The Reader in the Room

Samuel Pepys kept the most candid diary in the English language — nine years of daily entries describing the plague of 1665, the Great Fire of London, court intrigue, marriage troubles, and several affairs he clearly preferred not to share. He wrote all of it in a shorthand called tachygraphy, developed by a man named Thomas Shelton in 1620, that almost no one else could read.

When Pepys died in 1703, the diaries went to the Pepysian Library at Magdalene College, Cambridge. A scholar named John Smith spent three years, from 1819 to 1822, cracking the shorthand from first principles. He finished before discovering that Pepys had shelved the key to the system a few rows above the diary volumes all along. Smith had been searching for a code that was never meant to be found, while the answer sat at eye level.

What mattered was Pepys' belief. He wrote as he did — unguarded, specific, self-incriminating — because he believed he was writing to himself alone. That belief made the diary possible.

Add a reader, and everything changes.

Front stage, back stage

Erving Goffman published The Presentation of Self in Everyday Life in 1959, and the observation at its center is simple: we perform different versions of ourselves depending on whether we're in front of an audience or not. The front stage is where we manage impressions. The back stage is where we don't bother.

Private notes are backstage. You write them in your own shorthand — unfinished thoughts, idiosyncratic vocabulary, half-baked ideas you wouldn't say out loud. The act of writing without an audience is qualitatively different from writing for one. Not more truthful in some moral sense, just less edited. The gap between what you actually think and what you'd say to someone narrows when there's a someone.

The Hawthorne effect — named after factory productivity studies in the 1920s, though the original data turns out to be messier than the textbook version — describes the same phenomenon more bluntly: people change their behavior when they know they're being observed. The food diary that makes you eat less isn't just a record; it's a witness. Simply writing things down changes what you do, which is one of the weirder properties of notation.

Both findings point at the same underlying structure: the presence of an audience, even an imagined or implicit one, reshapes what gets expressed.

The AI as a persistent reader

AI tools that read your knowledge base are, by Goffman's definition, an audience. Not a human one, but a persistent one. When you connect a knowledge base to an AI assistant and know it's reading your person records, your preferences, your project notes — you have moved from backstage to something in between.

The behavioral shift is subtle enough that most people don't notice it happening. You write slightly more clearly. You leave out the thought you didn't finish. You phrase a preference in a way that sounds considered rather than reactive. You add a sentence of context you wouldn't have bothered with before, because you feel, without quite deciding to, that it should be intelligible.

None of these are conscious decisions. They're just what happens when you sense an audience.

So the AI ends up learning a mildly curated version of you. Not a false one — just one that's been tidied up, made legible, edited toward presentability. That's a real problem if you're relying on the AI to have accurate context for decisions you're actually making, rather than the decisions you'd prefer to appear to be making.

Pepys with a reading audience would have been a less useful historical source. A knowledge base written for an AI reader might be a less accurate map of how you actually think.

What this means for the design

The obvious response is scope control: decide in advance what the AI can read, keep some things off limits. If certain notes are clearly out of range, you can write them without the audience effect. This is genuinely useful. Pepys' solution — tachygraphy — was a version of the same thing: a permission layer, even if a crude one.

But there's a subtler design principle underneath it. The more opaque the AI's reading behavior, the more diffuse the audience effect becomes. If you don't know what the AI reads, when it reads, or why it accessed something, you end up optimizing every note against an invisible presence you can't locate. Generalized self-consciousness is harder to write around than a specific, visible reader.

Visibility helps here in a way that isn't immediately intuitive. If you can see exactly what the AI accessed and when — which queries triggered which reads, which documents it pulled — then the reader becomes legible. A legible audience is easier to account for. You can decide what belongs in view and what doesn't. The notes you keep inside that scope can be written for that reader without contaminating everything else.

John Smith, once he decoded the diaries, realized the key had been in plain sight the whole time. Pepys hadn't hidden it; he'd just assumed no one would know to look. That assumption collapsed the moment someone did.

The question for knowledge bases connected to AI isn't whether there's a reader. There is. The question is how clearly you can see them — and how deliberately you've decided what they're allowed to find.

---

Asgeir Albretsen is the founder of Harbor.