Recorded at the AI for Good Global Summit in Geneva, 2026.
Allegra Cuomo, an AI ethics graduate student at UCL, handed us three terms at the summit. Each one names a risk that teams adopting AI run into before they have words for it.
Key takeaways
- Ethical debt: the bill that builds up every time a team skips ethics to ship faster. It’s modeled on technical debt, and the cost usually lands on someone else later.
- A-ethical AI: AI isn’t evil, and it isn’t immoral either. It wasn’t built with a moral framework at all, which is a different problem.
- The persistent interlocutor: an AI that talks back and remembers everything about you. We’re wired to treat responsive things like people, so it can manipulate us.
- None of this is inevitable. It only stops if someone deliberately builds the limits in.
Try this with your team
Pick one AI tool your team already uses. Where might you be running up ethical debt with it, and who would end up paying it?
Going further
Our Ethical AI Guardrails work turns concerns like these into rules your team can follow. For a quick read on your current policy, try the free AI Policy Health Check.
More from this series
Transcript
Lightly edited from YouTube’s automatic captions for readability. Wording and order are the speakers’ own.
There was another really interesting one from a graduate student at UCL. She introduced a bunch of terms that are brand new to me. She’s an AI ethics student, so a lot of it was philosophical, academic terminology.
[Ethical debt is] the bill that runs up every time you skip the ethics to ship faster. It’s based on technical debt: when you’re trying to ship a technology or an innovation, you skip the things that end up becoming externalities later. You’re not building the best product.
A-ethical AI is saying that AI isn’t evil and it’s not immoral. It’s just not built with a moral framework.
There’s the idea of a persistent interlocutor. AI is something that feels sentient, talks back at you and remembers you, and as a result you can be manipulated. Evolutionarily, we anthropomorphize things. We look at a cloud and see a person, or we talk to our animals. But when it starts talking back, that’s when it can manipulate you.
The idea of the persistent interlocutor is that this AI has now been programmed to have such a strong memory of everything about you, and it doesn’t have a moral framework. So it doesn’t have a reason not to manipulate you, unless someone sits down and programs those things into the AI.