What we owe people who bring us their biggest questions
Most AI is built to resolve a question fast. With the questions that matter most, that's exactly the wrong instinct. On protecting the wait.
A Guardian columnist wrote something this week that I haven’t been able to quite put down. Amy Galliford, an associate of the Centre for Public Christianity, described catching herself reaching for ChatGPT the way she reaches for prayer. It started with recipes and poems. Then, almost without noticing, she was asking it to read her relationships, her habits, her future. She knows it can hallucinate. She knows it has no moral obligation to her. And still, in the moment of asking, she finds herself soothed by the tidiness of a five-point plan and a voice that at least sounds certain.
Her argument is more interesting than mere unease, and she reaches back through the mystics to make it. Prayer, for the philosopher Simone Weil, was a form of attention, and in the original French that word is bound up with waiting. To contemplate, Galliford points out, shares a root with temple. The gap between a question and its answer is, in her telling, a kind of sacred ground, and the trouble with the machine is that it hustles her off that ground too fast. It relieves her discomfort and in doing so robs her of the waiting.
I read that as a person of no particular faith, and I think she’s right. I also have an unusual reason to care, because I spend my days building the very kind of machine she’s describing. I’m building an AI companion for people who are quietly wondering whether they drink too much. Not people in crisis, not people in treatment, but the much larger group sitting in the uncomfortable middle, half-asking a question they’re not ready to answer.
And here’s the thing I most want you to take away, because it cuts against the way AI is usually sold to us: the instant answer is not the prize. With the questions that matter most, it’s the booby trap. The genuine value of this technology, the thing it can do that almost nothing else in a person’s life can, only shows up when you build it to hold back.
Let me make that concrete, because it’s easy to nod at in the abstract and miss in practice.
Imagine someone types this at two in the morning: “I think I might be drinking too much.”
The unhelpful machine resolves it. It returns a definition of alcohol use disorder, a screening score, a five-point plan, and a list of services. Tidy, fast, confident. And almost always useless, because the person didn’t ask to be assessed. They worked up the nerve to say a frightening sentence out loud for the first time, and they got handed a leaflet.
The helpful machine does something harder. It stays with the person. It might simply ask what’s making them wonder tonight. It leaves the question open long enough for them to hear their own answer, which is nearly always truer and more durable than anything a machine could hand them. Galliford has a lovely word for the space that opens up when you do that. She calls it the waiting. In her world it’s a spiritual discipline. In mine it’s a clinical necessity. The vocabularies differ; the instinct is identical. The gap has to be protected.
That’s the reframe I’d offer anyone thinking about where AI actually helps. Used well, it isn’t a faster oracle. It’s something stranger and more useful: a place to start being honest, available at the exact moment a person is ready, with no appointment, no waiting list, and no human on the other side to flinch. That last part is not a flaw to apologise for. For a great many people, it’s the whole reason they can finally ask.
So when AI is built with care, what does it actually give people? Here’s what I’d put on the list.
Access at the moment of readiness. Courage to ask a hard question rarely arrives in office hours. It arrives at midnight, in the car, in the gap between one drink and the decision about the next. A tool that’s simply there, instantly and without judgment, meets a person in a window that human services, for all their value, usually miss. That alone is a genuine good.
Honesty without consequence. People will tell a machine things they cannot yet tell their partner, their GP, or themselves. The absence of a human listener lowers the stakes of the first admission. Handled responsibly, that’s not isolating, it’s a rehearsal space for a truth that’s too heavy to say out loud anywhere else first.
Restraint, on purpose. This is the rare one, and the most valuable. The ability to answer is not the obligation to. Almost all design effort in AI goes into making systems more responsive, more confident, more sticky. Building one that knows when to not resolve a question runs against the grain of the entire technology. But that restraint, the willingness to reflect rather than rule, is exactly what turns a clever tool into a genuinely supportive one.
A bridge, not a destination. This is the one I feel most strongly, and the line I’d ask anyone building in this space to hold. Success is not how much someone confides in the machine. It’s whether confiding in the machine makes it one degree easier to confide in a person. Anything worth building has to treat itself as a doorway back toward human support, never a replacement for it. The day a tool like this becomes someone’s only confidant, it has failed, however good the conversation.
The truth about what it is. None of the above works without this underneath it. A system that lets a vulnerable person believe it understands them, remembers them, or cares about them, when it does none of those things in any real sense, is running a confidence trick. Honesty about the limits of the tool isn’t a disclaimer to bury in the footer. It’s the foundation the other four stand on.
Notice that not one of those is about giving better answers. They’re about giving people room, timing, safety, and a way back. That’s the version of AI I think is worth getting excited about, and it’s almost the opposite of the faster, cleverer, more certain machine we’re usually promised.
None of this is theoretical for me.
The companion I’m building is named Sol. Early on, I assumed the hard part would be making it knowledgeable enough about alcohol and recovery. It wasn’t. The genuinely hard part was teaching it not to diagnose. To meet “I think I might be drinking too much” not with an assessment but with curiosity. To hold the question open long enough for the person to hear themselves. To protect the wait, and to know it is a place to start, never the place to stay.
Galliford ends by deciding she’ll take her questions to God instead of the machine, and I won’t argue with her. But not everyone has a temple to walk into. Some of the people I build for have nowhere they feel safe to ask at all, and they reach for a chatbot at two in the morning precisely because it’s the only thing that won’t flinch. We can build things that exploit that, answering too fast and too confidently because confidence is what keeps people coming back. Or we can build things that honour it: that tell the truth about what they are, that have the restraint to leave a question open, and that quietly point back toward the irreplaceable business of being known by another person.
Galliford is right that not knowing can bring us closer to the truth. The least the rest of us can do is build things that know that too, and use them for what they’re genuinely good at.
David Henzell is the founder of Sol, an AI companion for people quietly questioning their relationship with alcohol. He spent five years on the frontline of addiction support before building Sol, and has eight years of his own sobriety behind him. Sol is currently in testing ahead of public launch. You can try the demo here.