AI embryo selection tools can support clinical judgment, but current evidence does not justify treating an algorithmic score as a clinical verdict.
A prediction is not a decision
Artificial intelligence is entering in vitro fertilisation at one of its most consequential points: deciding which embryo should be transferred first. The scale of the underlying need is substantial. The World Health Organization estimates that 17.5% of adults (roughly one in six people worldwide) experience infertility during their lifetime.[1][2] Against that background, tools that promise more consistent embryo assessment are understandably attractive to clinics and patients.
Computer-vision systems can analyse time-lapse images, detect developmental patterns and assign scores intended to estimate outcomes such as blastocyst formation, implantation or clinical pregnancy. They may reduce some of the variability associated with visual morphology assessment and process more images and variables than a person could review unaided. But a score is still a statistical estimate produced from chosen data, labels and endpoints. It does not determine what ought to be done.
That distinction matters because the algorithm is not ranking an abstract dataset. Its output may influence which embryo is transferred, cryopreserved or reconsidered. The decision engages reproductive autonomy, professional responsibility and the moral and legal significance attributed to embryos. Efficiency can inform the process; it cannot, by itself, justify the choice.
What the strongest clinical evidence actually shows
The case for caution is no longer based only on the absence of randomised evidence. In 2024, Nature Medicine published a multicentre, randomised, double-blind non-inferiority trial conducted in 14 IVF clinics in Australia and Europe. It enrolled 1,066 women under 42 who had at least two day-five early blastocysts. Clinical pregnancy occurred in 46.5% of patients whose embryo was selected with the iDAScore deep-learning system and in 48.2% of those assessed through standard morphology. With a risk difference of −1.7 percentage points and a 95% confidence interval from −7.7 to 4.3, the study did not demonstrate the algorithm’s non-inferiority within the pre-specified five-percentage-point margin.[3] This does not prove that AI is ineffective; it shows why promising retrospective performance cannot be treated as established clinical benefit.
The United Kingdom’s Human Fertilisation and Embryology Authority reaches a similarly restrained conclusion for time-lapse incubation with automated embryo analysis. Its current evidence rating is black: on balance, moderate- and high-quality evidence shows no improvement in the chance of having a baby for most fertility patients. The authority notes that time-lapse systems may still offer laboratory benefits, including continuous observation without removing embryos from the incubator, but that is different from proving better treatment outcomes.[4]
Endpoint selection is therefore critical. Blastocyst development, implantation, clinical pregnancy, live birth and the health of the child are not interchangeable. A model optimised for an intermediate outcome can appear more clinically decisive than the evidence supports. Reviews of AI embryo selection have warned that datasets, laboratory protocols, patient populations and validation methods vary markedly, limiting comparison and generalisability.[5][6]
Bias in IVF is a matter of safety and justice
An embryo-ranking model learns from the population and clinical practices represented in its training data. Age groups, ethnic populations, uncommon diagnoses, laboratory conditions and treatment protocols may be unevenly represented. A tool can perform acceptably on average while being less reliable for a subgroup. Before adoption, clinics should ask not only for headline accuracy but also for external validation, subgroup performance and evidence generated under conditions comparable to their own laboratory.
The commercial context matters as well. Proprietary tools may increase costs and concentrate access in well-resourced centres. Clinics should be able to determine what data were used, which outcome the model predicts, how the current version differs from earlier versions and how errors or model drift will be identified. Contractual confidentiality cannot be allowed to prevent a clinic from obtaining the evidence necessary to protect patients.
Selection criteria require separate ethical scrutiny. A system that estimates implantation potential is not equivalent to one that ranks embryos using genomic predictions of disease risk or complex traits. As the objective moves from estimating developmental potential toward preference-based selection, concerns about discrimination, diversity and eugenic reasoning become more acute. Technical feasibility does not settle ethical legitimacy.
Consent must include the algorithmic intervention
General consent to IVF should not be assumed to cover every clinically relevant use of AI. Patients need understandable information about the system’s purpose, the outcome it predicts, the evidence supporting it, its known limitations, the extent of human review and whether their health, genetic or embryo data may be used to train or refine a model.
Consent should be renewed when a material change occurs, for example, when a new model version is introduced, data are proposed for a secondary purpose or an algorithmic recommendation changes the treatment plan. Refusing AI support should not lead to abandonment of care. Nor should the system be presented as infallible or as the only rational option.
Explainability must be proportionate to the decision. A clinic may be unable to translate every mathematical operation into ordinary language, but it must still explain what the score means, what it does not mean, which outcome it addresses and how a qualified professional reviewed it. Technical opacity must not become ethical or legal opacity.
Responsibility cannot be delegated to software
Brazil’s Federal Council of Medicine has now established explicit professional rules. Resolution CFM No. 2,454/2026, published on 27 February 2026 and corrected on 5 March, entered into force on 26 August 2026.[7] It requires AI to remain a support tool; the physician retains final responsibility for clinical, diagnostic, therapeutic and prognostic decisions. The resolution also requires critical review of AI recommendations, recording the use of AI for medical decision support in the patient record, clear information to patients, protection of personal data and mandatory human supervision.[7]
For fertility services, this means that merely placing a physician or embryologist at the end of an automated workflow is insufficient. Meaningful oversight requires enough information, competence and institutional authority to challenge or reject the output. The record should identify the system and version used, the relevant result, the professional review and the final decision.
In the European Union, the AI Act creates a risk-based framework rather than a rule written specifically for IVF. An AI system that is itself a regulated product, or a safety component of one, may fall within the high-risk category under Article 6(1) when the product requires third-party conformity assessment. The resulting framework addresses risk management, data governance, technical documentation, logging, transparency, human oversight, accuracy, robustness and cybersecurity.[8] Application is phased, and the timetable has recently changed. Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on 27 July 2026 and defers the high risk obligations for stand alone Annex III systems to 2 December 2027, and for AI embedded in products already covered by European Union product safety legislation, including medical devices, to 2 August 2028.[9] Whether a particular embryo-ranking product meets that classification requires a case-specific regulatory analysis. The deferral affects when the obligations take effect rather than whether they apply.
A practical governance standard for fertility clinics
Before using an AI embryo-ranking selection system in patient care, a clinic should be able to answer seven questions:
What exact outcome does the model predict, and is that outcome clinically meaningful to the patient?
In which populations, clinics, incubators and laboratory protocols was it validated?
What were the prospective and external-validation results, including subgroup performance?
Can a physician or embryologist understand, contest and override the recommendation?
What information, alternatives and refusal rights are offered to patients?
How are model versions, decisions, incidents and performance drift recorded and audited?
Who is responsible for investigating an error and informing affected patients and regulators?
These questions should sit within a documented programme of multidisciplinary review, local validation, periodic audit, version control, incident response, cybersecurity and staff training. Procurement contracts should preserve access to performance evidence and establish duties when the vendor changes the model.
AI can improve IVF when it makes uncertainty more visible and supports better-informed professional judgment. It becomes dangerous when prediction is confused with certainty, ranking with worth or automation with neutrality. The defensible legal and bioethical position is neither rejection nor technological enthusiasm: AI should remain an accountable instrument within a human decision process that protects reproductive autonomy, acknowledges the significance of embryo selection and distributes the benefits and risks of innovation fairly.
Author Bio
