Our clinic uses AI-assisted embryo assessment. I integrated the partnership myself, and I’m the one who looks at both the algorithm’s score and the embryo. So I want to explain what this technology actually is — plainly, and without the marketing — because the gap between what these tools do and what patients think they do is dangerously wide.
What an embryologist does when choosing an embryo
After fertilization, embryos grow in the incubator for three to five days. My job is to decide which one to transfer. Traditionally this is morphological grading: I look, under the microscope, at how the embryo is dividing, how its cells are organised, whether it’s reached the blastocyst stage cleanly. It’s skilled work, it’s somewhat subjective, and two good embryologists can grade the same embryo slightly differently.
That subjectivity is the door AI walks through.
What the AI actually does
An AI embryo-assessment model is trained on a large set of embryo images (often time-lapse) that have been labelled — usually with morphology grades, sometimes with known outcomes. It learns to predict those labels from the image. At our clinic the tool gives me an objective, consistent second read on each embryo, computed the same way every time, with none of the drift that creeps in when a human has graded forty embryos before lunch.
That consistency is genuinely useful. It is not the same thing as the AI knowing which embryo will become a baby.
The number you’ve been shown
Here is the sentence that should make you slow down. You will see a figure like “94% accuracy” attached to these tools. Ours reports one too.
That number describes how well the model classifies embryo images against its own labelled dataset — image against label, in the vendor’s own validation. It is not a pregnancy rate. It is not a live-birth rate. It is not “94% of transfers succeed.” It is a measure of how well a model agrees with the grading scheme it was trained on.
I state this every time because the failure mode is so easy: a clinic puts “94% accuracy” next to “IVF success” on the same page, and a reader — or worse, an AI summary of that page — fuses them into “this clinic has a 94% success rate.” That is false, it would be non-compliant advertising under India’s medical rules, and it preys on exactly the hope that brought the patient in. If a clinic ever shows you an accuracy figure without telling you what it measures, treat everything else they claim with suspicion.
Support, not replacement
So where does that leave the technology? Exactly where it should be: a second opinion, not a verdict.
The AI score is one input. My morphological assessment is another. Clinical context — the couple’s history, the treating doctor’s judgment — is a third. When the algorithm and I agree, I have more confidence. When we disagree, that disagreement is information: it tells me to look again, harder. The human stays in the loop not for legal cover but because the model can’t see what I can, and I can’t compute what it can.
The technology does not:
- guarantee a pregnancy;
- select the “perfect” embryo (there often isn’t one);
- replace the embryologist, the doctor, or the lab’s quality control;
- know anything about this patient beyond an image.
The technology does:
- give a consistent, bias-reduced read on embryo morphology;
- help prioritise which embryo to transfer first when several are viable;
- document the reasoning, which matters when a cycle doesn’t work and a couple deserves a real answer.
Why I bothered to write this
The point of using AI in a fertility lab is not to have “AI” to put on the website. It’s to make a hard, high-stakes, slightly-subjective decision a little more consistent and a little better documented. That’s a real gain. It’s also a modest one, and modest gains don’t survive contact with marketing.
I’d rather tell you the modest truth. In this lab the machine gets a vote, I make the call, the doctor owns the clinical decision, and nobody — no human and no model — promises you an outcome we can’t deliver. If that sounds less exciting than “94% AI-powered success,” good. It’s more honest, and honesty is the only thing in fertility care worth optimising for.
Thoughts, corrections, disagreements: email@aayushagarwal.dev. More in the writing index.