Some time last week, the AI I use for organising my random thoughts ran out of room and quietly rewrote our whole conversation as a summary (it happens in my head too). It stopped reading the original, carried on from the summary, and neither of us noticed.
That’s compaction. Every chat tool does some version of it, because a model can only hold so much text at once. Some apps show a line saying it’s “organizing its thoughts”, most show nothing, and the model has no way of knowing what its summary left out. Afterwards, the summary simply is the conversation.
Compaction is old
People have always compressed: meeting minutes, the handover note at shift change, the top paragraph of a report that is the only paragraph anyone reads. Retell a story often enough and it drifts towards the shape you expected it to have, which psychologists noticed about a century ago.
Who holds the pen
What’s new is who writes the summary, and what for. A model compacts so it can finish the task. The feed compacts the world before you see it, and its task is keeping you on the feed, so fashion, a war, a sleep tip and a phone launch arrive in one thumb’s length, already cut down to the part that holds attention. Your inbox offers to turn forty emails into five lines, and your meeting tool has a summary out before the call has properly ended.
This year the summarise button was sold as the end of reading, and nobody’s launch slide mentioned what you stop knowing. I built one of those buttons, naturally. Our meeting bot at work emails whoever pressed record a summary and the full transcript, and I’d bet anyone on which one gets read, which is why it sends both.
The invisible cut
Forgetting leaves a gap you can feel: you know you read something about that once, you just can’t remember where. Compaction leaves nothing to feel. Five lines feel like forty emails, and a meeting summary feels like having been there.
That’s what the model and I now have in common. After compaction it feels caught up and answers confidently from the summary, and most of the time the summary is good enough. But “good enough” was decided by whoever wrote it, for their reasons, and you inherit the result without seeing the cut. Verification debt is impossible to pay off here, because you can’t check what you never saw.
Editing
After its compaction, my own conversation kept missing exactly the details I cared about: that a publisher’s name had to be spelled out in full, or that I’d asked to talk about an idea rather than research it. The summaries were reasonable every time. My corrections were the only part of the work that was unmistakably mine, and without them it would have drifted somewhere slightly wrong and carried on just as smoothly.
We all know the scale, roughly (it’s coarse, but it’ll do):
- The machine summarises and you skim.
- The machine drafts and you fix what’s wrong.
- You write the thing completely yourself.
Whether you use the tools matters less than where you sit on that scale, because that decides how much of what you know was chosen by you.
We trained these models on our writing, which was already a compressed version of us, and now we read their summaries of our writing and find them complete. Somewhere in that loop a person used to sit with a red pen. These days the seat is mostly empty, and from the inside it doesn’t feel empty at all.
/compact
⎿ Compacted. You are absolutely right! AI summaries conceal what they cut, giving us a seamless sense of completion while quietly taking over the editor's chair—leaving where we choose to edit as the only mark of what’s truly ours.