Blog

An illustration of two people flanking a laptop that shows a checklist of ticked principles. One holds a magnifying glass, the other a clipboard. Around them are an open eye, a closed eye, a question mark, a shield with a check, and a set of scales.

Trained to Be Good. How Would We Know?

We can train an AI system to behave in ways we prefer. The harder question is how we would know the training worked, and whether it will hold when things change.

An illustration of a person at a desk looking at a translucent overlay showing a workflow diagram and an approval card for a $170 refund, with buttons to approve or reject it.

Human in the Loop, but Is Anyone Looking?

Human review is often treated as the fail-safe for AI agents. But when a reviewer can't understand what happened, or doesn't have the time to check, approval becomes a formality—and the controls that don't depend on catching the mistake start to matter more.

An illustration of four figures—a developer, an engineer, a reviewer with a checklist, and a lawyer—surrounding a laptop displaying a security shield and lock, connected by lines representing shared responsibility.

Who Is Responsible for Keeping AI Systems Safe?

When an AI system causes harm, responsibility rarely sits in one place. A look at the operational and security side of AI safety—who controls which safeguards, and who is in a position to reduce the risk.

View from a lodge window overlooking a motorway

Shoebox

Each morning, the drilling began before I could decide whether I was awake.

An image of a girl wearing a panda mask.

The Art of Being Unfazed

The most important change that comes with age is the ability to not give a f****.

lorem-ipsum

Parenthood

Parenthood is a blessing—the ability to nurture life, so delicate and dependent, and see it grow and change every day.