Picking the 'best' embryo: Can we trust AI?

For 40 years, choosing which embryo to transfer during IVF has rested on a single trained eye looking down a microscope. Soon we will leave it to the machines, and perhaps that's a good thing.

Across fertility clinics in Australia, Brazil, Denmark, Sweden and the UK, artificial intelligence (AI) systems are now reviewing embryo images and footage alongside – or sometimes, instead of – embryologists. This adoption is happening quietly, with far less public debate than the technology deserves.

Embryo grading by looking under a microscope has always been an imperfect science. A systematic review of 20 studies comparing AI with embryologists found that AI models achieved a median accuracy of 75.5 percent when grading embryo morphology, compared with 65.4 percent for embryologists assessed on blinded test datasets. These results suggest that AI can grade embryos more accurately, but the studies used different datasets and methods, making direct comparisons difficult. Furthermore, grading an embryo is not the same as predicting whether it will lead to a successful pregnancy or live birth.

Part of the reason is data. The introduction of time-lapse incubators in the early 2010s meant that embryos could be filmed continuously rather than checked briefly once or twice a day, producing far more information about development than any embryologist could realistically absorb. AI was, in some sense, the obvious next step. It can make use of information that already existed, but had become too dense for human review. With a wealth of rich imaging data and a defined outcome to predict, we are in the computer's domain now.

What is harder to find in the coverage of this technology is perhaps its most interesting result. In a randomised trial across 14 clinics in four countries, an AI system called iDA selected the same embryo as the embryologist 65.8 percent of the time (see BioNews 1253). Clinical pregnancy rates were not significantly different between the two groups: 46.5 percent for AI-selected embryos compared with 48.2 percent for those selected by embryologists.

The trial did not, however, establish that AI was non-inferior to conventional selection in terms of clinical pregnancy rates. The AI was, however, dramatically faster, taking an average of 21.3 seconds to rank embryos compared with 208.3 seconds for an embryologist. This differs from the usual 'machines beat doctors' story we often read about – it is closer to parity in pregnancy outcomes, achieved at a fraction of the assessment time, and (presumably) the cost. For a treatment that is already expensive and exhausting for the people going through it, time and money may turn out to matter more than any accuracy gain.

Several of these systems are now commercially available and in routine use – iDA, Fairtility's Chloe and AIVF's EMA among them. This is not some pipe-dream software or just a proof-of-concept. These systems are increasingly being deployed in clinical settings around the world, with the potential to improve embryo selection for many patients. In fact, AI is already part of the decision-making process in a meaningful number of IVF cycles.

That raises a question that the efficiency statistics don't answer. As the bioethicist Dr Julian Koplin of Monash University, Melbourne, Australia, has argued, these systems mean that AI is beginning to make decisions about who is brought into the world. This is rather a stark way of putting it, but it is accurate. While many algorithms were created to assist with an embryologist's existing judgement, their speed and potential usefulness in the context of AI has now earned them a promotion. In many clinics, AI recommendations may carry considerable weight in the final decision.

One specific problem makes this harder to sit with than may initially be apparent. Many AI tools produce a score without clearly explaining why one embryo scored higher than another. A recent methodological review of the field identifies a lack of explainability, alongside limited external validation, as one of the field's central unresolved problems.

Patients are increasingly being told which embryo is predicted to offer the best chance, by a system that – if pressed – often cannot say why. Some systems offer tools to highlight features that influenced their predictions, but this is not necessarily the same as explaining why one embryo was judged better than another. Scientists and clinicians are expected to justify and explain their decisions. Is it ethically acceptable to sweep this under the rug for an AI tool, just because it performs well?

There is also a quieter equity question buried in here. A team in Brazil recently built and validated its own AI embryo selection model, MAIA (an acronym for Morphological Artificial Intelligence Assistance), rather than adopting an existing commercially available model. The researchers argued that imported embryo selection technologies may not adequately account for local demographic and ethnic profiles, and thus developed a model using embryo images from patients treated in São Paulo. That is a reasonable, responsible thing to do, and is also a reminder that the 'best embryo' is not simply a fixed biological fact waiting to be detected by AI. It is easy to forget that the output of AI is a pattern learned from whichever embryos, and whichever patients, happened to be included in the training data. Train the AI on a different population, and the definition of 'best' may shift with it.

None of this, to be clear, makes the technology a bad development. The honest case for AI embryo selection is a strong one: it may offer greater consistency than the traditional eye-and-experience method, it can work dramatically faster, and it has the potential to make IVF marginally less of a lottery for patients who are already under significant strain.

Few would seriously argue that we should go back to relying solely on a tired embryologist's subjective read of a screen at the end of a long shift, when AI can accomplish the same task in a fraction of the time and produced similar clinical pregnancy rates to conventional selections in a randomised trial, although non-inferiority was not established. All I suggest is that there is still work to do, kinks to iron out, and bigger questions to consider.

We should not be concerned simply because the decision is no longer made by a human alone. No one is calling for embryologists to lose their jobs. Rather, the deeper change is that 'best embryo' will quietly stop being a judgment made by one trained person in the room, and will become whatever pattern a model was optimised to recognise in whatever data it happened to be given. That is not necessarily worse. It may, on balance, be better for most patients, most of the time. But it is a different kind of decision to the one IVF patients may think they are consenting to, and clinics and regulators owe patients to them to say so plainly, rather than letting 'AI-assisted' stand in for a much bigger shift in who – or what – is actually doing the choosing.

---

The article above by Guy Davies won third prize in the Marcus Pembrey BioNews Writing Competition 2026. The competition is named in honour of Professor Marcus Pembrey (Patron of PET), and is generously supported by the London Women's Clinic.

The issues discussed in the article will be further explored at this year's PET Annual Conference, AI and Automation in Fertility and Genomics: Pipeline? Or Pipe Dream?, taking place in central London on Wednesday 9 December 2026. Find out more and register here.