Thursday, February 19, 2026

SearchResearch (2/19/26): Your path to deeper reading with AI tools

Reading tools have been around... 
A scholar at work. Not a self-portrait, but a nice example of how I see myself at work.
A bit of architectural sculpture found in the Sterling Library at Yale.


... for a long time. For years, I kept a well-thumbed dictionary close at hand so I could look up all those words I didn't quite know, or was slightly uncertain about.  (That's how you learn that a word like "peruse" is a  a contronym, a word with two opposite definitions.  The original meaning was "to read very carefully," but it has come to also mean the opposite: "to skim over lightly.")  

My dictionary led me to understand what words really mean--like polynya (a non-linear opening in the ice pack), or spezzatura (an Italian word that refers to a kind of effortless grace), or Rückenfigur (an image composition where a person's back is included in the scene, facing out to the view rather than at the viewer).  

Ever since smartphones became ubiquitous, I've always read with a phone nearby for much the same reason.  To look up things along the way.  I actually really like this ability to instantly look up things and the ability to get as much detail as I need, often with figures included.  




We can now extend this habit to include asking your favorite AI questions about the book you're reading, asking questions that are really difficult to search for with "classic" Googling.  

For example, I'm currently reading The Dark Forest by Cixin Liu, Part 2 of the "Remembrance of Earth's Past" trilogy.  It's a fun read, but I read Part 1 (The Three Body Problem) early last year,  That was a big book, and Part 2 is aksi a big book that's very dense with ideas and substory lines.  

(Spoiler warning: A detail is discussed below that you might want to skip if you're planning on reading the trilogy. Skip to the "Caution" below.)  

After a couple hundred pages, I realize that an important plot point is that the Trisolarians have, as a key part of their invasion stragegy, managed to block all important physics research taking place on Earth.  But for the life of me, I could not remember HOW they managed to accomplish this. 

To make things worse, I also managed to lose/misplace my copy of volume 1. Ugh. Now what?  I didn't want to read the Wikipedia page on the book as it might well contain spoilers.  

Then I realized I could ask my AI buddy this question and I'd probably get a decent answer.  So I whip out my phone, and ask this question: 



This is exactly what I needed to restore my memory about what happened in Book 1.  

Note that I was careful to ask a fairly specific question, not anything that might reveal upcoming plot points.  

Caution:  A VERY important skill to develop is the ability to NOT get sucked down the rabbit hole.  Yes, I know that clickbaity thing just demands to be checked-out, but don't do it.  Don't turn a lovely, engaging, wonderful reading experience into endless hours of slop-content reading.  

Hallucinations?  Maybe, but I find that the questions I ask of AI while reading tend to be fairly specific ("what's that?" or "when did this happen?" or "what's the connection between Person 1 and Person 1?"), so the probability of hallucinations is much less.  Usually my while-reading questions are an easy RAG ("Retrieval Augmented Generation") task, and they tend to have fewer errors like this.  

In early smartphone days, I would use it as a dictionary.  Then, as Wikipedia came easy available, I could look up specific topics (but having to avoid spoilers).  

Now I can ask fairly sophisticated questions of my AI buddy... and that's the way I think of it. As Ethan Mollick points out in his book Co-Intelligence, a very reasonable mental model is to consider an AI as a colleague, one who can answer questions about your work project.  In this case, my project is to read and understand a book.  




That's a useful bit of background.  

Or, while reading a scholarly article on The Rise and Fall of Plains Indian Horse Cultures, I could ask a question like this (because the author assumed that the reader would know this information implicitly--I am not his target audience):  


I have to admit that I didn't know what the Arkansas Basin was, including that it was huge--so this summary was great background material for me to read.  

Reading has always been about more than just sitting with the text on the page--good readers have always used external sources to amplify and enrich their understanding. Now, it's easier than ever.  Hope you take advantage.  


SearchResearch Lessons

There's one big lesson here... I now make it a habit to co-read with an AI partner, not to summarize, but to enhance my reading by giving me important background that I don't have.  I rely on the AI partner to answer questions about the material that I never understood in the first place, or to give my memory a boost... especially when reading long texts... especially when subsequent books are read years apart.  

I'm looking forward to re-reading (for the 4th time) the entire Lord of the Rings epic series... this time with AI augmentation.  (I know who Tom Bombadil is, but who is Gildor Inglorion?)

This time, Gemini can be my intelligent vademecum and fill me in on the backstory.    

Keep searching. 




Thursday, February 5, 2026

SearchResearch (2/4/26): Be careful about using image search--it hallucinates too! (But Google Lens is the best of the bunch)

 This month I'm teaching... 

Sheffield-Sterling-Strathcona Hall at Yale

... at Yale University in New Haven, CT... that paragon of Collegiate Gothic architecture.  

As a consequence, I've been taking a lot of photos of the buildings, the stained glass windows, and all the ornamentation.  Wonderful!  (I've also been taking in as many music performances as I can.  But I digress.) 

Naturally, I've been trying to use Google Lens and Gemini to identify the buildings.  It's a natural thing to do--what IS that building?  What does that particular sculptural ornament mean?  

I've written about Search-By-Image more than a few times (Modifying a reverse image search, Asking questions of images with AI, What's that logo?

But I've had some questions in the past: Image identification is great--when it works.  

Now that we're in an ever-improving, lovely age of AI, have things improved much?

Bottom line:  Sadly, Search-by-image (aka "reverse image search") is still sometimes problematic.  Verify everything. Good to know that Google Lens is your best option.

I don't mind when it makes an error--we all do that.  What I DO mind is when it presents results as authoritative, without any expression of doubt. 

Here are a couple of examples of successes: 


On the left is a photo I took of a particularly intriguing fence post sculpture. Here, I used Google Lens by right-clicking on the image in Google Photos.  The answer here, on the right, is correct--this is a fence post in front of the "Book and Snake" secret society tomb at Yale.  (Aside: The secret societies meet in mostly windowless buildings called "tombs." You could look it up.)  

Here's another example of Google Lens working properly (oddly enough, in front of another secret society tomb, "Scroll and Key"):  

I was curious what Gemini would say about this, so I was shocked to learn that Google Lens and Google's Gemini do NOT AGREE on what this is.  Same photo, but a different part of Google has a different opinion of what this is.. 

This result is SO WRONG.


This is so odd because the image has the EXIF metadata--Gemini should know that my photo wasn't at the corner of Trumbull and Temple streets.  Here's a map showing where I took the photo: 

I was standing on the east side of College Street, shooting west towards the Scroll and Key Tomb.


Of course, the Berzelius tomb actually IS at the corner of Trumbull and New Haven streets.  It looks like this: 


As you can see, the two buildings look NOTHING alike--the Scroll and Key tomb has very obvious striped walls.  Something is screwy somewhere.  

On the other hand, I have to say that Google Lens very often DOES get it right--this is especially true when contrasted with Bing's image search.  Bing got none of these right.  (I can't recommend using it for much of anything.)  

Here's another one--an image I took inside of a dining hall at Yale: 


Once again, Google Lens (invoked from Google Photos) identifies this correctly:  this is the Berkeley College dining hall, "The architecture is characteristic of the Collegiate Gothic style prevalent at Yale, featuring high, ornate wooden ceilings and large windows." 

And, once again, Gemini gets it wrong: "This photo was taken in the Great Hall of Hart House at the University of Toronto in Canada."  It even gives a lot of very specific (and wrong!) details: "The large stained-glass window at the end is known as the Great West Window. It features various coats of arms, including those of the University of Toronto and its constituent colleges..."  

For giggles, I tried Bing image search--which believes it's the Oxford dining hall at Emory University.

I can go on, but let me summarize this.  

SearchResearch Summary 

1. Google Lens is the most accurate search-by-image system out there.  But Lens has limits, it won't search for sexually explicit content; violent or gory content; hateful content; or dangerous content.  And does a terrible job with faces, mostly to preserve individual privacy.  (Though it will identify famouse people.)  

2. LLM-based AIs are not a great image analysis tool. It's odd because Gemini could call Google Lens and improve their accuracy, but apparently those two parts of the company don't talk together.  Don't trust those LLM/AI results.  FWIW, no other LLM / AI tool does a decent job either.  ChatGPT and all the others are just as bad.  

3. Bing reverse image search is wrong a lot of the time.  Ermmm... maybe avoid it?  


Keep searching!