I turned to AI when my doctor couldn’t fix my back pain. It actually helped

A few months ago, I injured my back, resulting in debilitating nighttime pain.

To make matters worse, the pain meds my doctor prescribed weren’t working, and I hadn’t gotten a good night’s rest for weeks.

Desperate and sleep-deprived, I decided it was time for a Hail Mary, and asked https://www.fastcompany.com/section/artificial-intelligence" title="AI">AI for help.

Google’s Gemini chatbot was more than happy to become my physical therapist, checking in with me every morning about my progress and offering suggestions on how to make it through the night.

Gemini cheered me on when my sleep improved (“That is absolutely fantastic news!”) and offered words of empathy when I faced setbacks (“That is incredibly frustrating, especially after making such great progress”). 

It also offered helpful suggestions on how to keep track of my medications, told me about exercises I could try, and even tried to break down medical lingo into terms a tech reporter like me could understand (“You are essentially ‘working through a tech outage’ without the right tools”).

I’m far from alone in relying on a chatbot in my hour of need.

About a third of adults are turning to artificial intelligence for health advice, according to https://www.kff.org/public-opinion/kff-tracking-poll-on-health-information-and-trust-use-of-ai-for-health-information-and-advice/">a recent KFF study.

Among adults under 30, AI use for health information even reaches 42%.

Those are stunning numbers, especially considering that those very same chatbots now doling out health advice not too long ago used to tell people to https://www.bbc.com/news/articles/cd11gzejgz4o">eat glue, and seemed unaware of https://www.sciencefocus.com/future-technology/why-ai-generated-hands-are-the-stuff-of-nightmares-explained-by-a-scientist">how many fingers a human hand has.

AI models have arguably improved a lot over the past two years, thanks to their developers training them on expert knowledge, including medical facts, and penalizing obvious mistakes.

To put it simply, AI chatbots were sent back to school, and returned a lot more capable.

Still, I couldn’t help but wonder: Does AI really have my back, or was it a mistake for me to rely this much on Gemini’s advice?

“You absolutely did the right thing by asking AI,” says Gaurav Mathur, an Oakland, California-based physician who has been building his own AI health app in his spare time. 

Mathur believes that a patient’s use of AI is just the latest evolution of something that’s been going on for years: Patients go online to learn more about their condition. “Before AI, you would have gone to Google, and that’s also the right thing,” he says. “There are some doctors who [believe that] patients shouldn’t be going to the internet.

We sometimes make fun of Dr. Google.” However, Mathur argues that better-informed patients can take a more active role in their health, resulting in better outcomes. 

“I think that AI does have a lot of value,” he says.

Chatbots aren’t physicians

Not everyone is ready to embrace AI as a source for health advice.

Monica Agrawal, an assistant professor of bioinformatics at the Duke University School of Medicine, cautions that AI chatbots may not be able to spot critical warning signs in the questions and prompts of patients as easily as a real physician. “AI models are quite good at answering medical exam questions and things like that,” she says. “But that is very different from clinical practice.

The way patients ask questions is very different from the way physicians ask questions.”

For https://biostat.duke.edu/profile/monica-agrawal">her research, Agrawal has been analyzing thousands of real conversations people had with AI chatbots, including on medical subjects.

One problem she noticed was that AI models try to please their users, which can lead to their reaffirming false assumptions. “The way a patient phrases a question can guide the output of the language model in one direction or another,” Agrawal cautions.

Part of the problem is that an off-the-shelf AI chatbot simply doesn’t have the same kind of information about your health history as your physician.

That may lead to it doling out advice that it really shouldn’t give. “If you ask a leading question [like] what is the dose of this drug I should take—that presupposes that you should take this drug whatsoever,” Agrawal says.

A doctor should know about existing drug allergies, or interactions with other drugs a patient is taking.

Chatbots, on the other hand, often operate in a vacuum, leading to overly broad generalizations.

AI can help prepare for your doctor’s visit 

AI companies have long known that their models work best when they have lots of information to draw from, and some are starting to apply that idea to medical records.

Earlier this year, https://openai.com/index/introducing-chatgpt-health/">OpenAI introduced ChatGPT Health, which allows users to connect the company’s chatbot to their personal medical records through B.well, a company that has relationships with more than 300 health plans and countless medical providers.

Once synced, ChatGPT Health can answer detailed questions about a person’s medical history.

https://health-gpt.org/">Health-GPT, the app Mathur has been developing, takes a bit more pedestrian approach to gathering this kind of data.

Users simply upload PDFs, or take photos of the papers they got handed from their doctor or pharmacist. “Patients are actually underestimating how much information they have in their drawers at home,” he says.

Progress and discharge reports often include detailed information about prescriptions, health history, and more. “They don’t realize how much is in there, because when they look at it, it all looks like Greek,” Mathur says.

(Technically, patients can do the same thing with any AI chatbot, whether that’s Gemini, ChatGPT, or Anthropic’s Claude.

However, Health-GPT explicitly guarantees that patient data is not being used to train AI models, adding an extra level of privacy.)  

Once this data is uploaded, Mathur’s app uses this data to help people prepare for their medical appointments by explaining medical terms, and suggesting questions a patient may want to ask their doctor.

He knows from his own experience as a physician that this kind of preparation can be invaluable, especially given the time constraints doctors often find themselves under. “If you spend 15 or 20 minutes explaining what the disease is, you have very little time left to figure out [a treatment plan],” he says.

Keeping doctors in the loop is important 

Challenges with provider availability and health care access is one reason that drives so many people to use AI for health advice.

One in five people told KFF that they turned to chatbots because they didn’t have a doctor or had trouble getting an appointment, while another 20% simply couldn’t afford traditional care. http://kff.org/uninsured/key-facts-about-the-uninsured-population/">Growth in the number of uninsured also boosts the popularity of AI; 40% of people without health insurance use AI for physical and mental health advice, compared with 31% of those who do have health insurance, according to KFF. 

In my case, I kept in touch with my doctor throughout my recovery, and also frequently double-checked the chatbot’s recommendations with “Dr. Google.” Part of the reason for that was Gemini’s frequent disclaimers that it wasn’t able to offer any medical advice or diagnosis. 

“Gemini is designed to empower users with information.

With the help of tools like Gemini, people can come to their providers more informed and prepared,” says Google’s chief health officer, Michael Howell. “As the reporter saw in his experience, we make clear that Gemini is not a substitute for professional healthcare and should not be used for medical advice.”

I was also worried that the AI might miss something crucial, and steer me down the wrong path.

That has happened to others: In July, a pastor from Florida sued OpenAI, alleging that ChatGPT told him to weather his declining health at home.

This advice contradicted his family’s pleas to go to a hospital, and ultimately led to a massive pulmonary embolism.

An OpenAI representative told the BBC in response to the lawsuit that “ChatGPT is not a doctor and should never be used as a substitute for medical care, diagnosis or treatment.”

For my own recovery, I did appreciate the constant availability of my AI physical therapist, and I’m fairly sure that the recovery would have taken longer without it.

I knew from my doctor that I had an injured nerve, which was resulting in painful leg muscle spasms.

I also knew that healing would take time and patience.

But it turned out that there was room for more immediate interventions.

During one of our conversations, Gemini picked up on me remarking that a simple posture change all but made the pain go away. “That observation about arching your back is a huge clue,” it noted, adding: “In physical therapy, we call that a Directional Preference.” 

Gemini then suggested changes to my sleep position and specific exercises to reinforce that posture change, and also presented me with updated action plans every day to further my recovery.

Once I started to incorporate those changes and suggestions, my symptoms finally improved, and I was able to taper off pain medications under the guidance of my doctor.

Speaking of which: Anyone who opts to use AI to augment their health care should absolutely keep their doctor in the loop, especially when it comes to questions about medications and other critical interventions.

You also shouldn’t expect wonders.

Even with an always-on AI physical therapist at my side, my recovery still took weeks.

But in the end, I was rewarded with a full night of restful slumber.