Artificial intelligence (AI) is increasingly mentioned in IVF — in laboratory brochures, news stories and online discussions — often with the suggestion that a computer can "pick the best embryo". For patients, the important questions are practical: what does the software actually do, and does it increase the chance of having a baby? This article explains how AI embryo assessment works, what the best available studies show, and why a prediction score is not the same as a guarantee.
Key takeaway: AI tools can rank embryos quickly and consistently, and in research datasets they often predict pregnancy at least as well as embryologists. However, randomised trials have not shown that AI-based or time-lapse-based embryo selection increases pregnancy or live-birth rates compared with standard assessment by embryologists. AI is best understood as decision support within a well-run laboratory, not a replacement for professional judgement — and no algorithm can guarantee implantation or a live birth.
How Are Embryos Chosen Today?
When a cycle produces more than one embryo, the laboratory must decide which to transfer first. Traditionally, embryologists examine embryos under a microscope at set times and grade their appearance (morphology). At the blastocyst stage (day 5–6), this usually means scoring how expanded the embryo is and the quality of two cell groups: the inner cell mass, which forms the baby, and the trophectoderm, which forms the placenta. Grading is useful but partly subjective, and two experienced embryologists may not always agree.
Time-lapse incubators add a camera that photographs each embryo every few minutes without removing it from the incubator. This produces a video of development and allows the timing of cell divisions (morphokinetics) to be measured.
What Does "AI-Assisted Embryo Assessment" Actually Mean?
AI in embryo selection usually refers to software that analyses embryo images — single photographs or time-lapse sequences — and produces a score.
- Machine learning means a computer program learns patterns from large sets of past embryos whose outcomes are known.
- Deep learning is a type of machine learning that uses layered "neural networks" to analyse images directly, identifying features that humans have not defined in advance.
- The output is a score or ranking — an estimate of the likelihood of a particular outcome — not a diagnosis.
An important detail is which outcome the program was trained to predict. Some tools predict blastocyst formation, some implantation or a fetal heartbeat, and relatively few live birth. A model that predicts an early outcome well may not predict live birth equally well.
What AI Can Do
Current evidence supports some genuine, if modest, strengths. These are listed below.
- Consistency: the same images produce the same score within a given model, reducing variation between observers.
- Speed: in a large randomised trial, the deep-learning system assessed embryos in about 21 seconds compared with about 208 seconds for manual assessment (Illingworth et al., 2024).
- Structured ranking: it can help order embryos that look similar to the human eye.
- Research value: large image datasets may improve our understanding of embryo development.
The Core Limitation: AI Can Only Choose Among the Embryos You Have
Embryo selection — whether by an embryologist, a time-lapse algorithm or AI — changes the order in which embryos are transferred. It cannot improve an embryo's quality or increase the number of embryos. If all embryos from a cycle are eventually transferred, the total chance of a baby from that cycle is determined by the embryos themselves, not by the order in which they were used.
The realistic benefit of better selection is therefore reaching a pregnancy sooner, with fewer transfers. That is valuable — it can save time, cost and emotional strain — but it still needs to be demonstrated in well-designed trials, not assumed.
Prediction Accuracy Is Not the Same as Clinical Benefit
Many studies report that AI "outperforms" embryologists. A systematic review of 20 studies found that AI predicted clinical pregnancy with a median accuracy of 77.8%, compared with 64% for embryologists, and 81.5% versus 51% when images were combined with clinical information (Salih et al., 2023). However, the authors noted that the studies lacked prospective clinical validation, mostly used local datasets and rarely focused on live birth.
There is also a structural problem with retrospective studies. Only embryos that embryologists chose were transferred, so the true potential of the embryos left behind is unknown — a form of selection bias. A methodological review also found that very few AI studies checked whether predicted probabilities matched real pregnancy rates (calibration), and that differences in embryo populations made comparisons between models difficult (Kragh and Karstoft, 2021).
In simple terms, a program can be good at predicting outcomes in historical data without changing outcomes when used in a clinic. Only randomised trials can answer that question.
What Do Randomised Trials Show?
| Study | Participants | Comparison | Main result |
|---|---|---|---|
| Illingworth et al., 2024 (Nature Medicine) | 1,066 patients; 14 clinics in Australia and Europe | Deep-learning embryo selection (iDAScore) versus standard morphology | Clinical pregnancy 46.5% vs 48.2%; non-inferiority not demonstrated. Live birth 39.8% vs 43.5% |
| Kieslinger et al., 2023 (The Lancet, SelecTIMO) | 1,731 couples; 15 clinics in the Netherlands | Time-lapse algorithm selection (EEVA) versus routine selection in a time-lapse incubator versus standard culture | Cumulative ongoing pregnancy 50.8% vs 50.9% vs 49.4%; no improvement |
| Bhide et al., 2024 (The Lancet, TILT) | 1,575 participants; 7 centres in the UK and Hong Kong | Time-lapse imaging and selection versus undisturbed culture versus standard care | Live birth 33.7% vs 36.6% vs 33.0%; no significant improvement |
The Illingworth trial is the most directly relevant, because it randomised patients to AI-based or embryologist-based selection. AI did not meet the pre-specified standard for being "no worse" than standard assessment (a margin of 5 percentage points). This does not prove that AI is harmful — the difference was not statistically significant — but it means that equivalence, let alone superiority, has not been demonstrated. The two large time-lapse trials likewise found no gain in pregnancy or live birth from algorithm-based selection. ESHRE's 2023 good practice recommendations on IVF add-ons concluded that available data on time-lapse technology do not show an improvement over conventional assessment.
Why Might a Good Algorithm Fail to Improve Live Birth?
Several reasons are plausible. These are listed below.
- Implantation also depends on the endometrium, transfer technique, maternal age and chance, none of which an embryo image captures.
- Experienced embryologists are already good at excluding poor embryos, so the remaining differences between top-ranked embryos may be small.
- Chromosomal abnormalities — a major cause of implantation failure — are not reliably visible in images.
- A model developed in one laboratory may not perform as well in another.
Validation, Generalisability and Bias
AI models learn from the data they are given: the clinics, patients, incubators, culture media and camera systems involved. When conditions differ, performance may fall. A 2026 study trained 50 versions of the same model on identical data, changing only a random starting value. The models agreed poorly on how to rank the same embryos, in roughly 15% of cases ranked a lower-quality embryo above a viable one, and made more errors when tested on data from a different centre (Thirumalaraju et al., 2026).
Bias is a related concern. If a model is trained mostly on one population, it may be less reliable for patients whose age profile, causes of infertility or treatment protocols differ. ESHRE's good practice recommendations on time-lapse technology advise that each laboratory should validate selection models in its own setting rather than simply adopting published algorithms (Apter et al., 2020).
Transparency and Explainability
Many deep-learning systems are "black boxes": they give a score without explaining why. Authors writing in Human Reproduction Open have argued that interpretable models should be preferred for embryo selection, because opaque systems make it harder to detect errors, assign responsibility and involve patients in shared decisions (Afnan et al., 2021). Patients are entitled to know whether AI was used in their treatment and how much weight its score carried.
The Embryologist's Role Remains Central
At present, AI is best viewed as clinical decision support. The embryologist reviews the images, checks that the score makes biological sense, and integrates information the algorithm does not have — such as genetic testing results, the number of embryos available and the couple's plans for future transfers. Human oversight is what catches the occasional implausible ranking. Overall laboratory quality — stable culture conditions, trained staff and quality control — matters at least as much as any single piece of software.
Regulation and the Indian Context
Regulatory requirements for AI software differ between countries, and approval generally concerns safety and performance rather than proof of a higher live-birth rate. In India, the Indian Council of Medical Research published ethical guidelines in 2023 for the application of AI in biomedical research and healthcare, providing a framework for ethical decision-making across development, deployment and adoption, including ethics review, governance and informed consent (ICMR, 2023). The major randomised trials discussed above were conducted in Australia, Europe, the United Kingdom and Hong Kong, so it is reasonable to ask how a tool has been validated in the laboratory where it is being used.
What AI Cannot Tell You
The limits of current AI embryo assessment are listed below.
- It cannot guarantee that an embryo will implant or result in a live birth.
- It cannot reliably confirm that an embryo is chromosomally normal.
- It cannot assess the uterus, the transfer or other non-embryo factors.
- It cannot increase the total number of babies a cycle can produce — it can only influence the order of transfers.
What Should You Ask Your Fertility Specialist?
The useful questions to ask are listed below.
- Is AI or a time-lapse algorithm used to select my embryos, and how much does it influence the final choice?
- What outcome was the software trained to predict — implantation, fetal heartbeat or live birth?
- Has it been validated in this laboratory, and with patients like me?
- What do randomised trials show for this specific tool?
- Is there an additional charge, and what would happen if I chose standard assessment instead?
Common Misconceptions About AI in Embryo Selection
"AI can find the perfect embryo." It ranks the embryos available; it cannot create a better one.
"A high AI score means the embryo will implant." A score is a probability estimate, not a guarantee.
"AI is more accurate, so it must improve success rates." Better prediction in historical data has not yet translated into higher live-birth rates in randomised trials.
"Using AI means the embryologist is no longer needed." Expert review remains essential for safety and interpretation.
Frequently Asked Questions
Does AI embryo selection improve IVF success rates?
Not on current evidence. The largest randomised trial of deep-learning embryo selection did not show that it was as good as or better than standard embryologist assessment for clinical pregnancy, and large time-lapse trials found no improvement in live birth.
Is AI better than an embryologist at choosing embryos?
In retrospective studies, AI often predicts outcomes as well as or better than embryologists. In the randomised trial that tested this directly, AI did not demonstrate non-inferiority.
Is AI the same as time-lapse imaging?
No. Time-lapse imaging is a camera-equipped incubator that records embryo development. AI is software that may analyse those images, or single photographs, to produce a score.
Can AI tell whether an embryo is genetically normal?
Not reliably. Preimplantation genetic testing (PGT) analyses cells from the embryo; image-based AI cannot replace it.
Is AI embryo selection safe?
Image analysis itself does not touch the embryo. The main risk is an incorrect ranking, which is why embryologist oversight and local validation matter.
Should I pay extra for AI embryo selection?
That is a personal decision, but it should be based on evidence. At present, it should be regarded as an emerging technology with unproven effect on live birth.
Medical Disclaimer
This article is for educational purposes only and does not constitute personalised medical advice. Fertility assessment and treatment depend on individual circumstances, including age, medical history and test results. Please consult a qualified fertility specialist before making any decisions about diagnosis or treatment.
References
- Illingworth PJ, Venetis C, Gardner DK, et al. Deep learning versus manual morphology-based embryo selection in IVF: a randomized, double-blind noninferiority trial. Nature Medicine. 2024;30(11):3114–3120. doi:10.1038/s41591-024-03166-5. PMID: 39122964.
- Kieslinger DC, Vergouw CG, Ramos L, et al. Clinical outcomes of uninterrupted embryo culture with or without time-lapse-based embryo selection versus interrupted standard culture (SelecTIMO): a three-armed, multicentre, double-blind, randomised controlled trial. The Lancet. 2023;401(10386):1438–1446. doi:10.1016/S0140-6736(23)00168-X.
- Bhide P, et al. Clinical effectiveness and safety of time-lapse imaging systems for embryo incubation and selection in in-vitro fertilisation treatment (TILT): a multicentre, three-parallel-group, double-blind, randomised controlled trial. The Lancet. 2024;404(10449):256–265. doi:10.1016/S0140-6736(24)00816-X.
- Salih M, Austin C, Warty RR, et al. Embryo selection through artificial intelligence versus embryologists: a systematic review. Human Reproduction Open. 2023;2023(3):hoad031. doi:10.1093/hropen/hoad031. PMID: 37588797.
- Kragh MF, Karstoft H. Embryo selection with artificial intelligence: how to evaluate and compare methods? Journal of Assisted Reproduction and Genetics. 2021;38(7):1675–1689. doi:10.1007/s10815-021-02254-6.
- Thirumalaraju P, Kanakasabapathy MK, Kandula H, et al. Stability and reliability of artificial intelligence models in embryo selection for in vitro fertilization. Fertility and Sterility. 2026;125(2):277–286. doi:10.1016/j.fertnstert.2025.08.021. PMID: 40876725.
- Afnan MAM, Liu Y, Conitzer V, et al. Interpretable, not black-box, artificial intelligence should be used for embryo selection. Human Reproduction Open. 2021;2021(4):hoab040. doi:10.1093/hropen/hoab040.
- ESHRE Working Group on Time-lapse Technology, Apter S, Ebner T, Freour T, et al. Good practice recommendations for the use of time-lapse technology. Human Reproduction Open. 2020;2020(2):hoaa008. doi:10.1093/hropen/hoaa008.
- ESHRE Add-on