The Self-Portrait You Didn't Sit For

James Pennebaker spent decades analyzing how people write, and his most persistent finding runs against intuition: the revealing words aren't the ones you chose carefully. They're the function words — the "I"s, the "but"s, the "anyway"s — produced without deliberate thought. When Pennebaker's LIWC system processes a large sample of someone's writing, it builds a surprisingly accurate personality profile from word frequency alone, and content carries less signal than the structure around it. The words you don't notice yourself using say more about you than the ones you meant to say.

I thought about this a lot while building Harbor.

What lives in a knowledge base

The obvious layer of a personal knowledge base is the one you put there: the notes, the decisions, the person records, the tasks. You wrote these things down because you wanted to remember them. They represent your thinking, at least the version of it you chose to surface.

But there are other layers underneath.

One is linguistic. The register you use in project notes versus personal reflections. The way you phrase things when something is going wrong. Whether you write "I should" or "I need to" or "I'm going to" — each carries different psychological weight and tends to be consistent in ways the writer doesn't notice. Pennebaker found that pronoun use alone — how often someone writes "I" versus "we" versus "you" — is reliably associated with depression, social status, and personality traits. Not because people are expressing those things intentionally, but because language patterns are mostly automatic.

The other layer is behavioral. What you return to. What you start and abandon at the same point in the process. Which tasks you reschedule indefinitely. Which contacts you actually maintain. Which project gets a careful weekly note and which one gets a status update once and then silence. None of this is written anywhere as a statement about yourself. But the aggregate tells a coherent story — one you probably haven't read, because you've only ever read the base one note at a time.

The prediction gap

In 2013, Michal Kosinski published a study in PNAS showing that Facebook likes predicted personality more accurately than close human relationships. A model with 70 likes outperformed a friend or roommate. With 300, it outperformed a spouse. The computer wasn't more empathetic or insightful than the people who knew the subjects well. It was just better at reading behavioral residue at scale.

Timothy Wilson's 2002 book Strangers to Ourselves explains why from the inside: we have significantly less access to our own motivations than we feel like we do. The adaptive unconscious — the mental processes that actually drive most of our behavior — operates below the level of introspection. We construct narratives about ourselves that feel accurate and are often not. Wilson's research is full of cases where people confidently explain why they made a decision, unaware that the actual cause was something entirely different, visible only from outside the moment.

The implication is a little strange. Your knowledge base contains both the self-portrait you deliberately painted and the behavioral record you accumulated without noticing. If Pennebaker and Kosinski are right — and they've replicated this extensively — the second one may be the more accurate picture.

A different question about access

Most thinking about AI and personal knowledge focuses on retrieval: can the AI find the note you wrote last September? Can it remember that you prefer short meetings and dislike status-for-status's-sake? These are valid questions.

But there's a less obvious question worth sitting with: what does an AI with full access to your knowledge base infer that you never explicitly stated?

Not the preferences you wrote down, but the patterns you didn't register as patterns. Which projects collapsed at the planning stage, not the execution stage. Which person you mention across many different documents without ever writing their name as an important relationship. The concern that shows up in three different contexts under three different labels, recurring enough that it must be structural rather than incidental.

Some of this is simply useful — an AI that can read your actual work patterns is more useful than one that only reads your stated intentions. But some of it is genuinely surprising to contemplate. Your notes, read in aggregate by something that can hold all of them at once, might describe a version of you that you'd recognize, mostly. A version that has decided certain things and keeps undeciding them. A version that says one thing in plans and does another thing in practice.

Not as a criticism. As a description.

The honest record

I don't think this is a reason to avoid building a knowledge base. The alternative — scattered notes across five apps, no persistent context, an AI that starts from scratch every session — is clearly worse. But it seems worth knowing what you're actually handing over when you hand over access.

Pennebaker's finding wasn't that people are lying in their writing. It was that the most honest signal is often invisible to the person producing it. The function word you didn't think about. The pattern in when you show up and when you don't. The aggregate of a thousand small decisions.

A knowledge base your AI can read is a powerful thing. It might also be the most accurate account of yourself you've ever accumulated — written in a language you weren't quite aware you were using.

---

Asgeir Albretsen is the founder of Harbor.