Building Responsible AI Is a Human Problem

When I started this course, I assumed AI ethics would be mostly about models, algorithms and technical controls. Six modules in, I’m coming to a different conclusion — many of the hardest problems in AI have very little to do with AI itself. They’re about people: who gets included, who gets excluded, who carries the burden when a system gets something wrong, who is responsible for fixing it. And perhaps most importantly — who gets to challenge the decision?

One of the biggest shifts in my thinking came from the module on fairness, non-discrimination and inclusion. Before this course, I probably would have measured fairness by looking at overall performance numbers. The course challenged that assumption: a system can perform well on average and still produce consistently worse outcomes for specific groups. The average can look healthy while the lived experience tells a completely different story. What matters isn’t only whether a model works — it’s who receives the errors, delays, scrutiny or barriers when it doesn’t. Fairness isn’t a property of the model. It’s a property of the experience.

Another idea that stayed with me was the difference between equality and inclusion. Treating everyone exactly the same sounds fair on the surface, but in practice it can create exclusion. A process that only works online disadvantages people with limited digital access. A process that only provides information in one language excludes others. An appeals process might technically exist, but if people can’t realistically use it, it’s not meaningful. The lesson wasn’t that every system needs a special rule for every scenario — it was simpler than that: fairness has to work in practice, not just on paper.

The privacy module introduced another uncomfortable idea. In the past, privacy was mostly about data we knowingly provided. AI changes that — modern systems can infer things that were never explicitly shared: mood, health indicators, personal preferences, risk profiles. The question is no longer just “what data did I give away?” It’s also “what can be inferred about me from what’s already available?” That feels like a fundamental shift, and the more capable AI becomes, the more important it is to ask whether collecting more data is actually necessary. One principle from the course captures this perfectly: don’t ask how much data you can collect — ask how much data you genuinely need.

That principle is data minimisation, and I found it particularly relevant as someone who enjoys building things. It’s easy to fall into a “just in case” mindset — store everything, keep it forever, maybe it will be useful later. The course argues for the opposite: only collect data that serves a clear purpose, only retain it for as long as it’s needed, be deliberate about how and why it’s used. Constraints can actually improve design. Sometimes the best solution isn’t collecting more information — it’s designing smarter systems with less.

The final module introduced a distinction I hadn’t considered deeply before. Transparency means people know AI is involved. Explainability means they understand why a decision was made. Accountability means someone owns the outcome when things go wrong. All three matter, and they work as a chain — a system can be transparent without being explainable, explainable without providing meaningful recourse, and it can provide information without anyone being accountable. Break one link and the others get weaker. The test I keep coming back to is simple: could an affected person understand what happened and take action if they needed to? If not, the explanation probably isn’t good enough.

Across the last few modules, a theme keeps appearing: AI governance is less about controlling technology and more about designing responsible systems around it. The questions I’m asking myself are changing — who might be excluded? Can someone challenge this outcome? Are we collecting data because we need it or because we can? Would a non-technical person understand this decision? If something goes wrong, who owns the fix? Those feel like better questions than simply asking whether the model performs well, because trust isn’t created by algorithms. It’s created by design, governance, accountability and the choices people make around the technology. The more I learn about AI ethics, the less it feels like a technology discipline and the more it feels like a human one.

Currently 6 of 10 modules completed. Next up: exploring the remaining modules and how ethical principles become organisational and societal practice at scale.