AI Isn’t the Hard Part. People Are.
I’ve spent a lot of time exploring what AI can do. This course has me thinking much more about what AI should do.
So far, the most valuable lesson hasn’t been a framework, regulation or technical control. It’s the reminder that the biggest challenges in AI aren’t technical at all — they’re human.
A simple framework from the course has stuck with me: who benefits, who is harmed, who decides, who can challenge, and at what scale. They’re deceptively simple questions.
Most AI systems feel harmless when viewed one user at a time. But AI rarely operates at that scale — it operates across thousands, millions, or even billions of interactions. A small design decision repeated millions of times can become a societal issue. That’s when AI stops being a product and starts becoming infrastructure.
The question isn’t just “does it work?” It’s “who does it work for, and at what cost?”
One of the strongest ideas from the course was reframing ethics as a steering wheel, not a brake. Compliance tells us the minimum we must do. Ethics asks whether we’re heading in the right direction in the first place.
That’s an important distinction as more teams rush to adopt AI. Shipping quickly is valuable, but shipping responsibly matters just as much. The course repeatedly reinforced that responsible AI isn’t something you bolt on at the end — it needs to be considered from the start, through design, testing, deployment, monitoring and governance.
The second module focused on human-AI interaction, and one principle stood out: humanisation over automation.
As AI becomes more capable, there’s a temptation to automate every decision possible. But the goal shouldn’t be removing people from the process — it should be helping people make better decisions.
I found myself reflecting on how often we optimise for efficiency without asking what we’re losing in return: context, empathy, judgement, human connection and accountability.
The best AI systems shouldn’t replace human agency. They should strengthen it.
Just because AI can make a decision doesn’t mean it should make the decision alone.
Another useful distinction was between AI safety and AI security. Safety is about unintended harm — what happens when the system gets something wrong? Security is about deliberate misuse — what happens when someone actively tries to exploit, manipulate or break the system?
What surprised me most was learning how AI failures can be invisible. Traditional software often crashes when something goes wrong. AI can continue operating confidently while producing incorrect or biased outputs. That makes governance, monitoring, testing and human oversight far more important than many people realise.
The Uber autonomous vehicle case study was probably the most memorable part of the course. Not because of the technology — because of the people.
The incident wasn’t caused by a single catastrophic decision. It emerged from multiple decisions about risk, testing, deployment, oversight, incentives and accountability. It was a reminder that many failures don’t happen because nobody cared — they happen because assumptions weren’t challenged early enough.
As someone who spends a lot of time working through delivery risks and dependencies, that lesson felt very familiar.
The course has shifted the questions I’m asking when experimenting with AI. Instead of focusing purely on capability, I’m increasingly thinking about responsibility: who might be excluded, what assumptions am I making, what happens at scale, how can someone challenge an outcome, and where does human judgement need to stay involved?
The technology itself is fascinating. But the ethics of AI isn’t really about machines — it’s about the choices humans make when building, deploying and governing them. And those choices matter far more than any model ever will.
The future of AI won’t be defined by what the technology can do. It will be defined by the judgement we apply when deciding how to use it.
Currently 3 of 10 modules completed. Looking forward to diving deeper into governance, regulation and responsible implementation in the remaining modules.