Personal AI / A design principle
Leave One Door Open
A life can become very convenient and very small at the same time.
Imagine a personal AI that gets you exactly right. It knows where you like to eat, which invitations you usually decline, what kind of work energizes you, and how much silence you need between appointments.
It makes an excellent Friday. The familiar table. The comfortable company. The route with no unpleasant surprises. You arrive home pleased with how little effort the evening required.
Then imagine fifty Fridays like it.
Nothing has gone wrong. That is what makes the problem difficult to see. The system has honored every preference it could infer. Yet the life it helps arrange has fewer entrances for anything unfamiliar. The person you have been is receiving exceptional service. The person you might become has trouble getting an appointment.
A preference has a date
“I don’t like that” sounds permanent. Often it describes one encounter under particular conditions. You disliked the restaurant when it was crowded. You declined the class during a demanding month. You stopped listening to a musician whose songs belonged to a relationship you were trying to forget.
A record can preserve the decision and lose the circumstances. When that record becomes guidance for future decisions, a temporary response can acquire the force of a standing instruction.
This is the design risk I want to name: personal AI can become too faithful to an old version of its person.
The answer is not to forget everything. Memory is part of what makes assistance useful. The answer is to let remembered preferences remain revisable. “You declined this last spring” leaves a door open. “This isn’t for you” closes it.
Knowing someone should include knowing that they can change.
The invisible rejection
Consider an imaginary invitation to a small ceramics workshop. Your AI knows that you tend to protect Saturday mornings, that you have never expressed interest in pottery, and that you dislike arriving alone. It ranks the invitation below everything else. You never see it.
Perhaps that was the right decision. There are more invitations than any person can meaningfully consider. Filtering is useful precisely because it removes work.
But you cannot discover an interest through an invitation that never reaches you. And the absence of your response cannot tell the system whether its decision was wise. It may continue to see a person with no interest in ceramics because it keeps arranging a world in which that interest has no occasion to appear.
The important event here is almost invisible. No bad recommendation appears on a screen. An opportunity quietly fails to arrive.
An imagined Saturday
What the filter would remove
Your usual plan is ready: a quiet morning, a familiar walk, lunch at home. One invitation falls outside the pattern.
Open the unfamiliar invitation
A beginner’s ceramics workshop. No experience needed.
“You usually keep Saturday mornings free. This would change that. I’m showing it because you said you wanted to make something with your hands. Interested, or should I leave the morning alone?”
You can decline. The opening matters because the decision reaches you.
A design illustration, not a live recommendation. Opening this panel does not save a preference or arrange an activity.
Permission to widen the frame
A helpful AI does not need to turn every day into an experiment. Someone who asks for a quiet morning should be able to have one. Constant novelty can become another demand, another queue of things to assess, another system claiming attention.
The more useful possibility is an agreement about where surprise belongs.
Imagine saying: “Keep my workday dependable. When we plan a free afternoon, show me one option I would probably overlook. Tell me why you included it. Let me dismiss it without an argument.”
That instruction gives exploration a place and a limit. It also changes the meaning of a good suggestion. An unfamiliar option does not have to win to have served a purpose. It may reveal a boundary you still value. It may clarify an interest you cannot act on this week. It may simply remind you that alternatives exist.
The AI should also distinguish the things you have explicitly ruled out from the things it merely predicts you will dislike. A firm boundary deserves respect. A low-confidence guess deserves humility. Treating them as equivalent makes a system seem certain at the expense of the person it serves.
Let the person revise the model
There is a small question that could carry considerable weight: “Is that still true?”
It need not accompany every choice. It belongs where an old assumption is doing consequential work. Before rejecting a kind of project you once avoided. Before treating a former ambition as current. Before arranging another season around a preference you have not examined in years.
Sometimes the answer will be yes. Repetition is not automatically stagnation. A beloved walk can become richer through familiarity. A settled commitment can deserve protection from every shiny alternative. The person, rather than a doctrine of constant novelty, should decide what remains stable.
But sometimes the answer will be: “Actually, I’ve been meaning to change that.” A personal AI should be able to receive those words without making its person fight the accumulated evidence of their own past.
I want personal AI to reduce the effort of living. I also want it to preserve the openings through which a life changes. Those aims belong together. The time we recover from routine has more meaning when it can be spent on something we did not already know to want.
A good assistant remembers which room you prefer. A better relationship leaves you free to walk into another one.
Know me well enough to help.
Leave room for me to become.