.
Auditability as architecture. An AI system is auditable when every decision can be traced to its inputs, its model version, and its configuration. This requires a decision log that records not just the output but the entire inference context: input features, model version, confidence score, and the alternative outputs that were considered. This log must be immutable, queryable, and retained for the period required by the relevant regulation.
The procurement lever. Public institutions have a unique mechanism for enforcing responsible AI: procurement. By specifying architectural requirements in tenders published on TED—from decision logging and fairness testing to human oversight mechanisms—institutions can ensure that the systems they acquire are responsible by design, not just by declaration. In the US, similar requirements are emerging through SAM.gov contract vehicles that reference the NIST AI RMF.
Responsible deployment is not about slowing down AI adoption. It is about ensuring that when public institutions adopt AI—and they will—they do so with systems that earn the trust of the citizens they serve.