A patient asked me recently if it was safe to ask ChatGPT for advice about her new bisphosphonate prescription. She had a long list of questions, her appointment was three weeks away, and she wanted something to do in the meantime.
It is a fair question, and one I think more of us should be asking out loud.
I use AI tools every day, both as a physician and as someone living with osteoporosis. They are genuinely useful for translating medical jargon, organizing questions before an appointment, and giving you a starting point on an unfamiliar topic. They are also, sometimes, confidently wrong, or quietly incomplete, in ways that are very hard to detect if you do not already know the answer.
Here is what I have learned about getting useful answers without getting burned.
The single most important habit: favor tools that show their sources
Not all AI tools work the same way, and most people have no idea which kind they are talking to. Some tools are essentially very fluent autocomplete, predicting what a reasonable answer might look like based on patterns in their training data, which may be a year or more out of date. Others search the live internet first and then summarize what they find, with links so you can verify each claim.
For health questions, the difference matters enormously.
A tool that shows sources can still be wrong, but the wrongness is auditable. You can click the link, read the original, and decide whether it actually says what the AI claims it says. A tool with no sources gives you a polished paragraph that sounds authoritative whether the underlying facts are correct, partially correct, or quietly missing something important.
How to tell which kind of AI you are using
The simplest way is to just ask. Before you trust an answer on anything medical, paste in a prompt like one of these:
- “Are you answering from your training data, or are you doing a real-time search of the web?”
- “Can you answer this with links to sources from a real-time search?”
- “What is the date of the most recent information you are using to answer this?”
A good tool will tell you honestly. If it admits it is working from training data with a cutoff date, you now know you are looking at an educated guess that could be a year or two stale. If it offers to search and cite sources, ask it to do exactly that and then check the links.
You can also watch for visual cues in the interface. Tools that are actually searching the web usually show you a “searching” or “browsing” step, and then list numbered citations next to claims in their answer. Tools that go straight from your question to a polished paragraph, with no search step and no links, are almost always running on training data alone.
When I am researching anything medical, I now strongly prefer tools that cite as they go, and I ask explicitly when I am not sure.
A small experiment you can run yourself
To show you what I mean, here is what happened when I recently asked two major AI tools the exact same simple question that thousands of newly diagnosed osteoporosis patients ask every day:
“How much calcium should I get per day if I have osteoporosis, and should I get it from food or supplements?”
The results, captured in June 2026, were instructive.
Both tools gave fluent, confident answers. Both recommended the standard 1,000 to 1,200 milligrams per day depending on age and sex. Both said to prioritize food over supplements. Both mentioned vitamin D. Both reminded the reader to talk to their doctor. By every surface measure, the two answers looked equally good.
But two things were quietly different.
Only one of the two tools flagged the cardiovascular concern. It included a short note that dietary calcium has not been linked to the same potential cardiovascular risks (such as arterial calcification) that some research has associated with high-dose calcium supplements. This is not an obscure detail. It is one of the most clinically important nuances in the entire conversation about calcium supplementation, and it is the reason many physicians now favor a food-first approach even more strongly than they did a decade ago. The other tool simply omitted it.
Only one of the two tools cited where its information came from. It listed eight sources, including the International Osteoporosis Foundation, MedlinePlus, and several major academic medical centers. A reader could click any of them and verify a specific claim. The other tool listed no sources at all. Its answer might have been drawn from current guidelines, an outdated article, or some mix of both, and a reader had no way to tell.
If you had only used the tool that omitted the cardiovascular point, you might walk away thinking the choice between food and supplements was purely a matter of absorption and convenience. You would not know there was a safety nuance worth asking your doctor about. And without sources, you would have no way to even check.
Both answers sounded equally trustworthy. Only one was complete enough to act on.
(One caveat worth flagging: AI behavior changes month to month as models retrain and as new information enters their search indexes. The specific results above are a snapshot in time, not a permanent verdict on any tool.)
How this maps to your osteoporosis questions
If a tool can quietly omit a major safety nuance on a basic question like calcium dosing, imagine what it can do with questions like:
- “Is it safe to take alendronate if I have a history of acid reflux?”
- “What are the side effects of romosozumab?”
- “Can I take vitamin K2 with my prescribed osteoporosis medication?”
- “How long should I be on a bisphosphonate before considering a drug holiday?”
A confident, fluent, unsourced answer to any of these could be perfectly correct. It could also be subtly out of date, mixing up two related medications, missing an important contraindication, or quietly skipping a safety point the way the calcium example skipped the cardiovascular one.
For questions where the wrong or incomplete answer could lead to a fracture, a serious side effect, or a worse long-term outcome, I would never act on an AI response that does not show me where its information came from.
Five rules I follow when asking an AI about my own health
1. Favor search-grounded tools for medical questions. If the tool is not pulling from live sources and showing you the links, it is making a more educated guess than you might realize. Use those tools for brainstorming questions to ask your doctor, not for medical decisions.
2. Ask for sources explicitly, every time. Even with search-grounded tools, prompts like “cite each claim with a link” or “what is your source for that?” tend to produce more careful, traceable answers.
3. Click the sources. I cannot stress this enough. AI summaries occasionally describe what a source says incorrectly, or extrapolate beyond what the source actually claims. The two-second habit of opening the linked page and skimming it catches most of these errors.
4. Ask the same question to two different tools. This is the single best trick I know for spotting when something has been quietly omitted. If one tool flags a nuance the other skipped, that is information worth bringing to your physician.
5. Treat any answer without a verifiable source as a hypothesis to test, not a finished answer. Bring it to your physician or pharmacist as a starting point for a real conversation.
What AI is actually great at, for someone with osteoporosis
I do not want any of this to sound like a blanket warning against using AI. It has changed how I research and prepare for appointments. A few of the places I find it most useful:
- Translating jargon from a clinical note or lab result into plain English you can actually act on
- Preparing for an appointment by drafting a list of questions based on your specific situation
- Drafting a follow-up email to your physician when you forgot to ask something during the visit
- Comparing two medications side by side in a way that is easy to skim
- Explaining a concept to a family member who is worried about you and needs to understand the basics
For all of these, the worst-case downside is that you get a slightly imperfect first draft, which you can refine. That is very different from asking “should I take this medication” and getting a confidently incomplete answer with no way to check what was left out.
The takeaway
The technology is real, the usefulness is real, and so is the risk of confidently incomplete information. The simplest way to protect yourself is to use tools that show their work, click through to the sources, ask the same question twice, and never treat an unsourced AI answer about your health as a final answer.
A good physician will not be threatened by a patient who arrives with a thoughtful list of AI-generated questions. A good AI tool will not pretend to know what it does not know.
You deserve both.
If you want an example of using AI well, paired with the lived human experience of this diagnosis, that is what my book is. Osteoporosis: The Book I Wish I’d Had When I Was Diagnosed is the clinically grounded, emotionally honest overview I wish someone had handed me on the day I was diagnosed.