Debugging my doctor: Using ChatGPT to fix a harmful prescription

This is the worst thing that ever happened to me—and AI is the only reason I got out of it safely.

I recently wrote about New York’s S7263 and New Hampshire’s SB640—two bills that would substantially limit how regular people can use AI to explore their health. In that piece, I shared some personal stories that illustrate why access to this information is so important. Today, I wanted to share another, more detailed personal story about how AI was critical in my healthcare.

Most of my health journey happened before ChatGPT and involved doctors, family friends, my own medical research, and online forums. I didn’t expect to rely on AI much—my issues had largely been resolved. Unfortunately, I finally had an opportunity to use AI extensively for a medical problem.

Misdiagnosis

Last year, I had just tapered off duloxetine—an SNRI I was prescribed for fibromyalgia. Before my fibromyalgia diagnosis at age 17, I was constantly getting infections. The disease is known to cause low serotonin, which in turn can cause low immunity. Duloxetine mediates this by increasing serotonin, and it effectively stopped me from having colds half of every year. After solving some of my underlying medical issues a year ago, my fibromyalgia symptoms disappeared and I no longer needed the medicine.

But at the same time, I was going through a rough personal situation, and ended up feeling depressed—something I hadn’t really experienced before. I wasn’t sure if those feelings were because of the situation or, I wondered, maybe all these years duloxetine had kept a latent depression at bay. Another possibility was withdrawal—my brain adjusting to the absence of the medication.

I eventually made an appointment with a psychiatrist to figure this out. She told me that because I talk fast, was experiencing depression, and because my mom has bipolar disorder, that she believed I might have “cyclothymia,” sometimes regarded as a milder form of bipolar disorder. Although I was skeptical of her diagnosis due to my lack of mood swings, she warned me that the condition would worsen if I didn’t treat it, and assured me that there was no harm in trying the medication.

That did not turn out to be the case.

It had horrendous effects. Lurasidone blocked most of my brain’s dopamine receptors, which caused a long list of severe symptoms and blocked my brain’s ability to think clearly. I won’t get into the details of how bad it was, but it was the worst thing I have ever experienced—worse than any of my autoimmune diseases. I then found out from my primary care provider that the psychiatrist had documented a medical history I did not recognize and had not reported in order to justify her diagnosis (AI isn’t the only one hallucinating). So, I switched to another psychiatrist who told me that the medication was doing no harm, but that the diagnosis was wrong and I didn’t need that medication. This new provider cut the dose in half, and an already severe set of symptoms became catastrophic.

Until this point, I had actually believed that my own emotions were responsible for the new range of mental and physical symptoms I was experiencing, because nothing else could explain them. Quick research had shown me that lurasidone “balances” dopamine—but in reality, it blocks dopamine receptors. That can indeed balance dopamine activity in a brain with too much of it or too many/too sensitive dopamine receptors, but not in a brain where dopamine behavior is normal, like mine. Because of my own lack of knowledge here, I didn’t understand the medicine could cause the symptoms I was experiencing.

Undoing the harm with ChatGPT

I then realized the new doctor was wrong: not only was the medication causing all the new strange symptoms I developed, but reducing the medication that much at once was making them far worse. She went so far as to say that the medication couldn’t cause what I was experiencing, despite the medicine’s warning that these were potential effects. This same doctor also encouraged me to take benzodiazepines as regularly as I needed to while I was going through the withdrawal instead of extremely sporadically, only when I couldn’t handle without. That class of medicine can become dependence-forming after mere days and, if I had listened to her, I might have had to figure out how to taper off of this, too. Thankfully, this time I researched it more in-depth before following her advice.

Because my doctors did not seem to know what they were doing, I consulted ChatGPT and learned about “hyperbolic tapering”—a method of decreasing each next dose in a taper by a small percentage of the last dose (this can be, say, 10% or even 5% of the last dose instead of 50%). The method first emerged on online forums and has since been shown to be effective in both preventing relapse and minimizing withdrawal symptoms in various studies. Psychiatric drugs affect receptors in a nonlinear way: small dose reductions at low levels can have much larger effects than bigger reductions at high doses.

Moreover, through research surfaced for me by ChatGPT, I learned that the issue with going off the medication too quickly was that my body couldn’t handle unblocking so many receptors at a time. ChatGPT found me a study that mapped dopamine receptor occupancy (what percent of your body’s receptors are blocked by the medication) to dose. I then asked ChatGPT to extrapolate this data into four possible curves, and provide me the receptor occupancies for different doses in order to plan out my next dose reductions and ensure each one doesn’t change receptor activity on any curve by too much.

It worked.

I even used ChatGPT to log how often and what doses of the prescribed benzodiazepine I was taking in order to help me gauge the possible risk for dependency. Because my provider was not nearly risk-averse enough, this was a more helpful check. This worked effectively, too.

I’m fine now, with only modest symptoms accompanying each dose reduction and very mild symptoms from the drug itself at the lower doses. But to be clear, I am still in the process of reducing doses and it will have taken a total of eight months to come off of this medication.

Doctors are still in charge

For all the concerns about AI guiding health decisions, it’s important to note that I could not have executed this approach without the doctor agreeing to prescribe to my graph. On my own, I could not have cut a 20mg pill into 8.5 or 3mg doses. AI could not have called in the prescription to the compounding pharmacy—it could only work with me to make a plan to suggest to my provider. Thankfully, the doctor agreed to follow this plan. And soon after, I found an extremely knowledgeable new provider who understands pharmacology better, was already familiar with hyperbolic tapering, and has read the same research that ChatGPT had shown me. I explained how I made the tapering chart, and he agreed that it made sense and is continuing to prescribe the medication in accordance with it.

Without ChatGPT’s help, I likely would have followed standard tapering advice: cutting too quickly, worsening symptoms, and potentially causing lasting harm to my body. It is possible that I would have come off the rest of it cold turkey—I was already at the smallest prescribed dose—and suffered through the rest of the reduction which would have unblocked five times the amount of receptors that the first cut unblocked.

That’s what’s at stake when access to AI for health questions is restricted. This isn’t so much about my own singular experience, but all the other people who will benefit from taking charge of their own health using new technology.