What the Summary Loses

In 1932, Frederic Bartlett gave British participants a Native American folktale called "The War of the Ghosts" and asked them to recall it repeatedly over weeks and months. The story was deliberately strange, with an unusual narrative structure, ghostly elements, and cultural references that didn't map onto anything familiar. With each retelling, it became shorter, more internally consistent, and more culturally legible. The ghostly elements were dropped or rationalized. Odd phrases became conventional ones. By the end, the story was fluent and coherent, but substantially different from what he'd originally given them.

Bartlett called this leveling and sharpening. Details that didn't fit were leveled out. Details that confirmed the teller's existing understanding got sharpened, made more central. Memory isn't a recording; it's an act of reconstruction, and reconstruction bends toward what the reconstructor already understands.

This is not a historical curiosity. It's what AI summarization does to your notes.

What summaries delete

Ask an AI to summarize a folder of notes about someone you work with, and it will produce something clean and useful. "Alice is a senior designer who prefers direct feedback and works best in the morning." Probably accurate. But the entry from eight months earlier might not make it: "Alice asked a question in the retro that I didn't have a good answer to. Worth returning to." Neither might the note about her seeming more guarded after the Meridian project ended.

Those get leveled. They're anomalies. They don't fit the pattern strongly enough to survive compression.

Gary Klein spent decades studying how firefighters, military commanders, and emergency room doctors make decisions under pressure. His Recognition-Primed Decision model, published in 1985, found that expert decision-making is mostly fast pattern recognition followed by a quick mental simulation of the obvious response. The speed comes from the pattern-matching. The accuracy comes from the richness of the pattern library.

And the most valuable parts of that library aren't the rules. They're the exceptions. The cases where the obvious response failed. The anomaly that, two years later, turned out to be an early signal of something important. Klein found that expert insights arise when people detect contradictions and anomalies — when something doesn't fit and they dig further.

Summarize away the exceptions and you get a clean rule set. You also get something closer to what everyone else already believes.

What compression optimizes for

Summarization — including large language models doing it — optimizes for coherence and coverage. Coherence means the output hangs together. Coverage means the major themes are represented. Neither criterion has anything to say about preserving your anomalies.

Anomalies hurt both. They're incoherent, by definition. They're single data points in a sea of confirming examples, so they fail coverage tests too. The model has no way to know that the one time a vendor relationship went sideways in an unusual way is the most important thing you ever noted about that vendor.

The summary it produces is legible to itself — trained on averages, on what's typical. Your exceptions don't survive the trip.

What structure is actually for

There's a reason I find typed, structured knowledge useful as a hedge against this — not as an aesthetic preference for tidy data, but as a practical response to compression-driven loss.

A typed entity is harder to summarize away than a paragraph. If a person record stores "relationship: complicated — see note 2024-08-14" as a discrete field, that field survives because it's not a candidate for compression. It's already discrete. The anomaly is load-bearing by construction rather than floating in prose where it can be quietly dropped.

This isn't a complete solution. Most interesting knowledge doesn't arrive pre-categorized. But explicit structure preserves the things that matter enough to specifically record, in a form that doesn't require an AI to correctly guess what's important about them.

The alternative — storing everything as prose and trusting AI to surface what matters — is Bartlett's experiment running continuously in your knowledge base. Clean, fluent, losing a little of the weird stuff every time.

An asymmetry worth knowing about

There's something uncomfortable about the direction this is heading. AI summarization is getting better. The outputs are more fluent, the coverage more reliable. And so the summaries are becoming more convincing, which makes it harder to notice what's been left out.

A bad summary is obviously bad. You can tell something's missing. A very good summary is harder to interrogate. It accounts for everything it includes and gives you no signal about what it didn't.

The exceptions that shaped your thinking are still in your notes somewhere. But if you're querying a summary, you'll never know to look for them.

---

Asgeir Albretsen is the founder of Harbor.