Sora AI Failed to Capture My IVF Journey

OpenAI's Sora is making waves in video production, promising the ability to generate high-quality clips from simple text prompts. As a writer currently going through IVF, I was curious to see if this technology could help me tell a deeply personal and scientifically complex story. My idea was to create an explainer video about the fertility industry, using Sora to generate B-roll footage and help me overcome my shyness on camera.

However, I quickly discovered that when fertility meets machine learning, a lot gets lost in the algorithm. My hands-on experience revealed Sora's significant limitations, especially when it came to accuracy and context. In short, it couldn't even spell the word "uterus."

My AI Directorial Debut

I started by giving Sora a broad prompt describing my project: "I am going through IVF as a reporter/writer. I want to produce an explainer video...create a montage of IVF clips, pull in news headlines and create custom graphics."

The result was strange, to say the least. The video showed an "embryologist" literally face-planting into a petri dish—something that would never happen with embryos worth tens of thousands of dollars. The clip was surrounded by nonsensical text like "Brakfoots of tecmofolitgy" and "Breaknctve tennology," clear giveaways of its AI origin.

A second attempt was slightly better but still felt machine-generated and lacked context. For instance, it showed a fertility doctor with her hair down and a stethoscope, details that don't align with the sterile environment of an IVF lab. The clips resembled generic stock medical imagery.

Drilling Down Into AI's Flaws

I realized I needed to generate one clip at a time. My prompt for an embryo developing into a blastocyst produced something that looked like a mass of cells, but it wasn't scientifically accurate.

Things got worse when I asked for a diagram of the female reproductive anatomy. Not only was the image inaccurate, but Sora also misspelled "uterus" as "uteryus."

At this point, it was clear Sora struggles to produce legible text and lacks a fundamental understanding of human anatomy.

I shifted from science to more emotional imagery, asking for "a close-up of a newborn, with golden light and purity." Finally, a success! The resulting clip looked accurate.

But this success was short-lived. When I asked for clips of a newborn's feet, Sora produced images with the wrong number of toes—a classic AI struggle. It also added extra hands in a later prompt for a baby growing in a belly.

A prompt for a table of IVF medications looked passable at first, but then I noticed the needles were filled with black fluid. For comparison, here is what my actual medications looked like.

Even ChatGPT Gave Bizarre Advice

Looking for better prompts, I turned to ChatGPT, but its suggestions were even weirder and far too metaphorical to be useful. It suggested prompts like, "A garden of glass flowers representing embryos growing in a lab" and "Butterfly emerging from a test tube." Given the strict sterile protocols of an IVF lab, these ideas were simply ludicrous.

The Verdict: Can Sora Tell a Human Story?

While I didn't create a cinematic masterpiece, I did manage to get a few usable clips that could serve as B-roll in a larger video. The four clips below were the best of the bunch.

Would I use Sora again? Possibly, but I'll wait for a more advanced version. This experience revealed the serious blind spots of the technology. For a story as delicate and precise as my fertility journey, there can be no room for errors or inconsistencies. AI just isn't there yet.