Responsible AI Is Not Just About Regulation—It Starts in the Design
Responsible AI is not only a matter of rules or governance. It is also a technical challenge. That was one of the main messages from Clara Higuera, BBVA’s Responsible AI Lead, at QCon London, a leading international conference for engineers, software architects, and technology leaders. Her talk focused on how to turn concepts such as fairness, transparency, and security into methods, metrics, and controls that can be applied throughout the AI development lifecycle.
When people talk about ethics in artificial intelligence, the conversation often centers on regulation or governance frameworks. But many of the choices that shape how fair, transparent, secure, or sustainable a technical solution will be are made much earlier—during design and development. The data used to train a model, the metrics considered valid, and the safeguards built into the system are, in the end, scientific and technical decisions.
"From this perspective, ethics is no longer something added at the end of the process. It becomes a practical discipline embedded in technological development", stated Clara Higuera in her presentation.“It is not just about asking what is allowed, but what should be built, how it should be built, and under what conditions”, the expert points out. As Langdon Winner argued in his essay Do Artifacts Have Politics?—and as Clara also explains in her talk—these decisions become part of the technology itself, which in turn shapes the experience of the people who interact with it. In other words, technology is not neutral when it comes to values. AI systems are decision-making systems, and many of their ethical implications stem from the choices made during development.
Right now, AI is moving through a stage society has already seen with other technologies, such as aviation and electricity. Both expanded rapidly at first, and only later developed the safety standards and shared frameworks needed to support broader adoption. That is often how innovation matures: it begins as an experimental technology and, over time, becomes reliable infrastructure. Electricity followed that path. Society reacted to its failures and inconsistencies, but technical progress—and a growing focus on people’s safety—made the grid more robust, until it eventually became part of everyday life.
From the ethical foundations of AI to decisions in technical development
If ethics, transparency, and security are understood as core to AI, the next question is how to turn them into practice. The first step is to adopt the same mindset used in reliability engineering: design systems with the full lifecycle in mind, from initial design through implementation and ongoing monitoring. “Bias can appear at many points in the process: in historical data, in how the population is represented, in the way variables are measured, or in monitoring once the system is already in production. Assessing fairness therefore requires a continuous, end-to-end perspective,” says Clara Higuera.
In this context, explainability and metric evaluation become practical tools. At BBVA, that means applying quality reviews and evaluation methods designed to ensure that AI solutions meet standards for security, privacy, and transparency. Teams have hands-on guides on explainability and fairness, along with metrics and libraries such as mercury-explainability and mercury-monitoring. These tools help explain the decisions made by AI models and ensure they remain accurate and reliable when working with real-world data. This work is also supported by applied research, including the development of a stress test to measure bias in generative AI. That allows us to assess how large language models perform when responding to user queries.
One of the most important lessons from our work on AI development has to do with fairness in machine learning. There is no universal definition of what is fair. Fairness depends on the context, the specific use case, the groups involved, and the potential harm. That is why, in high-impact models, teams must explicitly decide, based on the use case, which fairness criterion or metric is most appropriate in each case—and explain why.
Ultimately, AI systems encode our values, whether we intend them to or not. Recognizing that is the first step toward designing them in a safer, more transparent, and more responsible way.