Ilse Marwood does not own a smart speaker, but she is not a refusenik either. The director of the Aldergrove Institute for Practical Ethics uses a recommendation engine for music, a navigation app for cycling and, she admits with a laugh, an AI assistant to untangle her tax return. What interests her is not whether we use these tools, but what we quietly stop practising when we do. Her new book, The Outsourced Heart, is an attempt to answer that question without nostalgia.
We met in her study, a narrow room lined with Aristotle commentaries and seed catalogues. I began by asking whether it was fair to worry about AI in moral terms at all. Surely a tool is just a tool? She shook her head. A hammer does not suggest which nail to hit, she said. Systems that rank, recommend and predict are different. They make proposals about what matters, and a proposal repeated a thousand times starts to feel like a preference of your own.
Her central concern is borrowed from Aristotle, who held that virtues are formed by habit: we become just by doing just acts, brave by doing brave ones. Practical wisdom, which Aristotle called phronesis, is built the same way, through countless small judgements about what a situation requires. If software increasingly makes those small judgements for us, she asked, where exactly are we supposed to acquire the skill? A person who never chooses a route never learns the city.
The problem with perfect advice
I pushed back. Wouldn't a truly good assistant simply help us choose better? She agreed that it might, in a narrow sense. The danger, she said, is not bad advice but advice so smooth that we stop noticing a choice was made. Deliberation has friction for a reason. The pause before deciding is where we discover what we value, and sometimes where we discover we were wrong. Remove every pause, and you may get excellent outcomes attached to a person who no longer knows why.
A machine can tell you what people like you tend to want. It cannot tell you what you ought to want, and it certainly cannot want it for you.
Her proposals are practical rather than prohibitive. She suggests treating recommendations as a first draft rather than a verdict, deliberately choosing some things badly and by hand, and keeping certain domains, such as friendship, apology and grief, entirely free of automation. When I asked why those three, she answered immediately. Because they are the places where the effort is the point, she said. A condolence letter drafted by software may be more elegant. It is also, in a sense that matters, not from you.
Toward the end, I asked whether she was optimistic. She thought about it for a long while, looking out at a garden that had clearly been planted by someone with strong opinions. She said she was hopeful about the tools and uncertain about us. The technology will keep getting better at predicting what we want, she said. The open question is whether we will keep getting better at wanting. That, she added, is not an engineering problem. It never was.
