Don't give me the summary

I've been thinking a lot about the value of summarization as an act of cognition.

Some of the most prevalent examples of AI in the wild are tools that will summarize your email, transcribe and summarize your meetings, summarize news articles, summarize Wikipedia articles, or a wide smattering of articles from across the interwebs, etc.

In the face of information overload, we just want the Cliff Notes, but I think that's a problem. The prepackaged summary always seems comprehensive, but it misses a crucial element: your relationship to the source material.

To be fair, I use AI for summarization all the time. I appreciate how I can sit through a small mountain of documentation, Reddit threads, and random blogs to give me a high-level overview of a new software that I'm thinking about trying, a technical approach I'm considering, or just the contents of a Google Sheet I haven't opened in a hot minute. In those cases, I just want the gist because maybe I'm just getting familiar with a new domain or I need a quick insight into a file.

But the problems surface quickly when we count on the AI summary for everything. I've had Gemini record and transcribe several meetings. It creates a full transcript of the meeting with the dilligence of a court reporter along with a summary and checkbox list of things it classifies as action items from the meeting.

I save the recording notes because I think I'm going to go back in reference it, but truth be told I hardly ever do it. The list I write—whether it's in my head, or on paper, or in a doc—is more powerful than what the AI gives me because I am the one who synthesizing the information.

I'm the one who can suss out what the one to three important action items are out of an hour long meeting. I'm the one who can identify the real blockers and the action items that were glaringly apparent to me even if they were unspoken or unidentified on the call.

I don't need a computer to give me a list of 12 things that don't even hit the mark.

The value of writing a précis

Once upon a time, I took a media theory course in grad school, where we had to read a bunch of dense, philosophical essays and books from folks like Roland Barthes, Jean Baudrillard, etc. Implosion of reality, semiotics, cultural analysis, propaganda, and the like—basically light bedtime reading.

Every week we had an assigned piece to read. We had to write a 200 to 300 word summary or précis of the main argument, which requires you to:

1. actually locate the main argument (often requires the sensitivity of a bloodhound)
2. understand the salient points (sometimes needle in the haystack, sometimes too many to choose from)
3. synthesize several thousands of words into a just a few hundred that capture the essence of the piece

The first few weeks were grueling because I found much of the material opaque. Many of the readings were written decades ago or translated from another language to English, so the diction could be a barrier in itself. Sometimes I didn't know enough about the historical background of the piece to make heads or tails of the argument so I had to read and re-read and sometimes learn about the context to understand what was going on.

And even when I finally got the gist of the piece, there was the challenge of putting my understanding and interpretation into a prescribed word count.

Every part of the process was full of friction: reading, interpreting, analyzing, writing, editing for clarity.

But along the way—probably week 4 or 5—things started to click. I started to actually understand the articles. It became easier to identify the argument and make connections to previous pieces and class discussions. I became more confident in my ability to articulate my thoughts about the piece. And as my understanding of each individual piece increased, so did my appreciation of media theory as a field of study.

To date, it's the hardest course I ever took, but it has also been the most rewarding. I still reference the readings from that course all the time.

Summarization is not a "low level task"

The prevalence of AI generated summaries seems to suggest that it's a low-level task that's best offloaded to computers, because after all, anything that a computer can do is dull and useless and rote and uninspiring and not worth our time, right?

What's really happening is that a high-level cognitive process is being offloaded to basic semantic analysis. A Python script somewhere is tokenizing the content into little bits, stripping out the nuance, and delivering "summary" with about as much character as a bowl of Malt-o-meal.

AI often highlights what's unimportant and devalues what's actually important. It has no relationship to the material. It's just working off of a massive corpus of training data, including dictionary definitions, and statistical probabilities.

The AI doesn't understand who sent you that email, why it matters, and where it actually fits in the priority list. It can never understand the context of the 52 weekly meetings that preceded this one and why the topics discussed show significant movement for the company. It can't give a great project status update because it can't filter the signal from the noise.

But you can.

Writing a summary is one of the best ways you can do your future self a solid. The summary that's given obscures. The summary you write clarifies—both for yourself and your team. Take the time. Embrace the friction. You'll thank yourself later.