A proposal to meet specific criteria before fully integrating Artificial Intelligence (AI) into the IVF lab

Abstract

Artificial intelligence (AI) is generating genuine excitement in reproductive medicine for good reason. The promise of more consistent embryo assessment, reduced inter-observer variability, better outcome prediction and meaningful support for the embryologist’s judgment represents a real opportunity to improve care for patients who have already invested enormously in the pursuit of a family. Realizing that promise, however, depends on something the field has not always done well which is insisting on rigorous validation before widespread adoption. This editorial is written in that spirit, as advocates for AI who believe the technology’s long-term acceptance by patients, clinicians and embryologists will be determined by how carefully we deploy it now and how we proceed going forward. How will we know when AI is ready to fully deploy in the IVF lab? The answer requires distinguishing between being safe and being safe and effective and the medical community has not yet reached consensus on either standard for most applications. Professional societies will need to lead that process. However, in the meantime, with commercial adoption accelerating and stakeholder incentives not always aligned with patient welfare, we propose criteria as a starting framework for that discussion.

Artificial intelligence (AI) is generating genuine excitement in reproductive medicine, and for good reason. The promise of more consistent embryo assessment, reduced inter-observer variability, better outcome prediction and meaningful support for the embryologist’s judgment represents a real opportunity to improve care for patients who have already invested enormously in the pursuit of a family. Realizing that promise, however, depends on something the field has not always done well which is insisting on rigorous validation before widespread adoption. This editorial is written in that spirit, as advocates for AI who believe the technology’s long-term acceptance by patients, clinicians and embryologists will be determined by how carefully we deploy it now and how we proceed going forward.

Caution is warranted. It is not because AI poses some existential threat [1] but because medicine has a long history of adopting technologies before adequately understanding them. The irreversible consequences in this field where each decision can determine reproductive outcomes are not recoverable by a patch or software rollback. As E.O. Wilson observed, “the problem with humanity is that we have Paleolithic brains, medieval institutions and 21st century God-like technology.” [2] The gap between the speed at which AI tools are being commercialized and the pace at which the field is developing the standards to evaluate them is exactly the kind of institutional lag Wilson was describing. We believe closing that gap is the responsibility of the clinicians and embryologists who practice reproductive medicine, not the companies selling the tools.

AI applications in IVF span a wide range of tasks. They include image analysis for embryo and oocyte assessment, decision support for fertilization checks and embryo grading, time-lapse morphokinetic annotation, sperm selection, prediction of aneuploidy, outcome modeling for live birth probability, and administrative functions such as patient matching and scheduling [3]. What most of these tasks share is that they operate on structured biological data, images, measurements and clinical records, drawn from a relatively contained field. The entire United States generates roughly 300,000 IVF cycles per year, setting the data ceiling for the whole field and most commercial models train on a fraction of that volume. These are not the large-scale frontier systems that concern AI safety researchers working on general intelligence. The case for caution in our field stands on its own clinical evidence, which is substantial, and that is where the argument belongs.

The potential benefits of AI in reproductive medicine are specific and meaningful. More consistent embryo and oocyte assessment, reduced inter- and intra-observer variability, standardized grading across clinics, lower embryologist workload, better-informed transfer decisions and richer datasets to build prognostic models are all realistic near-term contributions of well-validated AI tools. These are achievable goals with genuine patient impact. Realizing them requires resisting the pressure to deploy tools before they have been adequately tested for safety and effectiveness in the populations and settings where they will actually be used.

Practitioners of IVF have demonstrated a track record of early adoption of new and novel technologies used in IVF with the intent of improving outcomes. Critics might say sometimes it has been appropriate but, on several occasions, it has been added before safety and/or effectiveness was demonstrated and properly evaluated. Thus, it is believed by some that we may be more likely to prematurely adopt the full integration of AI in the embryology lab before establishing specific standards prior to its implementation. The pressures are real. They include growing case volume, insurance expansion, corporatization by private equity, increasing competition between clinics for market share and a growing cohort of patients without classical infertility seeking treatment. Each of these creates incentive to move quickly. None of them is a substitute for validation and the costs of a premature false start in this field falls squarely on our patients.

Governmental regulation of AI remains minimal [4], despite calls from developers and researchers for meaningful guidance and guardrails. The competitive race to build and own AI products before another person, company, or country is a significant driver of that regulatory gap. In its absence, the responsibility for protecting patients falls on the practitioners who use these tools. Self-discipline is not a comfortable substitute for regulation, but for now it is the mechanism available to us [5].

Not all AI applications in IVF carry the same clinical weight and the field will benefit from being precise about this distinction. Beyond embryo grading, AI is being applied across each step of the IVF laboratory workflow. Embryo selection for transfer is one of the most consequential tasks. Ranking embryos within a cohort to identify the best candidate is also more time-consuming than any single assessment and it is precisely where the evidence raises serious concerns. Published data demonstrates weak correlation rank agreement across models, with a poor quality embryo ranked highest for proposed transfer in roughly one in seven cases even when high quality blastocysts were available [6]. Models with similar aggregate accuracy can produce entirely different rank orders while drawing on different image features. Aggregate discrimination metrics measure population-level performance and cannot detect ranking errors at the individual patient level [6]. Several important applications remain in development and are not ready for routine clinical use, including sperm selection for ICSI, oocyte quality assessment, robotic/automated ICSI and micromanipulation, non-invasive aneuploidy prediction, automated cryopreservation, comprehensive outcome prediction of live births and embryo witnessing. Challenges to overcome going forward are that most AI tools in IVF have been developed and validated retrospectively while large-scale prospective randomized controlled trials are lacking [5–7]. Data-sharing barriers between clinics, lack of standardized imaging protocols, limited demographic diversity in training and validating datasets and regulatory uncertainty remain significant barriers to wide clinical adoption [5–7].

Not all AI applications in IVF carry the same clinical weight and the field will benefit from being precise about this distinction. Beyond embryo grading, AI is being applied across each step of the IVF laboratory workflow. Embryo selection for transfer is one of the most consequential tasks. Ranking embryos within a cohort to identify the best candidate is also more time-consuming than any single assessment and it is precisely where the evidence raises serious concerns. Published data demonstrates weak correlation rank agreement across models, with a poor quality embryo ranked highest for proposed transfer in roughly one in seven cases even when high quality blastocysts were available [6]. Models with similar aggregate accuracy can produce entirely different rank orders while drawing on different image features. Aggregate discrimination metrics measure population-level performance and cannot detect ranking errors at the individual patient level [6]. Several important applications remain in development and are not ready for routine clinical use, including sperm selection for ICSI, oocyte quality assessment, robotic/automated ICSI and micromanipulation, non-invasive aneuploidy prediction, automated cryopreservation, comprehensive outcome prediction of live births and embryo witnessing. Challenges to overcome going forward are that most AI tools in IVF have been developed and validated retrospectively while large-scale prospective randomized controlled trials are lacking [5–7]. Data-sharing barriers between clinics, lack of standardized imaging protocols, limited demographic diversity in training and validating datasets and regulatory uncertainty remain significant barriers to wide clinical adoption [5–7].

How will we know when AI is ready to fully deploy in the IVF lab? The answer requires distinguishing between being safe and being safe and effective and the medical community has not yet reached consensus on either standard for most applications. Professional societies will need to lead that process. In the meantime, with commercial adoption accelerating and stakeholder incentives not always aligned with patient welfare, we propose the following criteria as a starting framework for that discussion as summarized in Table 1.

Table 1.

Nine criteria for evaluating Al tools before integration into the IVF lab

1. The algorithm has been trained and validated on demographically diverse populations, including patients with a range of infertility diagnoses and non-infertile patients and its performance has been tested at clinics that had no role in its development. Currently many AI models are trained on homogenous datasets from large volume, high resource, economically advantaged clinics. They may perform well on training data but have degraded performance in different clinics, patient populations and clinical conditions. There is a critical lack of evidence on the ability to replicate performance of these technologies under different field conditions with greater diversity in patient populations. This can result in overfitting of AI models creating demographic bias and inaccurate results. Published data on model stability show that critical error rates do not scale predictably across centers. Models with low error rates at one institution did not reliably produce low rates at another, with error variance increasing by 46% when the same models were evaluated on data from a different fertility center using a different embryoscope generation [5]. External validation, meaning testing at a clinic that had no role in training, is not optional. It is the floor, not the ceiling, of what should be required before clinical deployment. Clinics adopting these tools should demand site-specific performance data, not only the developer’s aggregate validation statistics. The obligation to monitor and re-validate does not end at purchase; it belongs to the adopting program on an ongoing basis.

2. Demonstration that implemented AI is free of systematic misclassification, data leakage and fragility to instill confidence in its users. The term “hallucination” applies specifically to generative language models; image classifiers of the kind predominantly used in IVF-AI do not hallucinate. Their failure modes are different and in many ways harder to catch. They misrank embryos, produce spurious correlations, and generate confident probability scores that may not reliably identify the correct top-choice embryo. A model can post impressive aggregate accuracy while placing a compromised embryo at the top of the ranking in a meaningful proportion of cases. Validation must therefore assess decision consistency and rank accuracy at the patient level, not just summary classification metrics of a population.

3. Programs understand what the model was trained on, what it was optimized to predict, and under what conditions its performance degrades, before clinical deployment. Conclusions of “black box” decisions must be as transparent and traceable as the technology allows, so that practitioners can interrogate and troubleshoot AI guidance.

   Full post-hoc explainability of deep learning, meaning the ability to say exactly which image features drove a given recommendation, is not currently achievable in any practically meaningful sense. This does not mean we should accept opacity, but we do need to define an honest and reachable standard. Before adoption, a program should understand what population the model was trained on, what outcome it was optimized to predict, under what conditions its performance degrades and how frequently it makes errors of the most consequential kind. If a vendor cannot answer these questions with specificity, that is itself a “red flag”. Tools such as gradient-weighted activation mapping can provide partial insight, but clinicians and embryologists should not depend on these to catch problems. Population-level performance data and ongoing re-evaluation against local cases are what actually protect patients.

   Opacity is not acceptable as a default, but traceability in deep learning means something different than it does in rule-based systems. A program cannot trace why a model ranked one embryo above another the way it can follow a calculation or decision tree. What it can require is that vendors disclose the conditions under which the model has been shown to fail, that those disclosures be specific rather than generic and that the program maintains the capacity to detect degrading performance over time. The distinction worth holding is between explainability, which is largely unavailable in current deep learning systems, and accountability, which is fully demandable. Insisting on the latter is how programs build the understanding and trust that black box conclusions alone cannot currently provide.

4. Clinical outcomes are equal or better and are validated for the specific patient populations they are applied to. Most AI studies in IVF are retrospective, use surrogate endpoints and lack prospective randomized controlled trials demonstrating improved clinical outcomes. If past is prologue, then AI may likely follow a pattern of premature commercialization and adoption as several “add-ons” to IVF have been made prior to proper validation. Examples of such premature “add-ons” have included intrauterine hCG infusion, endometrial scratching, endometrial receptivity testing, intralipid infusions, platelet-rich plasma infusions, indiscriminate routine use of PGT-A to all IVF cycles. Rigorous non-inferiority, safety studies are encouraged before AI is integrated and determined to be of real value. The most rigorous prospective trial to date, a randomized double-blind noninferiority study across 14 centers comparing AI-assisted embryo selection to standard morphological assessment, was unable to demonstrate noninferiority for clinical pregnancy rates, though it did show a roughly tenfold reduction in assessment time [8]. The workflow benefit is real, but clinical equivalence sufficient to justify replacing embryologist judgment with algorithmic ranking was not established. Beyond initial outcome trials, validation must also account for model updates. These tools are periodically retrained, and a version validated one year may not be the version running the next. Programs should require developers to provide updated performance data before any new version is clinically deployed.

5. Effective cybersecurity is implemented to maintain privacy and prevent data breaches of patients’ DNA profiles, parental identities and any patient sensitive information. IVF labs become the primary targets for ransomware as they contain “reproductive big data” with the most sensitive data of DNA profiles and parental identities. Connecting lab hardware to the cloud for AI analysis creates concentrated opportunity targets for hackers. This would be highly vulnerable to breaches of patient confidentiality. There is concern that such technologies are being implemented in clinical settings faster than vetting them for safety or regulation that would provide guardrails. In practice, the higher-probability threat may be more immediate than sophisticated ransomware: clinicians and staff who paste patient-identifiable data into consumer AI tools during routine case review. This is already happening and represents a direct HIPAA exposure requiring no external attacker. Programs adopting AI tools of any kind should establish and enforce clear policies on what data may be transmitted to external services, train all staff accordingly, and audit compliance. This is the same obligation that applies to any other electronic health record workflow.

6. Maintenance of “human-in-the loop” as part of the implemented AI to supervise and retain appropriate ethics, empathy and critical clinical judgement as part of clinical decisions. Without a human-in-the-loop the risks are too great for error and poor outcomes. The lab is not a factory. It requires human empathy, ethics and manual skill sets to trouble shoot when AI systems processes tilt. The human-in-the-loop principle must mean genuine clinical judgment, not the appearance of oversight. This applies across the full range of AI-assisted decisions in the IVF lab: fertilization assessment at the pronuclear stage, recommendations regarding whether to biopsy or freeze an embryo, ranking for transfer, and decisions about which embryos to discard. Each of these carries direct consequences for the patient, and in each case the AI output should serve as a prompt for clinical review rather than a directive. A clinician or embryologist who consistently defers to the algorithm without independent evaluation is not exercising meaningful oversight; they may be incurring liability for a decision the algorithm made. Many commercial AI tools are deployed with terms that explicitly place responsibility for clinical outcomes on the user. When market pressure toward adoption is strong and the practitioners’ understanding of the tool’s limitations is thin, the human in the loop becomes a legal buffer rather than a safeguard, particularly during this early period when the evidence base for many of these applications is still being established. The real test is whether the clinician or embryologist can recognize when the AI recommendation is likely wrong: which patient characteristics, imaging conditions, or cohort compositions degrade the model’s reliability. That requires training that goes well beyond reviewing vendor materials.

7. Define what entities are legally accountable for errors and accidents that may occur. Responsibility can become diffuse and fragmented if accountability is ambiguous. When rare errors occur the clinic may blame the software developer while the developer blames the hardware manufacturer for mechanical failure and the manufacturer points back to user error or unpredictable software. This leaves patients and legal systems with ambiguity of what entity to hold accountable. The principle that should resolve this ambiguity is the same one governing all other clinical decision-making: the clinician or embryologist who acts on a recommendation carries responsibility for that action. Vendors cannot be expected to bear clinical accountability, and the regulatory framework for software as a medical device, particularly tools that update continuously, does not yet provide clear standards. That vetting burden falls on the clinical user. This is not an equitable arrangement and it is one of the strongest arguments for our field to accelerate the development of shared validation standards rather than leaving each clinic to navigate vendor claims independently. Programs should ensure that informed consent discussions with patients include a transparent description of what AI tools are used and what the actual capabilities and limitations of those tools, and that ongoing monitoring is in place to detect performance changes over time. The patient who is harmed by an unreliable recommendation did not meaningfully consent to that risk if she was only told that the clinic uses AI to help with embryo selection.

8. Programs should evaluate whether AI tools provide cost-effective value, recognizing that cost reduction alone is not a substitute for safety or efficacy evidence.

   Lower cost is an adoption rationale, not a safety criterion; a tool can be inexpensive and dangerous, or expensive and excellent. Cost pressure is, however, one of the forces driving premature adoption, and programs should be clear that workflow efficiency gains, real as they may be, do not constitute evidence of clinical safety or equivalence.

9. Programs maintain the institutional capacity to vet, monitor, and re-validate AI tools on an ongoing basis, not only at the point of adoption. A validation study conducted at the time of purchase is a snapshot of a moving target. The model may change through vendor retraining, the clinic’s patient population may shift, imaging equipment may be upgraded and laboratory protocols may evolve, any of which can alter model performance without producing an obvious signal. Programs should maintain a local reference dataset representative of their own patients and imaging conditions, define a monitoring schedule, and establish performance thresholds that would trigger suspension of AI guidance pending re-evaluation. Without that infrastructure, a program cannot meaningfully vet what it is purchasing or catch problems after deployment. Only the buyer can protect the buyer.

These criteria are not a final checklist but a starting framework for a conversation our field needs to have now, before commercial adoption outpaces the capacity to evaluate what is being adopted. Our field also faces a structural question that extends beyond any individual tool or clinic: who controls the data used to train and benchmark these models? A path worth consideration is collective data governance through national registries organization such as SART or ICMART, which could manage shared, appropriately consented datasets, allow models from competing developers to be benchmarked against common standards. This would shift the competitive advantage from whoever holds the largest proprietary dataset toward whoever builds the most rigorous and generalizable model.

Without shared data governance, the evidentiary standards for AI in IVF will be set by the companies selling it rather than by the clinicians and embryologists responsible for patient outcomes. This would be a harder road than letting commercial entities set the pace independently, but it is the path on which reproductive medicine retains ownership of its own standards. The window for our field to shape this technology on its own terms will not stay open indefinitely. Letting commercial developers define what counts as good enough is already a well-traveled road and it does not end where we would want it to. The alternative is a field that asks the grounded questions, does the homework, builds the shared infrastructure, and earns the trust that comes from having protected its patients rather than its market position. This requires clinicians and embryologists to be drivers rather than passengers.

Authors’ contributions

Conceptualization DS; original draft preparation DS; Writing, editing and reviewing CB, DS; CB prepared Table 1. All authors reviewed the manuscript.

Funding

No funding was provided for this work.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

Dr. Bourmann: Fujifilm Inc., Nexspring Health Royalties/License payments made towards patents that are held and managed by MassGeneral Brigham. Payments are made to MassGeneral Brigham. WO/2024/249861, WO/2022/056374, WO/2021/046521, WO/2019/068073. Patent families related to this work. DBS is co- Editor-In-Chief of RB&E . All decisions on this manuscript were made by another senior editor.