Efficiency's Blind Spot

What AI search takes away when it gives us exactly what we asked for.

Previously on Giuseppe’s Glimpse: In the last episode, we explored what happens when a company hands a decision to AI for so long that nobody remembers how to make it without help. Missed it? Catch up here! ✨

Buongiorno everyone 👋

Do you remember the last time you got lost in a library? Not looking for a specific book, just wandering around the shelves led by the titles.

I don’t think I remember mine. It’s been a while since I searched for something without knowing, more or less exactly, what I’d find.

That’s not a complaint about libraries closing or habits changing. It’s something narrower. Something that, once again, I started thinking about while noticing how I use AI.

I think by now most of us share the experience of asking an AI tool a question to get a quick answer instead of looking for it ourselves. Knowledge is starting to feel quick and targeted by default, almost regardless of the subject.

If I think back to the way I used to learn, in school, at university, later on my own, it looked a lot different from this.

The efficiency trade

AI-powered search is remarkably good at one thing: getting you to an answer. Ask it something and it synthesizes a handful of sources into a clean paragraph, usually in a few seconds.

The old way made you do that synthesizing yourself. On paper, it meant pulling three or four books off a shelf, checking their indexes, reading the relevant chapter in each, and noticing when two authors disagreed with each other. Online, it meant something faster but not fundamentally different: open ten links, read the ones that looked credible, run into two sources that contradicted each other.

Either way, you had to decide who to trust, or hold conflicting information in your head until a third source broke the tie.

That decision, repeated enough times across enough subjects, is roughly how you built a working model of something rather than just an answer about it.

AI search skips straight to the tie already broken. That’s the upgrade that cuts down time and effort, because that’s the part that used to be the work.

What gets lost along the way

There’s a term researchers in information science have used since the 1990s for this: information encountering, coined by Professor Sanda Erdelez to describe the accidental discovery of something useful while you’re looking for something else entirely.

It’s the paragraph you weren’t searching for that ends up reorganizing how you think about the paragraph you were, or the source that contradicts the one you just read, forcing you to sit with the disagreement instead of accepting the first answer.

Traditional search was full of friction, and the friction is where a lot of this happened. You had to compare sources because nothing handed you a synthesis. You noticed when two experts disagreed because you were reading both of them instead of a summary that had already reconciled them on your behalf.

Researchers who study how journalists come up with story ideas have found they rarely describe a deliberate method for it. There’s no query that reliably produces a good one. What they describe instead is putting themselves in places and conversations where something interesting tends to turn up, without knowing in advance what it will be. That’s the opposite of asking a targeted question and getting a targeted answer back.

I’m not trying to be nostalgic about how things used to be better. These are simply facts describing how certain kinds of insights are more likely to come about.

AI search resolves the disagreement before you see it. That’s the value proposition and also the cost, since the resolving used to be what turned an answer into an actual mental model of the subject.

Where this appears most

I notice this most in my own work, usually when I’m under time pressure, which is exactly when the shortcut is hardest to resist.

If I ask an AI tool to summarize a market, I get the consensus view: the competitors everyone already tracks, the trends already in every report on the topic. It’s accurate, and it would take me an hour to assemble the same thing myself.

What I don’t get is the layer underneath it, like the odd regulatory filing nobody’s flagged yet or a line in an earnings call transcript that doesn’t quite fit the narrative. Those things used to surface because you had to read fifty documents to write a two-page brief, and got pulled off track by document thirty-one somewhere along the way. At the time that felt like wasted effort, but looking back, it was often where the real value came from.

Hiring is another good example. A colleague told me she now uses AI to shortlist candidates against a job description, which is efficient and mostly fair to everyone in the pile.

What she says she misses though is the resume that didn’t fit the brief at all, the one that used to stop her and make her see the position in a different way.

So the question is not that the AI’s answer is wrong. The issue is that it’s way thinner than the version you’d have by getting there yourself, and the thinness is easy to miss because the answer looks complete.

What do we keep?

I don’t think the answer is to avoid AI search, and I’d be suspicious of anyone telling you to go back to typing ten queries by hand out of principle. I use these tools daily and I’m faster because of them.

What I’ve started doing instead is paying attention to why I’m reaching for AI in a given moment. Sometimes it really is to save time, other times it’s just to skip past the discomfort of not knowing something yet.

While they feel the same in the moment, these are two very different things, and knowing which one you’re doing is most of what separates using AI well from using it on autopilot.

The practical version of this, for me, has been leaving a few pockets of the work deliberately unautomated. I’ll read a report nobody asked me to read, or follow a citation that has nothing to do with what I’m working on, or just let a piece of research take longer than it needs to.

Sure, this might not be efficient, but that’s sort of the point. It gives me room to engage with information differently, to notice things I wasn’t looking for, and every so often one of them turns into something I wouldn’t have gotten to any other way.

Designing for it on purpose

Obviously it’s hard to scale this smoothly into a policy. A company can’t mandate serendipity, and a rule that says “spend twenty percent of research time browsing” would probably just get gamed or ignored.

What companies should pay attention to though is which tasks get treated as pure retrieval and which ones get treated as understanding in progress.

Retrieval is fine to hand over completely. Understanding needs the chance for a detour left in, at least occasionally, or the person doing the work ends up informed about a great many things but truly familiar with very few of them.

That’s true for a company deciding what to automate, and it’s true for me deciding whether to just ask AI or dig in myself on any given afternoon. Same question, smaller scale, but you have to start somewhere.

Stay curious 🙌

gs.

---

Oh, wow! You made it to the end. Click here to 👉 SHARE this issue with a friend if you found it valuable.