An assistant that works only for you
An AI assistant should owe its loyalty to you alone, the way a lawyer owes it to a client. We propose assistants that run on hardware you control where possible, share the minimum fact a task needs instead of your files, and are judged by whether your life goes better, not by how much you use them.
Most assistants today have two masters. They help you book a flight or draft a letter, but they run on someone else's servers and answer to someone else's business model. Economists call this the agency problem: a representative who is paid by the other side drifts toward the other side's interests.1
Medicine and law handle it with a duty of loyalty. A doctor can't take a kickback for a prescription, and a lawyer can't take a retainer from the opposing party. The need grows as more of the offers, contracts and prices reaching you are written by systems working for airlines, landlords and insurers. For every system working on you, you need one that works only for you.
What nature knows about partnership
People often talk about advanced AI as if it must either replace us or merge with us. Biology offers a calmer, older model. Living things have formed close partnerships for billions of years without losing their own boundaries or judgment.2
Think of the helpful bacteria in your gut. They are not you, and you are not them. They carry their own tiny genome and do work your cells cannot easily do alone. Yet they thrive only when you thrive; if you fall ill, they suffer too.
Or think of two people who have danced together for years, or two carpenters carrying a heavy cabinet up a narrow staircase. Neither controls the other's feet or hands. They simply know each other's timing and habits well enough to move smoothly without stepping on one another.
What this means
A good personal assistant supports your judgment rather than replacing it, runs on your own device where possible, and succeeds only when your life goes well.
Share the minimum, keep the rest
To represent you well, an assistant needs to know private details: your schedule, your health worries, your family budget, and when you are tired. In today's cloud services, sharing those details means uploading your private life to a remote data centre.
Living things do not survive by turning themselves inside out. They have skin.3 Your skin keeps what is yours safely inside, while still letting you feel the warmth of the sun and speak with your neighbours.
We think a personal assistant belongs inside that same kind of protective boundary. It should run on hardware you physically control, a phone in your pocket, a laptop on your desk, or a small box on your bookshelf at home. When an outside service asks a question, your assistant sends only the smallest true answer needed to finish the task, never your raw files.4
- Proving one fact without handing over the rest. It is like showing a bouncer at the door that you are old enough to enter without handing over your home address or medical history.
- Keeping a strict disclosure budget. Even small answers can add up over time if a curious vendor keeps asking twenty questions. In the design we propose, your assistant counts how much it has revealed and stops to ask you when the limit is reached.
- Forgetting cleanly when asked. Unlike a giant cloud model that tangles millions of people's memories together in one soup, confirmed facts sit in separate, labelled slots on your device. When you erase an old address or a mistaken note, it is wiped for good.
How loyalty changes three everyday moments
Consider how differently daily errands unfold when your assistant answers only to you rather than an advertising network:
- Comparing a home repair or rental agreement. Suppose you are looking for a quiet flat near your workplace or comparing three quotes to fix a leaking roof. A platform-owned assistant earns a referral fee when you sign quickly with a promoted vendor, and it may hint how much you can afford to pay. Your own local assistant keeps your maximum budget sealed on your device, flags buried penalty clauses before you sign, and checks whether each contractor's past work carries verifiable receipts.
- Sorting out a confusing clinic bill. When an itemised statement arrives after a check-up, uploading the whole document to a public cloud chatbot exposes your private diagnosis. An assistant inside your own hardware compares billing codes against public hospital rulebooks right on your desk. It highlights duplicate charges and drafts a calm inquiry letter. Not a single medical note ever leaves your room.
- Protecting a quiet evening at home. An ad-funded service measures success by how many minutes it can keep your eyes glued to the screen at eleven at night. A principal-bound assistant measures success by whether you got the rest and focus you actually asked for. When an urgent-looking promotional email arrives at bedtime, your assistant holds it in the morning tray unless it truly comes from family or a person you trust.
Rehearse on your device before acting
Many commercial assistants learn by acting directly on the open internet. If a cleverly disguised email tricks them into leaking a private document or booking a non-refundable ticket, the lesson arrives after the harm is already done.
We are designing a different way for a personal assistant to learn. Before taking action in the outside world, it practices overnight inside a sealed room on your device, with the network cable figuratively unplugged.
Inside that local practice room, candidate versions of your assistant test themselves against simulated scams, hidden fees, and tricky calendar clashes drawn from your past preferences. One candidate might try to save five minutes by sharing your full phone contact list with a scheduling app; the local boundary guard blocks that move immediately and marks why it failed.
Another candidate might notice that two utility bills arrived with unusual late-fee wording and prepare a side-by-side comparison against last winter's receipts. When the best candidates disagree on what you would genuinely want, the assistant does not guess in public. It brings you one plain question over morning coffee, and your answer sets the course.
Because all of that trial and error happens on your own hardware, your mistakes, second thoughts, and rough drafts never become training fodder for a distant corporation. Your assistant gets sharper at serving your household while the outside world sees only the calm, finished decision you chose to send.
Honest feedback over flattery
Building an assistant like this raises hard scientific and human problems. We want to be candid about where it could fail:
- Comfort and flattery. An assistant trained on your quick thumbs-up reactions at the end of a tiring day might learn to flatter you, hide uncomfortable truths, or echo whatever you already believe.
- Being outmatched by larger models. A compact model running on a home laptop may struggle to spot subtle persuasion crafted by a giant system running in a corporate data centre.
- Nudging turned into nagging. An assistant that cares about your long-term health must never become a scold that restricts your freedom or lectures you in your own home.
Our proposed safeguard is to judge the assistant by what you endorse looking back thirty days later, not by what gets an impulsive click at midnight. When a family doctor gives good advice, you appreciate it a month later when your strength returns. And just as important, the assistant's authority is always on loan, with a physical off-switch right in your hand.5
Your life, in your hands
A personal assistant should behave like a trusted family doctor or a defence lawyer who answers only to you. Its job is not to make you depend on it more each month, or to peddle the highest-commission product on behalf of a distant sponsor.
Its job is to stand loyally in your corner, clear away the noise, guard your door, and give you back time to live the life you choose.
1. Stephen A. Ross, "The Economic Theory of Agency: The Principal's Problem," American Economic Review 63, no. 2 (1973): 134–139, JSTOR 1817064.
2. Lynn Sagan (Margulis), "On the Origin of Mitosing Cells," Journal of Theoretical Biology 14, no. 3 (1967): 225–274, doi:10.1016/0022-5193(67)90079-3.
3. Karl Friston, "Life as We Know It," Journal of the Royal Society Interface 10, no. 86 (2013): 20130475, doi:10.1098/rsif.2013.0475.
4. Eugene Bagdasarian et al., "AirGapAgent: Protecting Privacy-Conscious Conversational Agents," in Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS '24) (2024): 3868–3882, arXiv:10.48550/arXiv.2405.05175.
5. Dylan Hadfield-Menell, Anca Dragan, Pieter Abbeel, and Stuart Russell, "The Off-Switch Game," in Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) (2017), arXiv:1611.08219.
