Session on 08/06/2026 – AI Symposium
Pieter Pangat, Research Lead
Interest in “AI” bumped along quietly for fifteen years, mostly a researcher’s concern. Then ChatGPT launched, the graph decided to become a vertical wall, and AI stopped being a niche topic almost overnight. That wall is the wave we are all paddling in right now, and like Hokusai’s fishermen, most of us are somewhere between the crest and the boat.
I find it helps to sort what AI does into three buckets, with apologies to Sergio Leone.
The Good is when performance and ethics line up. AlphaFold predicting protein structures, retinal scans flagging early signs of dementia, an app that signs along to children’s stories. Genuinely useful, and hard to argue with.
The Bad is a performance problem. The model simply does not do what it should. A delivery chatbot that can be talked into swearing at customers and writing poetry about how useless its own company is. An airline chatbot that invents a refund policy the airline then has to honour. The damage here is reputational, the company takes the hit, and afterwards we all get to have a laugh about it.
The Ugly is the same thing with an ethics problem attached. The line between the two is thinner than it looks. It is the moment the failure starts having tangible, negative impacts on individuals. Chatbots implicated in real harm to vulnerable people. Generative tools that turn a famous face into anything you type. A government that used an algorithm to hunt benefits fraud and wrecked thousands of lives in the process. The systems often did roughly what they were built to do, which is exactly the problem. And nobody is laughing.
Which brings me to the line I kept coming back to: compliant is not the same as ethical. The EU AI Act gives us a useful risk pyramid. But legislation moves at the speed of committees, and technology does not. Ticking the legal boxes is necessary. It is nowhere near sufficient.
Put differently: compliance is an attitude driven by ethics, not a checklist we wait to be handed. We should not sit on our hands until the law catches up. We should be building these systems with a moral and ethical compass from the start, so that by the time the rules arrive we have already cleared the bar.
So, four things I try to hold onto, especially in high-stakes assessment:
- Just because AI can, doesn’t mean it should. Capability is the start of the conversation, not the end of it. Before asking whether a model can grade, flag, or decide something, ask whether it should be the one doing it at all.
- Know what your model can and can’t do. Knowing what your model is good at is the easy part. Knowing where it breaks matters more. Every model has a competence boundary, and most failures happen when we quietly forget where it is. Map those limits before you deploy, not after a candidate has been wrongly flagged for fraud.
- A human in charge, not just in the loop. AI has the IQ side of intelligence down well by now, and on plenty of narrow tasks it outperforms us comfortably. What still sets us apart is EQ, the judgement and feel for context that no model really has, and that is why a human has to stay in charge rather than just nod along.
- Compliance is the floor, not the ceiling. Regulation tells you the minimum you can get away with. It says nothing about what you should aspire to. Treat it as the starting line, and let your own standards carry you well past it.
The thread running through all four: the higher the stakes, the higher our responsibility. In assessment the stakes are someone’s diploma, their visa, their career. That is not a place to let the wave do the steering.
The full deck is available below. It has more cowboys than a conference talk strictly needs, which felt appropriate.
Download the full presentation below.