An AI Unplanned Pregnancy?

Details of the method and case have been changed to protect the privacy of the couple

Large Language Models (LLMs) like ChatGPT, Grok, Claude, etc, are being used at an increasing rate in a myriad of ways. Most people I speak with use these models like a giant search engine - a master Google that can google things as never before. Then there are more specific uses of trained LLMs that use Agent Harnesses (if you don’t know what that means, then ignore it) to speed up specific work tasks (seemingly) exponentially. I, myself, use specific AI tools for very certain tasks – and it has increased my work output in a way I never imagined. But behind the scenes, I am very careful not to outsource my thinking to AI. I don’t use it to brainstorm or analyze things I am educated in. And I never rely on it for ideas or concepts that I am not educated in. Unfortunately, many people have started using AI for that exact thing. Nothing could make that more apparent than a recent situation I encountered at my diocesan work.

I was contacted by a couple that was very frustrated because their NFP method “failed” them, and they were now pregnant when they were avoiding. They contacted me to plead their case that they had a special circumstance that would allow them to use contraception. Interesting aside, I am regularly contacted by people who think that I, or the Bishop, have the power to “waive the rule” on contraception. That could be a blog post all its own. Spoiler alert – it’s not a rule made by some old guy in the Vatican – the Church’s teaching on human sexuality and the Principle of Inseparability is God’s design for marriage and cannot be waived by anyone. Not me, not the Bishop, not the ethics committee at the hospital, and not the Pope. Now back to the story…. After listening to this couple, it did appear that they had what Creighton Practitioners would call a “method-related pregnancy,” meaning they got pregnant using the method while avoiding despite being taught correctly and following the instructions correctly. This is very rare in Creighton, but it can be more common in other methods.

Something about what the wife said made me ask a few additional questions, however. She said, “And I even ran it through [LLM], and it said I shouldn’t be pregnant.” Well, surprise, AI was wrong. Because she was 100% pregnant. I asked to see the chart and, according to the chart, there was no specific day that would have resulted in pregnancy. But being a trained expert in calculating the estimated time of conception, I used their pregnancy dating ultrasound to narrow down some potential days she could have conceived. I went through each day, one by one, and asked them to tell me about their observations, what they thought the instruction was, and their behavior. After a lot of back and forth, I found out that she used an LLM to help her determine if a specific day was fertile because it was much earlier in her cycle than normal, so she didn’t believe it was correct. She told the LLM all the reasons it shouldn’t be fertile (previous cycle length history, age, medical conditions, medications, what the observation looked like, how many (or few) times she saw it, etc.), and the LLM determined that her likelihood of getting pregnant on that day was likely zero. It even told her what to put on the chart. So, she wrote down that observation. But just to be “safe,” they waited two days before intimacy. If you’re an experienced charter in practically any method, you know that you can still get pregnant two days after a fertile sign. The mystery was solved. She outsourced her knowledge of their NFP method to an LLM, and it gave her bad information. When I troubleshoot unexpected pregnancies, I always classify them for the couple. The most common classifications are “condom-related pregnancy” or “withdrawal-related pregnancy.” These are when couples use condoms or withdrawal on days of fertility in addition to their NFP method, thinking that they can’t get pregnant using them. This also could be an entire blog post, but I will resist. In this case, I classified it as an “AI-related pregnancy.” Wow. The times they are a-changin’, Bob (if you know what I’m talking about, then you can appreciate my age).

Now I don’t want to just leave it at that. I want to educate you on why using AI in this manner is likely to give you this scenario. This is not a one-off. I will argue that this is an expected outcome. The majority of LLMs that people use are open, meaning that they aren’t trained on any specific thing. They use ALL the information on the Internet. Even if you give an LLM an instruction ahead of time like “looking only at Marquette/Creighton/Sympto-Thermal, tell me X” they cannot give you accurate information. They are not trained in specific models of NFP. They cannot ignore the variety of forums, blogs, and opinions of NFP users and teachers on the internet who have posted scenarios. They also cannot ignore all the published research that lumps those methods together and assesses their methodology and effectiveness rates. Additionally, the teaching methods and materials of most methods are proprietary, and the instructions for all the different methods are not widely available on the Internet. This is especially true for Creighton and Marquette. I posted a 2-part blog on the extensive training your Creighton Practitioner goes through to teach you.

But to be sure that my assumptions were correct, I decided to run some scenarios through several LLMs to see the answers. Using the Creighton Model FertilityCare System, I told them I only wanted information from education programs, certified practitioners, etc. I told them I was a Certified Creighton Practitioner and that I wanted to see if it was a useful tool in analyzing observations for my clients. Each of them was completely incapable of giving me correct information. Every time they would give me an incorrect output, I would correct it, explain why it was wrong, and narrow the parameters. Each time I did that, the responses became more incorrect. So, I broadened the parameters. I even started over, using a different browser login (so they couldn’t use previous chats as data) and pretended to be a client. Each time, they were completely wrong. They consistently mixed other ovulation models with Creighton. Among the many, many errors, they almost ubiquitously conflated Lubrication with smooth. Commonly, they assigned the Vaginal Discharge Recording System incorrectly. At one point, I submitted a picture of a 10KL (from the picture dictionary) to see what it would classify it as. The result? Tacky, cloudy. Let’s put this into perspective. If you see one day of tacky/cloudy (8C) pre-peak (and you are not in a 3 count already), then the next 2 days are infertile. But I showed it a 10KL. So even one day of 10KL pre-peak requires a 3 count. If you wrote 8C and not 10KL and had intercourse 2 days later, then you just had intercourse on a day of fertility. Yikes.

Another common thing I ran into in my research was that every LLM consistently reported that Creighton uses sensation when walking around and also internal examinations. Both are completely inaccurate. This was an extremely eye-opening experiment for me. I knew that LLMs were not capable of accurately analyzing charts based on my extensive knowledge of how they work. However, I never suspected they would be SO bad – especially when given good parameters. So, what do I recommend? I recommend that you do NOT use LLMs (AI) to assist you with charting. If you’ve been doing it already, then stop it. Seriously. If you have already learned a method from a qualified instructor, then apply that knowledge and don’t outsource it to an LLM. If you haven’t, then learn a method from a qualified instructor and apply the instructions you learn. To get started with Creighton, contact one of our practitioners today!