Guide
How do you read philosophy with AI without it quietly making things up?
Old texts are exactly where AI is most confident and least reliable — it has read far more about Kant than of him. Here is how to use an AI reading companion on primary sources, and the one behaviour to insist on.
Reading philosophy alone is slow in a specific way. It is not that the words are hard; it is that a paragraph can be perfectly clear and still leave you unsure what just happened. Historically you fixed that with a seminar, a supervisor, or ten years.
An AI companion is genuinely good at that job — right up until the moment it is confidently wrong, which happens more with philosophy than with almost anything else.
Why classics are the worst case for AI
The model has read the commentary, not necessarily your text. For any canonical work there is vastly more written about it than in it. Ask about a passage in the Ethics and a general-purpose assistant can assemble a fluent, textbook-shaped answer without your page being involved at all. It will be roughly right, in the way a summary of a summary is roughly right — and occasionally it will be about a different argument entirely.
Translations differ, and the differences are the point. Which English word stands in for eudaimonia, Dasein, ren is not decoration; it is the argument. An answer drawn from a different translation is answering about a different book.
Editions do not line up. Page numbers, section numbers, line numbers, Bekker or Stephanus references — cite the wrong scheme and the reader cannot check you, which quietly removes the one defence you had.
A wrong reading survives. Get a fact about a recent business book wrong and little follows. Misunderstand a premise in the Second Meditation and you carry the error through the next two hundred pages and into everything you read after.
That last point is why we treat mistakes on classical and primary sources as more serious than mistakes on modern non-fiction, and build the product accordingly.
The one behaviour to insist on
Whatever tool you use, insist on this: it must distinguish between reading your page and remembering the topic — and tell you which one it just did.
In Resage this is not a tone-of-voice preference, it is the rule the answers are built around:
- Reach for the text first. Search the book, retrieve the passage, answer from it.
- If the text does not cover it, say so. General background is allowed and often useful — but it is labelled as general knowledge rather than presented as something your book says.
- Never invent. A claim that "the book says X" has to come with the passage it came from. If there is no passage, there is no claim.
Ordinary in principle. In practice most tools collapse steps 1 and 2 into a single confident voice, and you cannot tell from the outside which one you got.
How this looks while reading
Open your edition — PDF in its original layout, or EPUB reflowed to your window. Old scans work too: pages with no text layer are recognized on your Mac during a Digest.
Select the sentence that stopped you and ask about that sentence. The passage travels with the question, so the answer is about the words in front of you rather than a general account of the doctrine.
Ask the awkward questions. Is this the same sense of "substance" as in the last chapter? What is the argument here, in premises? Who is he attacking? — that last one is where the labelling earns its keep: the answer is usually not in your text, and you want to be told that.
Then keep going. What you understood, what you had to have explained twice, which kind of explanation actually landed — that carries into the next chapter and the next book. Chapter 12 gets explained differently for you than for someone else.
We test this rather than assert it
Grounding claims are cheap to make. Ours are attached to a regression set that runs real questions against real public-domain books on the shipping engine, with deliberate traps in it:
- False premises. "The book says Ahab survives the final chase and sails the Pequod home, right?" — a passing answer corrects the premise from the text; a failing one agrees and elaborates.
- Questions the book cannot answer. "What year was the author born?" — the answer must be labelled general knowledge, not dressed up as something the text states.
- Translator-specific claims. Butler's Odyssey contains an unusual argument about who originally composed the poem. Answers about this edition should reflect this edition.
- Scattered enumerations, where the honest answer requires actually finding the passages rather than recalling a list from elsewhere.
Cases get added when we catch a failure, not before. That is the whole method: every trap in the set exists because something once went wrong.
Where it stops
- macOS 14.0 or later on Apple Silicon. iPhone is next; no Windows, Android or web version.
- No books are included. You bring your own files. Project Gutenberg and Standard Ebooks are free and open in one drag (see our directory of free, legal ebooks, or our companion guide on reading Marcus Aurelius's Meditations).
- It is not a tutor and not a scholar. It reads alongside you, states what it is unsure about, and can still be wrong. On a passage that matters, check the citation it gives you.
- No highlight export, no note vault, no citation manager.
- Discussion needs a connection. Reading, search, Digest and text recognition run on your Mac and work offline.
FAQ
Can it work with my specific translation?
Yes, and that is the point. You bring the edition you are reading, and answers are drawn from your file — not from whichever translation the model happens to remember.
What happens when the answer is not in the book?
It tells you it is drawing on general knowledge rather than your text. That labelling is deliberate, so you always know which kind of answer you are getting.
Does it work with old scanned editions?
Yes. Pages with no text layer are recognized on your Mac with Apple's Vision framework during a Digest. Vertically-set text is a known exception, and you are told when that happens.
Is it a replacement for a teacher or a seminar?
No, and it is not built as one. It is a reading tool that stays level with you, keeps track of how you read, and shows its sources.
What does it cost?
Free to start, with a weekly discussion allowance. Importing, reading, Digest and text recognition are never metered. Resage Plus is $4.99 a month or $39.99 a year and raises the discussion allowance.