Disclosure: the author is CEO of MyBabyBridge, an IVF company. AI embryo selection has not been shown to raise IVF success rates. In the first large randomized trial, published in Nature Medicine in 2024, a deep-learning ranker produced a 46.5% clinical pregnancy rate against 48.2% for embryologists, and failed to prove it was no worse. It was far faster. Faster is not the same as better.
Picture the moment the technology is built for. A patient has five blastocysts on day five. An embryologist can transfer one. The rest will be frozen, and the order in which they are thawed may decide whether this person is pregnant in March or in September, or whether they run out of money before the embryo that would have worked is ever used. Everyone in the room wants a better way to pick.
That wish is why embryo-ranking software has spread so quickly through fertility labs. The honest summary of the evidence is less comfortable. Retrospective studies make these tools look excellent. Randomized trials, the only design that can tell us whether patients actually do better, have so far told a humbler story.
The thesis of this piece is simple, and I think it holds up: embryo-selection AI is a genuine advance in consistency and speed, and so far an unproven one in outcomes. Those are different claims, and patients are routinely sold the second on the strength of the first.
How accurate is AI at picking the best embryo?
Most embryo AI is trained on time-lapse images or static photographs of embryos whose fate is known: did the transfer produce a heartbeat, a pregnancy, a birth? The model learns which visual patterns go with success, and its accuracy is reported as an AUC, the probability that it ranks a successful embryo above an unsuccessful one.
The development work behind one of the most widely used tools, Vitrolife’s iDAScore, shows why that number needs reading carefully. In a 2022 paper in PLoS One, Berntsen and colleagues trained the model on 115,832 embryos from 18 clinics. Across all embryos, including the many that embryologists discarded, the AUC was 0.95. Restricted to embryos that were actually transferred and had a known outcome, the AUC for predicting a fetal heartbeat fell to 0.67.
That gap is the whole story in miniature. An AUC of 0.95 largely reflects the easy task: telling a thriving blastocyst from one that has arrested or fragmented, which any trained embryologist can already do. The 0.67 describes the hard task, the one patients are paying for: choosing among several embryos that all look good. That is a real signal, better than a coin toss, but it is modest, and it is the number that matters.
What did the randomized trials find?
The most important such trial so far was led by Peter Illingworth and published in Nature Medicine in November 2024. It enrolled 1,066 women under 42 across 14 IVF clinics in Australia and Europe, each with at least two good blastocysts on day five. Half had their embryo chosen by iDAScore; half by embryologists using standard morphology grading and a predefined prioritization scheme. Neither patients nor clinicians assessing outcomes knew which arm was which.
The trial was designed to show the AI was not meaningfully worse, with a margin of five percentage points. It could not show that. Clinical pregnancy was 46.5% in the AI arm and 48.2% in the morphology arm, a difference of −1.7 points with a 95% confidence interval running from −7.7 to 4.3. Live birth, a secondary outcome, was 39.8% versus 43.5%. Neither difference was statistically significant, so the trial does not prove the AI was worse. It simply failed to prove it was as good, and the point estimates leaned the wrong way.
To be fair to the tool, the same trial found something real. The AI took about 21 seconds per evaluation against roughly 208 seconds for embryologists, an almost tenfold saving, and it gives the same answer to the same embryo every time, which human graders do not. A commentary in Human Reproduction drew the conclusion bluntly: blastocyst morphology remains the standard to beat.
There is a related body of evidence on time-lapse imaging, the camera-equipped incubators that feed many of these algorithms. The TILT trial, published in The Lancet in 2024, randomized 1,575 participants in the UK and Hong Kong. Live birth was 33.7% with time-lapse imaging and selection, 36.6% with undisturbed culture alone and 33.0% with standard care. The investigators concluded that time-lapse imaging did not improve the odds of live birth. The UK regulator, the Human Fertilisation and Embryology Authority, rates time-lapse imaging and incubation “black” in its add-ons scheme, meaning moderate or high quality evidence shows it has no effect on treatment outcome.
TILT tested time-lapse systems, not every AI model trained on their images. But it closes off the easiest argument, that watching embryos more closely must improve selection. Much of what decides whether a good-looking embryo implants, including chromosomal status and the uterus it enters, is not in the picture at all.
Why can’t a better ranking produce more babies?
In modern IVF, most clinics freeze surplus embryos and transfer them one at a time. In that world, ranking changes the order in which embryos are used. It does not change which embryos exist. If a patient has one embryo capable of becoming a baby among five, a perfect ranker finds it first and a poor one finds it third. The cumulative chance of a live birth from that egg collection is the same either way, provided the patient keeps going and nothing viable is thrown away.
A worked example (hypothetical). A 36-year-old has four frozen blastocysts. Suppose, unknown to anyone, only embryo C can produce a live birth. The embryologist ranks them A, C, B, D; an algorithm ranks them C, A, D, B. With the algorithm she has a baby after one transfer. With the embryologist, after two. Both routes end in the same place. The algorithm saved one transfer cycle, several weeks and a cost that matters a great deal to her. It did not add a baby.
That is a real benefit, and it should not be dismissed. Time to pregnancy is important, especially for older patients, and every failed transfer carries emotional and financial cost. Many patients stop treatment before they use all their embryos, so faster success can translate into more births in practice.
But notice what this means for evidence. The claims to test are time to pregnancy and cumulative live birth per egg collection, measured in randomized trials. Per-transfer pregnancy rates from retrospective data cannot answer either. And there is one way ranking can reduce cumulative success: if a low AI score persuades a clinic or patient to discard an embryo that would have become a baby. Any tool used to decide what is thrown away, rather than what goes first, carries a burden of proof it has not yet met.
What is the strongest case for AI embryo selection?
The case for these tools is serious, and anyone writing about them from inside the fertility industry owes it a fair statement.
First, one trial is one trial. Illingworth’s study compared the AI against embryologists at established clinics working to a predefined scheme, which is a demanding comparator. In a lab with less experienced staff, higher turnover or heavy workload, the gap between a consistent algorithm and a tired human could be larger. The trial did not test that setting.
Second, consistency is a clinical good in its own right. Morphology grading varies between embryologists and within the same embryologist on different days. A tool that removes that variation makes lab quality easier to audit and outcomes easier to compare across sites.
Third, the time saving is not trivial. Freeing ten-plus minutes of skilled embryologist time per case adds up in a busy lab and could be redirected to the steps where human hands still decide outcomes.
All three points are fair, and newer models that combine images with clinical data deserve their own trials. None of them is evidence that patients get pregnant more often. The appropriate response is not to dismiss the technology but to describe it accurately: a standardization and efficiency tool with unproven outcome benefit, pending better trials. For more on why diagnostic AI tends to look strongest on narrow, well-defined tasks, see our earlier piece on where the evidence for AI diagnostics is strongest.
Should patients pay extra for AI embryo selection?
This is where the evidence should shape practice most directly.
When a technology saves the clinic time and gives no proven gain to the patient, the natural economic arrangement is for the clinic to absorb it as part of lab operations, the way it absorbs a better incubator. Charging patients a separate fee turns it into an add-on, and IVF add-ons have a poor record. The HFEA’s rating system exists precisely because patients were being sold extras, from endometrial scratching to immune treatments, on the strength of plausible biology and hopeful retrospective data.
A reasonable test for any patient-facing claim is whether it names the outcome and the study design. “Our AI identifies the embryo most likely to implant” invites the retrospective-AUC confusion described above. “In a randomized trial, AI selection did not improve pregnancy rates compared with embryologists, but it made our grading faster and more consistent” is accurate and, frankly, still a reasonable thing for a lab to want.
Real-world data will also be tempting to cite. Clinics that adopt AI often report better results afterwards, but they are usually also upgrading incubators, protocols and staff at the same time, so a before-and-after comparison is weak evidence for a marginal effect in a field with this many moving parts. Clinics buying these systems should ask the questions we set out in buying digital health: the questions hospitals ask too late, starting with what outcome the vendor has actually shown in a controlled comparison.
What this means for clinics, patients and developers
For clinics, the honest position is available and defensible: adopt AI selection if it improves consistency, workflow and audit, report it to patients as a lab tool rather than a success booster, and do not charge for it as though it were proven to raise live birth rates. Keep embryologists in the loop on every transfer decision, and never use a score alone to discard an embryo.
For patients, the useful questions are concrete. Has this tool been tested in a randomized trial? What was measured: pregnancy per transfer, live birth, or time to pregnancy? Will it change which embryos are kept or only which is used first? Is there an extra charge? A clinician who answers those plainly is a good sign whatever the answer.
For developers and investors, the next credible step is randomized trials powered for cumulative live birth or time to pregnancy, in the settings where the tool is expected to help most, with results published whatever they show. Regulators and professional societies can help by treating embryo-ranking claims as the outcome claims they implicitly are.
Frequently asked questions
Does AI embryo selection improve IVF success rates?
Not on current randomized evidence. The largest published trial, in Nature Medicine in 2024, found clinical pregnancy of 46.5% with deep-learning selection versus 48.2% with embryologist grading, and could not show the AI was non-inferior. Live birth rates were 39.8% and 43.5%. The AI was much faster and more consistent, but a higher success rate has not been demonstrated.
Why do AI embryo tools report such high accuracy?
Headline accuracy often includes embryos that embryologists would never transfer, which are easy to rank. In one large development study, the AUC was 0.95 across all embryos but 0.67 among transferred embryos with known outcomes. The second number reflects the real clinical choice between good-looking embryos, and it shows a modest signal rather than a decisive one.
Is time-lapse imaging worth paying for in IVF?
The TILT trial in The Lancet (2024) found no improvement in live birth with time-lapse imaging, and the UK’s HFEA rates time-lapse imaging and incubation black, meaning good evidence of no effect on outcomes. Some clinics include it as standard lab equipment. Patients should ask whether it carries an extra charge and what outcome evidence supports it.
Can AI embryo selection ever increase the number of babies from one IVF cycle?
Only indirectly. With single transfers and frozen embryos, better ranking changes the order in which embryos are used, not which ones are viable. It may shorten time to pregnancy, which could help patients who would otherwise stop treatment early. It could reduce births if low scores lead to discarding viable embryos, so scores should not decide what is thrown away.
What should I ask my fertility clinic about AI?
Ask whether the tool has been tested in a randomized trial, which outcome was measured, whether it affects which embryos are kept or only the transfer order, and whether there is an extra fee. Discuss the answers with your fertility specialist, who can relate them to your own situation and embryo numbers.
The fertility field has been here before: a clever technology, persuasive retrospective data, confident marketing, and randomized trials that arrive years later with a smaller effect than promised. Embryo AI can still earn its place. It will do so by publishing trials on cumulative live birth and time to pregnancy, not by quoting an AUC that mostly measures what embryologists already knew.
If you work in an IVF lab or have been through treatment, how was AI embryo selection explained to you, and did the explanation match the evidence?
