The AI annotates your book sparsely, on purpose
MindMax's AI annotates a book sparsely on purpose: it marks only where a busy expert would actually stop, tells you up front how often it plans to stop, and two of its four lenses ask you questions instead of handing you conclusions. Fewer, better marks are the ones you reread.
The first reaction to an AI reading pass is often mild disappointment: it went through a whole chapter and left two notes. Surely it could find more than that?
It could. We decided it should not, and the reasoning is worth writing down, because "more annotations" is the kind of improvement that makes a product worse in a way users do not immediately attribute to the product.
Highlighting has a known failure mode
Look at a used textbook. The first chapter is a wall of yellow, the middle is patchy, and the last third is clean — not because the argument got less important, but because highlighting is a decision made before you know what matters, and the energy for making it runs out.
A tool that marks everything of possible interest reproduces that failure automatically and at scale. What you get is a document where the marks carry no information, because a mark that appears everywhere tells you nothing about anywhere.
So the bar is not "is this passage interesting". It is would a busy expert reading through the lens you picked actually stop here — which is a much higher bar, and it is why a chapter can legitimately go by with two notes.
Expectation management is part of the design
The problem with a high bar is that sparse output and broken output look identical from the outside. If you were picturing a glowing page, two notes per chapter reads as "it did not work".
That is why a pass opens by telling you what it intends to do before it does it: what kind of document it thinks this is, roughly how often it plans to stop, and which chapters it rates as worth your time or safe to skim. You get the number to judge it against up front. A pass that says "this is a dense methodological paper, I will stop six to ten times, mostly to question the assumptions" and then stops seven times has done what it said.
This turned out to matter more than any change to the annotation quality itself. The same output, with and without that opening paragraph, gets read as two different products.
Two of the four lenses refuse to give you the answer
The second decision people notice is stranger. Of the four lenses, two — critique and study — deliberately withhold conclusions. They ask; you answer.
The temptation to fix this is strong, because an AI that hands you a verdict feels more capable than one that asks you a question. But for a document you are trying to learn from rather than process, a supplied conclusion is precisely the thing that prevents the learning. You read it, you agree with it, and nothing happens in your head. The generation effect is one of the better replicated findings about how people come to know things: producing an answer yourself does something that reading one does not.
So the critique lens marks a premise and asks what follows if it does not hold. The study lens sets an exercise and does not grade it until you have tried. Meanwhile the guide and review lenses do hand you conclusions, because processing a stack of documents is a different job from learning one. Choosing between those two families is the actual decision at the start of a pass, and it is why there are four lenses rather than one with a quality slider.
What we do not claim
We have no numbers about retention, comprehension or reading speed, and we are not going to invent any. What is claimed here is narrower and checkable: the bar is set high, the pass tells you where it set it, and you can delete every annotation you disagree with — they are ordinary marks sitting next to your own.
If it stops in the wrong places for you, the fix is not a setting. Save the ones that were right and dismiss the ones that were not, and the next pass over that document leans toward what you kept.
The four lenses are described on the annotation page, the book-scale version lives in reading a whole ebook, and the end-to-end workflow is read a book and end up with notes you can use. The mechanics are in reading and annotating.