Article written by Keshav Malhotra – Embryologist · Chair Embryology, ISAR · Chair, Embryology SIG, ASPIRE · Executive Member, Alpha Scientists in Reproductive Medicine
A colleague of mine, a senior embryologist with over two decades in the lab, sat across from me at a conference dinner last year and said something I have been thinking about since. “They are not replacing pipettes,” she said. “They are replacing us.” She was not being dramatic. She was being specific. She had watched an AI grading tool get introduced into a clinic without any conversation, any training, or any explanation of how its outputs would be used in clinical decisions. She was asked to sign off on embryos she had not assessed. She refused.
I want to be clear about something before we go further. She was right to refuse.
The conversation around AI in reproductive medicine has developed a bad habit. It tends to begin with capability and end with inevitability. The embryologist who raises a concern gets labeled as resistant to change, as someone who will be left behind. That framing is not only unfair; it is counterproductive. If you lead a lab, manage a clinic, or sit on a scientific committee deciding how these tools get adopted, the embryologist who pushes back is not your problem. She is your early warning system.
The Threat Is Real. Just Not the One They Are Describing.
When an embryologist says AI is a threat, they are almost never worried about the algorithm itself. They are worried about the social and institutional context in which it lands. Will this tool be used to justify reducing headcount? Will my clinical judgment be overridden by a confidence score? Will I be held responsible for an outcome I had no real say in? These are not paranoid questions. They are the right questions.
The data on AI-assisted embryo selection is genuinely interesting. Studies comparing AI grading tools to morphological assessment alone have shown improvements in consistency and, in some cohorts, in live birth rates. But here is what those papers do not always say plainly: these tools were developed in high-volume centres with large, well-annotated datasets. Their performance in your lab, on your patient population, with your time-lapse system, may look quite different. The embryologist who has been grading blastocysts for fifteen years in a mid-sized clinic in a tier-two city is not being unreasonable when she asks to see validation data specific to her context.
The challenge for leaders, then, is not to persuade reluctant embryologists that AI is harmless. The challenge is to create the conditions in which honest, rigorous evaluation can happen, and in which the people doing the clinical work have genuine authority over what gets adopted and how.
Why the Standard Pitch Does Not Work
Most AI adoption efforts in IVF labs follow a predictable sequence. A tool is presented at a department meeting. Slides show accuracy metrics. Someone senior endorses it. Embryologists are expected to get on board. Resistance is interpreted as a training problem.
This approach fails for three reasons that have nothing to do with the technology.
First, it bypasses the epistemic authority of the people who know the most about the problem. Embryologists who have spent years refining their morphological assessment have developed an intuitive understanding of their patient population, their lab conditions, and the edge cases that algorithms rarely see enough of. Asking them to defer to a tool, without first asking them what they know, is not a technology problem. It is a respect problem.
Second, it asks for trust before it offers transparency. Most commercial AI grading platforms are not fully open about their training data, their exclusion criteria, or the clinical contexts in which they underperform. An embryologist who asks “how does this model handle day-six blastocysts with fragmented trophectoderm?” and gets a vague answer about proprietary methodology is not being obstructionist. She is doing her job.
Third, it conflates adoption with endorsement. Embryologists understand that using a tool is not a neutral act. In a clinical context, using a tool implies some level of confidence in its outputs. Asking someone to use something they do not understand and cannot evaluate is asking them to compromise their professional integrity. That is a significant ask.
A Different Starting Point
The labs that have navigated this well share a common approach. They do not begin with the tool. They begin with the problem the tool is meant to solve, and they ask the embryologists to define that problem themselves.
What does that look like in practice?
- Name the specific gap first. Is the problem inter-observer variability in blastocyst grading? Inconsistency across overnight shifts? Difficulty prioritising embryos in large cohorts? When the problem is concrete, it becomes possible to evaluate whether a tool actually addresses it, rather than assuming it does because the vendor says so.
- Run a parallel assessment period. Before any AI tool influences a clinical decision, run it alongside your existing process. Compare outputs. Document disagreements. Ask your embryologists to keep notes on cases where the tool’s recommendation felt wrong to them and why. This data is valuable. It tells you something about both the tool and your lab’s context.
- Make the disagreement visible. The most dangerous AI implementation is one where the tool’s recommendation is accepted by default, because it takes effort to override it. Good implementation keeps disagreement easy and auditable. An embryologist who overrides a recommendation and documents her reasoning is practising good science. That should be encouraged, not treated as friction.
- Separate the tool from the decision. AI output is information. The clinical decision belongs to the embryologist and the clinician. This is not a semantic point. It has real consequences for how errors are reviewed, how liability is understood, and how professional authority is maintained in the lab.
- Give the sceptics a formal role. The embryologist who is most critical of a new tool is often the best person to lead the internal validation process. Her scepticism is an asset. Giving her formal authority over evaluation sends a message to the rest of the team: this organisation takes the concerns seriously, and the people raising them are respected for it.
What I Have Learned From the Sceptics
I have had the benefit of working with, and learning from, some of the most thoughtful scientists in this field. What I notice about the ones who approach AI with caution is that they are not opposed to the question the technology is trying to answer. They are opposed to a particular kind of certainty that has not been earned.
One senior colleague put it this way: “I do not object to using a model to assist with selection. I object to using a model whose failure modes I do not understand, in patients I am responsible for.” That is a scientifically sound position. It is also, if you look at it carefully, an invitation. She is telling you exactly what she needs to trust the tool: transparency about how it fails.
That is tractable. If the vendors building these tools are serious about clinical adoption, and not just commercial adoption, that is the conversation they should be having with embryologists, not the one about percentage improvements in clinical pregnancy rates from cherry-picked cohorts.
The Embryologist Is Not the Last Line of Defence Against a Bad Algorithm. But She Is a Line of Defence.
There is a version of the future where AI in IVF labs is adopted carelessly, where embryologists are deskilled, where clinical authority shifts from trained scientists to confidence scores on a screen, and where the first indication that something has gone wrong comes in an audit three years later. That future is not inevitable. But it does not get avoided by optimism or by insisting that the technology is safe. It gets avoided by the same thing that produces good science: rigour, scepticism, and the willingness to say “I need to see the data.”
The embryologist who thinks AI is a threat is not wrong to be worried. She is wrong only if she stops at the worry and does not take the next step, which is to engage with the technology on her own terms, to demand transparency, to document what she observes, and to bring her clinical expertise into the evaluation process rather than leaving it at the door.
And for those of us in leadership: our job is to make that next step possible. Not to persuade her that her concerns are unfounded, but to build the kind of institutional environment where those concerns become the engine of a better adoption process.
That is not the easiest path. But it is the one that produces outcomes you can stand behind.
Want to get in touch? Click here to contact us.
