TheChek News Flow

When you connect medical records and health data to a chatbot, you get results. But you must understand the risks.

As the lab technician prepared my vials, I told him, “I’m going to upload my blood work to AI!”

His response: “Oh boy.”

I agreed. I was more nervous about opening up my medical records to a chatbot than getting my arm pricked.

Others are already doing it. One-third of U.S. adults have turned to artificial intelligence for health advice in the past year, and of those, around 41% uploaded personal medical information to the chatbot, according to a recent survey from the health nonprofit KFF.

Data sharing is getting even easier. In recent months, the most popular AI apps—OpenAI’s ChatGPT, Anthropic’s Claude, Perplexity and Microsoft’s Copilot—unveiled ways to connect your wearables and medical records.

A headache can be benign or life-threatening depending on the context. In the white-hot AI race, a model that can help us navigate that medical complexity—and make sense of our growing pile of health data—could gain an edge.

I let Claude and Perplexity access my health data to search for hidden patterns in my personal metrics, and better understand what the risks are. After lots of prompting, I found that the bots are expert explainers, though occasionally prone to overstatement. I awarded the top chatbot a 🏆 in three scenarios.

And the experience drove home a truth I have long warned about: Gaining AI-enabled medical clarity means losing some privacy. And no, HIPAA doesn’t apply merely because these companies can access your health network’s data.

Sharing the data

Any chatbot will give you an answer to “My back hurts. What should I do?”

These new health-focused AI features offer more personalized responses based on everything you’ve shared. If, say, you’ve had a stomach ulcer, the bot would recommend a painkiller that’s easy on the gut.

OpenAI and Microsoft haven’t made their services widely available yet. Claude and Perplexity’s are in beta, but already available to paying U.S. subscribers (starting at $20 a month) right now. The integrations are off by default—you have to opt in. Sharing Apple Health data was as simple as flipping a switch.

To link my medical records, I had to identify myself to health data connectors that work with the AI companies. For Claude, HealthEx needed my birth date, phone number, face scan and a driver’s license for verification. For Perplexity, B.well took me directly to my medical provider’s portal to grant access.

I could select which categories of information to share with the AI bots, such as medications, and lab results. The more you give the AI, the more personalized the response can be, so I granted full access.

Prompting the bots

Accuracy is essential for health queries, and AI bots are known to make up information. So I consulted Dr. Sumant Ranji, a professor of clinical medicine at the University of California, San Francisco.

Ranji was clear he didn’t endorse this practice: “I would definitely not recommend that patients upload their medical records to any AI due to lack of privacy standards.” But he went along with my experiment.

🤖 Prompt #1: What’s my risk for cardiovascular disease? The bots scanned my lipid panels and Apple Watch heart-rate data. The verdict: I’m low risk. I preferred Claude’s scannable format, which used tables and emojis for clarity.

However, Ranji said Perplexity’s response was closer to how he counsels patients. He approved of the AI requesting age and other information for context, and said its “bottom line”—that long-term risk is never zero—was correct.

Claude’s advice, while similar, overemphasized data from my smartwatch, Ranji said. “VO2 max” is a health measurement that suggests how fit you are based on metrics like heart rate during activity. Claude thought mine was “trending volatile.”

Ranji dismissed the “scary-sounding” phrasing, noting that wearable data isn’t clinically reliable. An Anthropic spokesman said the AI is designed to help people navigate complex information and isn’t a substitute for professional medical care.

🏆 The winner: Perplexity

🤖 Prompt #2: What supplements should I take for improved sleep and fitness? For his own recommendation, Ranji consulted an AI bot popular among doctors: OpenEvidence. The free tool pulls its responses from medical journals.

OpenEvidence doesn’t have my information, but its advice to take magnesium and creatine was consistent with suggestions from Perplexity and Claude. It even gave extra female-specific advice, including omega-3 for post-workout recovery.

Perplexity and Claude did point out that my iron levels haven’t been tested, and if they’re low, a supplement could increase energy levels. While Claude suggested a few more supplements than Perplexity, their overall responses were similar.

🏆 The winner: OpenEvidence

🤖 Prompt #3: I’ve been fatigued for three weeks and woke up with a sore throat. Help? Both bots identified dips in my heart-rate variability data and an increase in resting heart rate from my smartwatch, which could indicate my body has been fighting an illness. Perplexity ruled out infection, based on my recent lab results.

Perplexity advised rest and fluids, while Claude also recommended zinc lozenges and saltwater gargles. And it reminded me to pause my Peloton training, recalling an earlier conversation about workouts.

Ranji said the advice from both bots was “practical.”

🏆 The winner: A tie!

Ranji pointed out that none of these cases were emergencies. A recent study in Nature Medicine found that, in early-access testing, OpenAI’s new ChatGPT Health had trouble with emergencies. In one instance, it referred a case of impending respiratory failure to overnight evaluation instead of an ER. An OpenAI spokeswoman said the study doesn’t reflect typical health conversations in ChatGPT.

Ranji also worried past conversations, such as one on homeopathic treatments, could steer the AI away from evidence-based results.

Perplexity said it keeps health chats siloed from general conversations. Anthropic said while Claude health conversations are part of the chatbot’s memory, any specific memory can be deleted in settings at any time. Claude is also designed to provide evidence-based health information, regardless of prior conversations, the spokesman said.

Health data is considered confidential to prevent insurance discrimination and employment bias.

Dr. Ami Bhatt, a cardiologist and the chief innovation officer of the American College of Cardiology, said this data is vulnerable when patients use chatbots, which don’t offer healthcare-grade privacy and security protections. Bhatt wants patients to understand before using chatbots how their health data will be stored or shared.

SHARE YOUR THOUGHTS

Have you ever consulted AI for health issues? How was the outcome? Join the conversation below.

After this experiment, I’ll purge any health-related memories and conversations. AI is moving fast, and I don’t want my data caught up in a version of the technology we haven’t met yet.

Anthropic and Perplexity say that users’ health data isn’t used to train models and can be disconnected at any time. (Anthropic offers a HIPAA-ready enterprise version of Claude for healthcare providers.)

Bhatt recognizes that the desire to turn to AI isn’t just the allure of chatbots—it’s a reaction to the current state of healthcare. “We have an access problem,” she says. “We have fewer clinicians than we do patients, and that disparity is going to increase.”

I relate. My results were normal so my doctor didn’t follow up. But I wanted to know more about my numbers, and AI filled that gap.