What AI Gets Wrong About Reproductive Health

The accuracy problem is worse than you'd think

Let's start with the numbers. ChatGPT's accuracy on medical questions varies wildly by specialty -- from as high as 95% in some areas to as low as 20% in others. For gynecologic oncology, the accuracy rate sits at about 45.6%. That's essentially a coin flip on questions about cancer.

But accuracy isn't even the biggest concern. It's the failure mode.

When AI gets a medical answer wrong, it doesn't say "I'm not sure." It delivers the wrong answer with the same confident tone as the right one. There's no hedging, no uncertainty signal, no clinical judgment about when to defer. A sub-50% accuracy rate delivered with 100% confidence is worse than no answer at all.

Fabricated references and invented science

The hallucination problem in medical AI is genuinely alarming. A study evaluating AI responses to cancer-related queries found a 39% hallucination rate. That means roughly four in ten responses contained information that was simply made up.

It gets more specific than that. One systematic evaluation found 592 hallucination instances across bibliographic fields in AI-generated medical content. These weren't misattributed citations. They were completely invented -- fake authors, fake journals, fake DOIs, fake findings. A physician checking the reference list would find papers that don't exist, cited in support of claims that may or may not be accurate.

In another case, ChatGPT invented gene identifiers when asked about BRCA genes -- the very genes involved in hereditary breast and ovarian cancer risk. It didn't misinterpret existing research. It fabricated genetic nomenclature that doesn't correspond to anything in the human genome.

These aren't edge cases. They're what happens when a language model optimizes for fluent, plausible-sounding output without any mechanism for verifying whether the content is true.

Patients are already acting on AI health advice

Here's the part that should concern every practitioner: people aren't just reading AI health responses out of curiosity. They're making decisions based on them.

Research shows users have changed their birth control method based on AI advice. Others have adjusted medication dosages. Some have delayed seeking care because an AI tool told them their symptoms were likely benign. These aren't hypotheticals. They're documented behaviors.

For reproductive health specifically, AI tools have misrepresented the safety profile of medication abortion and provided inconsistent guidance on fertility treatment options. Only 1.6% of published studies on AI in reproductive health are experimental -- meaning the vast majority of what we know about AI's performance in this space comes from observational or descriptive studies, not rigorous testing.

We're deploying these tools at consumer scale without having adequately tested them in the domain where patients are using them most.

The one-size-fits-all problem

Even when AI gets the facts right, there's a more fundamental mismatch with how restorative reproductive medicine works.

AI gives population-level answers. Ask it about endometriosis treatment, and you'll get the standard options ranked by how commonly they're prescribed. Ask about irregular cycles, and you'll get the textbook differential. The responses are generalized, aggregated, averaged across millions of data points.

RRM's entire model runs in the opposite direction. Each woman's Creighton chart is unique. Treatment protocols are built from individual charting patterns, hormone profiles, surgical findings, and the patient's own goals. A NaProTechnology evaluation for two patients presenting with the same chief complaint might lead to completely different treatment plans based on what their charts reveal over three to six months of observation.

That level of individualization is exactly what AI can't do. Not because the technology isn't sophisticated enough -- but because the methodology requires sustained observation of a specific person over time, interpreted by a trained clinician who knows that patient's history and goals. There's no shortcut to that, and a chatbot can't approximate it.

AI tools aren't useless. But the limits matter.

This isn't an argument that AI has no role anywhere in healthcare. For certain applications -- literature search, administrative tasks, population-level research analysis -- these tools can be genuinely useful. They're fast, they can process volume, and they don't get tired.

But for reproductive health, where the stakes are high, the medicine is individualized, and patients are making real decisions about fertility, pregnancy, and surgical intervention, the current accuracy isn't close to adequate. The technology gives confident-sounding answers to questions it doesn't understand well enough.

When your patient comes in and says "I asked ChatGPT about my treatment options," the right response isn't alarm. It's an opportunity to explain what individualized care actually means -- why her specific chart matters more than a generalized answer, and why the time you spend reviewing her cycles isn't something an algorithm can replace.

Your patients deserve better than a coin-flip accuracy rate on questions about their reproductive health. And they already have it. That's the point of the medicine you practice.