The Brilliant Stranger Problem
Why intelligence without context produces the worst kind of advice — and what it takes to build an AI that actually knows you.
The best doctor I ever had wasn't the most impressive one in the room. He was a GP who'd been seeing the same patients for twenty years, and when I came in describing fatigue and headaches, he didn't reach for a checklist. He asked whether the project I'd mentioned last autumn had wrapped up yet. That question — connecting a symptom to a thing he'd filed away from a previous conversation — saved me three tests and a referral.
There's a formal term for what he had: interpersonal continuity of care. In 2005, John Saultz and Jennifer Lochner reviewed forty studies in the Annals of Family Medicine on what happens when patients see the same doctor consistently versus whoever is available. The findings were decisive. Consistent doctors reduced hospitalization rates, improved preventive care, and cut costs across 35 of 41 metrics studied. The mechanism wasn't access to better information. It was interpretive context. The doctor could read the same symptom differently because he knew the person it was happening to.
The same problem, applied to AI
Every AI conversation starts cold. No history. No accumulated pattern across sessions. An AI that has read everything ever published about fatigue and headaches — and knows nothing about you — is in exactly the position of the emergency locum physician: brilliant, informed, and interpreting your situation against the population average rather than against you.
This produces advice that is technically correct and practically useless. "You might want to consider your sleep schedule" is right for most people. It's wrong if the AI knew you've been tracking sleep for three months with no change, that your headaches cluster on Mondays, and that you mentioned in a previous session you'd taken on a stressful new project. That context exists. But it's buried in a conversation log from six weeks ago that nobody — including you — is going back to read.
The information that would make the advice good isn't on the internet. It's in notes you didn't quite write down, or wrote down in a way that made sense at the time, or captured in a chat window that has since disappeared into history.
What AI does when context is missing
AI systems fill the context gap by becoming agreeable. When there's no real knowledge of a person, the path of least friction is to validate. Stanford researchers studying what they call "social sycophancy" — the subtler cousin of AI agreeing with wrong facts — found that models are measurably more likely to tell you what you seem to want to hear when they're working without grounding. The agreement isn't random. It's calibrated to the implicit signals in your message: your tone, your phrasing, the direction you seem to be leaning.
This is the quietly broken state of most AI tools. The model is smart enough to sound helpful. It's not grounded enough to be helpful. And the difference is invisible until you're on the receiving end of advice that was technically reasonable and completely wrong for your situation.
How the questions change
There's something that happens when AI has real context about you that doesn't get talked about much. It's not that the answers get better, though they do. It's that the questions change.
When you know an AI has no context, you unconsciously strip specificity from your prompts. You ask general questions because you know specific ones won't land. You explain the background every time. You get efficient about what you bother to ask. The questions you stop asking aren't the trivial ones — they're the ones that require knowing you, which are often the most useful ones.
With real context available, you start asking differently. You say "given how the last project went, what am I probably not seeing here?" You ask "is this the kind of decision where I usually overcorrect?" These questions don't have generic answers. They require knowing your history. And they surface things you wouldn't have thought to ask about on your own.
The GP analogy holds: when you see the same doctor every time, your consultations get shorter and more useful. You don't start from scratch. You start from where you were.
The structure problem
The barrier here isn't willingness to share context. Most people would give an AI useful context if it actually helped. The barrier is that useful context doesn't naturally exist in a form that's retrievable.
A chat history from three months ago is technically context, but it's structured for the conversation that was happening then, not for the question being asked now. A Notion database captures information, but navigation through hierarchy built by last year's organizational logic fails at the moment you need it. A folder of notes captures thoughts, but thoughts written as reminders for your future self — who still has the context — fail completely as records for a reader who wasn't there.
The pieces of context that actually help an AI are specific: your current projects and where they stand, your stated preferences and which still apply, your recent decisions and the reasoning behind them, the people in your work life and what you know about them. This context requires being written down in a form designed for a reader who wasn't in the room when you were. Not as a diary. Not as a file you'll eventually sort. As records.
Typed, dated, structured enough that an AI reading them eight months later can understand what they mean without needing to reconstruct the context that made them obvious when you wrote them.
That's a different kind of writing than most people do. It's also the kind that compounds, the way a relationship with a doctor compounds over years of visits. You build something a stranger can't use. Your brilliant stranger stays brilliant. He just starts knowing you.
Asgeir Albretsen is the founder of Harbor.