A discharge summary, a radiology report, or a lab printout is not written for the patient. It is written by one clinician for another, in a shorthand built over decades of medical training - abbreviations, reference ranges, and phrasing like “unremarkable” or “no acute findings” that read as ambiguous or even alarming to someone encountering them for the first time. The instinct to paste the whole thing into a chatbot and ask “what does this mean” is understandable. It is also the exact point where this task needs a firmer method than a single open-ended question.
This is a four-column method for working through a medical document: the exact original text, a plain-language paraphrase of the language, what remains genuinely uncertain, and a specific question for your clinician. AI is useful for the paraphrase column. It should never be asked to fill in what a finding means for your case, whether a value is concerning, or what to do next.
Do not ask AI to tell you whether a result is normal, what a finding suggests about your health, or what you should do in response. A paraphrase of medical language is not the same as a medical opinion, and a model has no way to weigh your result against your full history, your other conditions, or your clinician’s own read of the same document.
Step 1: Work in short sections, not the whole document at once
Long documents invite a model to summarize broadly, which is exactly the wrong instinct here - a broad summary smooths over the specific wording you actually need preserved. Work section by section: one finding, one paragraph, one lab panel at a time.
Here is one section of a medical document: [paste the exact section].
Do not summarize or shorten it. First, repeat the exact text back to
me unchanged. Then, in a separate paragraph, paraphrase any clinical
terms or abbreviations into plain language. Do not tell me whether
anything in this section is normal, concerning, or requires action -
only translate the language.
Repeating the exact text back first, before the paraphrase, keeps you anchored to what was actually written rather than the model’s summary of it - and makes it immediately obvious if the model has silently dropped or altered a detail.
Step 2: Build the four columns for each finding
For each distinct finding, term, or value, capture four things side by side: the exact original wording, the plain-language paraphrase, what remains uncertain even after the paraphrase, and a specific question this raises for your clinician.
Here is the exact text and its plain-language paraphrase: [paste
both]. Help me write one specific, answerable question I could ask my
clinician about this finding - something that could reasonably be
answered in a sentence or two. Do not answer the question yourself or
suggest what the likely answer is.
The “uncertain” column is often the most valuable part of the worksheet. It is where you note things like “I don’t know if this reference range applies to me specifically” or “I don’t understand why this test was ordered” - exactly the kind of gap a paraphrase alone will not close, and exactly what a short appointment is for.
Step 3: Treat abbreviations and reference ranges with extra care
Medical abbreviations are especially prone to a specific failure: the same short string can mean different things in different specialties, or nothing standard at all outside its original context. Ask for the paraphrase, and separately ask the model to flag its own confidence.
This document uses the abbreviation "[abbreviation]" in this context:
[paste the surrounding sentence]. What does this abbreviation
typically stand for in this kind of document? If it could plausibly
mean more than one thing in this context, say so explicitly rather
than picking one.
Reference ranges printed next to lab values are calibrated by the lab, not by a general-purpose model, and they can vary by age, sex, method, and even by which lab ran the test. Do not ask AI to judge whether your value falls inside or outside a healthy range - that comparison, and what it means, belongs to whoever ordered the test.
Step 4: Keep a running list of clinician questions
As you work through the document section by section, the specific questions from step 2 accumulate into exactly the list you need for your next conversation. This connects directly to the appointment-preparation method - see preparing for a medical appointment in 20 minutes for turning this question list into a one-page brief, and after a diagnosis, organize questions if this document arrived alongside a new diagnosis.
A medical document is one of the most sensitive files you can hand to a consumer AI tool, and uploading a full PDF or photo is a different privacy decision than typing a short excerpt. Before uploading any document, check what the specific tool does with uploaded files, whether training is off by default, and whether a business or health-specific account with stronger data terms is available. See what ChatGPT remembers, sees, and shares for the general mechanics, and prefer typing short excerpts over uploading a full identifiable document wherever that is practical.
When paraphrase is not enough
Some documents genuinely need a professional reader, not a plain-language pass - a complex pathology report, an imaging report with several findings interacting with each other, or anything where the stakes of misreading are high. If a paraphrase leaves you with more uncertainty than clarity after two or three attempts, that is itself useful information: it means the document needs your clinician’s direct explanation, not another AI pass. Bring the original document to that conversation, not just your worksheet - the worksheet is preparation, not a replacement for the source.
The full medical document literacy worksheet gives you the four-column structure ready to use, plus a short checklist for what to do when a section resists paraphrasing.
Working from a photo or scan
If your document exists only as a printed page or a photo rather than typed text, decide how you will get the text into a workable form before you start. Typing a short section by hand is slower but keeps you looking closely at every word - often useful for the first pass on a document you know is dense. If you use a tool’s image or document upload feature instead, treat that as a separate privacy decision from typing text, and apply the same data-handling check from the checklist below before uploading anything that identifies you alongside clinical detail.
I am going to type out one section of a document from a photo rather
than uploading the image. Here is what I typed: [paste your
transcription]. Before we continue, does anything in my transcription
look like it might be a typo or misreading of a clinical term - flag
anything ambiguous rather than assuming I typed it correctly, but
don't guess at what the correct term should be.
Common pitfalls
- Pasting the whole document and asking “what does this mean.” This produces exactly the broad, interpretive summary this method is built to avoid - work section by section instead.
- Letting “normal” or “concerning” language slip into the paraphrase. If the output includes a judgment about a value or finding, delete it and note the gap in your uncertain column instead.
- Assuming one abbreviation always means one thing. Ask the model to flag ambiguity rather than silently picking the most common meaning.
- Comparing your own value to a reference range using AI. That comparison depends on context a model does not have - leave it to whoever ordered the test.
- Uploading a full document without checking the tool’s data handling. A quick excerpt typed by hand is often the safer choice for anything identifiable.
Try it today
Pick one section of a document you have been putting off reading closely. Work through the four columns - exact text, plain paraphrase, what’s still uncertain, and a specific clinician question - using the prompts above. Use the medical document literacy worksheet to keep the columns consistent as you move through the rest of the document.
