

More than 40 million people ask ChatGPT health questions every single day, according to data OpenAI shared in early 2026 (NPR, 2026). So if you have ever typed "build me a workout plan" or "is this supplement worth taking" into a chatbot at 10 PM on a Sunday, you are in very large company. The question that matters is no longer whether people use AI for health advice. It is whether the answers deserve your trust.
Here is the honest short answer: AI is good at explaining health and fitness concepts, organizing information, and generating useful questions, and it is unreliable for specific numbers, for the studies it claims to cite, and for anything that depends on your personal medical history. Independent research backs this up. In one study, ChatGPT's exercise recommendations were 90.7% accurate against a gold-standard reference but only 41.2% complete, which means the answers were usually right and almost always missing something (Zinzow et al., 2024). In another, nearly half of chatbot answers to health questions were rated problematic, and not one chatbot produced a fully accurate list of references (Tiller et al., 2026).
That combination, mostly right on the surface and unreliable underneath, is exactly why "can I trust ChatGPT for health" has no single yes-or-no answer. What you can do is learn where the tool helps, where it fails, and how to check anything before you act on it. Those three things are what this article covers.
AI is at its best when you need shape and language, not decisions. If you have a vague goal like "I want to get stronger" but no idea what a sensible week could look like, a chatbot can turn that goal into a structured outline: which days to train, what a balanced session includes, and what a realistic progression might resemble. It is a planning skeleton, and skeletons are genuinely useful when you are starting from a blank page. The same goes for meal planning frameworks, grocery lists built around a budget, and weekly schedules that fit around work and family.
Vocabulary is the second strength. Chatbots are very good at translating jargon into plain English, and at adjusting the reading level until it clicks. If progressive overload, macronutrients, or energy balance have never made sense to you, asking an AI to explain it "like I am new to this" usually works better than a search results page, because you can keep asking follow-ups until it actually makes sense. You can even ask it to quiz you afterward.
The third strength gets less attention: question generation. A good time to use a chatbot is before a doctor's appointment or a session with a personal trainer, because it can help you organize what to ask and what to track. Adam Rodman, a hospitalist who researches AI at Harvard Medical School, put it simply in an NPR interview: there are "no downsides to better understanding your health," and using a chatbot to prepare before and after an appointment helps you partner with your provider instead of burning the visit on basics (NPR, 2026).
There is real evidence behind this "good at the basics" pattern. Zinzow and colleagues (2024) compared ChatGPT's exercise recommendations for 26 clinical populations against guidelines from the American College of Sports Medicine, which is the reference standard in that field. The recommendations were 90.7% accurate, meaning most statements did not contradict the guidelines. But they were only 41.2% comprehensive, and only 2 of the 26 answers cited a single source. Think of it this way: the surface is usually solid, and the depth is usually missing. A plan that is 90% accurate but 40% complete is a starting outline, not a program.
The failures fall into three predictable buckets, and knowing them puts you ahead of most people using chatbots right now.
Specific numbers. When a chatbot prescribes a precise dose of exercise, such as how many days per week of cardio a condition calls for, it can be confidently wrong. Zinzow et al. (2024) found exactly that: the most common inaccuracy involved medical clearance to exercise, and the bot overprescribed aerobic frequency for fibromyalgia and Alzheimer disease compared with the published guidelines. It also made errors on intensity, type, and progression. None of this means the tool is useless. It means you should not act on a specific number you cannot trace to a real guideline. It should be noted that this study tested an earlier model, and chatbots have improved since, but the pattern of "mostly right, sometimes wrong, rarely cited" is the pattern to expect.
Studies it invents. Tiller and colleagues (2026) audited five popular chatbots, including ChatGPT, Gemini, Meta AI, Claude, and Grok, with 250 health questions across five areas: cancer, vaccines, stem cells, nutrition, and athletic performance. Nearly half of the responses, 49.6%, were rated problematic, and 19.6% were rated highly problematic. Worse, hallucinations and fabricated citations meant that no chatbot produced a fully accurate reference list. The chatbots performed worst in exactly the categories that matter for this website: nutrition and athletic performance (Tiller et al., 2026). And the bots almost never flagged their own uncertainty. Out of 250 questions, there were only two refusals to answer. If a chatbot names a study, assume it may not exist until you have searched for it yourself and found it.
Anything that needs your medical history. A chatbot only knows what you typed, and most people type less than a clinician would ask. In a study published in Nature Medicine, researchers had ordinary people use chatbots on medical scenarios, and participants identified the hypothetical condition correctly only about a third of the time; only 43% made the right call on what to do next, like whether to seek emergency care (NPR, 2026). Wording changed everything. In one scenario, a person who described "the worst headache I've ever had" was told to go to the emergency room, while a person describing the same situation in milder terms was told to take aspirin and stay home. That first description was a life-threatening red flag (NPR, 2026). A separate study found chatbots treated true emergencies as less serious than they were in 52% of cases (NPR, 2026). Your job, if you use AI for health questions at all, is to keep it away from triage and diagnosis and bring those to a human clinician.
Large language models, the technology behind chatbots like ChatGPT, are prediction machines. They generate the most likely next words based on patterns learned from enormous amounts of text, not from a database of verified facts about your situation. Even OpenAI, the maker of ChatGPT, has acknowledged that its training and evaluation processes reward guessing over admitting uncertainty, and that hallucinations "remain a fundamental challenge" even in its newest models (OpenAI, 2025).
Confident tone is therefore a terrible signal of accuracy. You have probably noticed how polished and authoritative AI answers sound. That polish is free to produce, which is also why low-quality AI content, sometimes called AI slop, has flooded search results and social feeds. The average answer you meet online is not getting more trustworthy just because it reads well. Treat every chatbot answer the way you would treat a confident stranger at the gym: possibly helpful, worth hearing, and not something to act on without a second look.
Before you follow any plan, number, or claim an AI gives you, run it through these five questions. It takes about five minutes.
The single biggest upgrade most people can make is how they ask. A generic prompt like "give me a workout plan" gets you a generic plan, because the chatbot has to guess everything about your life. An engineered prompt is a carefully built request that supplies the context, constraints, and outcomes up front, so the AI has less to guess.
Compare the difference. Generic prompt: "Give me a workout plan." Engineered prompt: "Act as a certified personal trainer and create a 12-week progressive strength training program for a 35-year-old office worker with mild lower back sensitivity from prolonged sitting. I have a home gym with adjustable dumbbells up to 50 pounds, a resistance band set, and a yoga mat. I can train 45 minutes, four days per week. My goals are building functional strength, improving posture, and increasing energy. Include warm-ups, form cues for key exercises, progression rules, and modifications for movements that might bother my lower back. Tell me when to increase weights and how to track progress. Tell me what to verify against official guidelines, and refer me to a professional if any of this is out of scope."
The engineered version will produce a plan you can actually react to and refine, because it contains your equipment, schedule, history, and goals. Notice the last sentence, though. Engineered prompts fix relevance, not truth. A perfectly engineered prompt still deserves the five-question check above, and the guardrails in that last sentence ("tell me what to verify" and "refer me to a professional") belong in every health prompt you write.
Most courses either ignore AI or hand you a pile of copy-paste prompts and call it a skill. The Innova Vita Health and Wellness Foundations Course treats it as a literacy you build, which is why engineered prompts are a core part of the curriculum rather than an afterthought.
Here is what that looks like inside the course. You get detailed AI prompt templates for planning health habits, written so you can see every field that matters: your constraints, your equipment, your goals, and the guardrails that keep the output honest. Then a dedicated module teaches you how prompting works so you can engineer your own prompts for situations the templates do not cover. And because templates fix relevance but not truth, the course also covers what most people skip: what AI hallucinations are, what health information never belongs in a chat window, and how to verify an answer before you act on it. That module doubles as a misinformation course, teaching you to evaluate claims from any source, whether that is a chatbot, an influencer, or a supplement ad.
If you want to see the teaching approach first, we wrote a primer on how AI is used in our health and wellness courses, and the same engineered-prompt approach shows up in the course's interactive learning activities. Optional AI activities let you use a free chatbot to get feedback on hypothetical case studies, or to organize your questions before a doctor's appointment, so you practice checking an AI answer while the stakes are low. No paid AI subscription is needed, since the templates work with the free tier of most chatbots.
The course was built by Chris Bigelow, M.S. Kinesiology, a college educator with more than a decade teaching health and exercise science, and it assumes no science background. If you want to start smaller, any of the free mini-courses is a good taste of the format, and completing one earns a discount on the flagship course.
Use AI to prepare, not to decide. Ask chatbots for explanations, outlines, and questions to bring to a professional, and keep diagnosis, triage, and treatment decisions with humans. An easier version: next time you have a health question, ask the chatbot "what should I ask my doctor about this" instead of "what should I do about this."
Upgrade one generic prompt into an engineered prompt today. Take the fields from the example above (role, context, equipment, schedule, goals, guardrails) and rebuild a prompt you have used before. If that feels like a lot, start by adding just two things: your real constraints and one guardrail sentence.
Run the five-question check on anything you plan to act on. Keep the list somewhere you will see it, like your phone's notes app. If five feels like too many, at minimum run questions one and three, since a fake source and a guideline contradiction are the fastest red flags to catch.
Verify every number against a named source. Any calorie target, heart-rate zone, or training volume the AI gives you should be checked against a calculator or guideline you can name. Our free macronutrient calculator and BMR calculator are useful cross-checks for the nutrition side.
Keep your medical details out of the chat window. Treat a chatbot like a public forum: hypotheticals in ("a 35-year-old with mild lower back sensitivity"), personal identifiers out. If a plan needs your full history to be safe, that is a sign the conversation belongs with a doctor or a registered dietitian.
NPR. (2026, March 11). ChatGPT might give you bad medical advice, studies warn. https://www.npr.org/2026/03/11/nx-s1-5744035/chatgpt-might-give-you-bad-medical-advice-studies-warn
OpenAI. (2025, September 5). Why language models hallucinate. https://openai.com/index/why-language-models-hallucinate/
Tiller, N. B., Marcon, A. R., Zenone, M., Kidd, K. E., Jeukendrup, A. E., Master, Z., & Caulfield, T. (2026). Generative artificial intelligence-driven chatbots and medical misinformation: An accuracy, referencing and readability audit. BMJ Open, 16(4), e112695. https://doi.org/10.1136/bmjopen-2025-112695
Zinzow, A. L., Burns, R., Colwell, K. J. T., & Pescatello, L. S. (2024). Comprehensiveness, accuracy, and readability of exercise recommendations from ChatGPT for 26 clinical populations: Mixed methods study. Journal of Medical Internet Research, 26, e52993. https://doi.org/10.2196/52993
AI and the Future of Personal Wellness — The first article in our series on AI and personal health, and the foundation this article builds on.
Behind the Scenes: Lifestyle Chronic Disease Prevention — How we designed the flagship course's prevention module and where AI fits into it.
Navigating Scientific Literature in Health and Fitness — How to judge the quality of any source, including the studies chatbots cite (or invent).
Evaluating Health and Fitness Information — Our hub for spotting bad advice, from logical fallacies to survivorship bias in fitness marketing.
This article is for education only and is not medical advice. For personal health decisions, talk with a qualified healthcare professional.

Chris Bigelow holds a master’s degree in kinesiology from A.T. Still University and has taught kinesiology and exercise science at Bryan University for more than 10 years. He founded Innova Vita Fitness to make evidence-based health, fitness and research education more accessible to adult learners. His work focuses on exercise science, health literacy, sustainable behavior change and helping beginners evaluate health information more confidently.
Click Here