Artificial intelligence in human-assisted reproduction: a paradigm shift still in search of clinical impact

Artificial intelligence in human-assisted reproduction has attracted intense interest and inflated expectations, with proposed applications ranging from ovarian stimulation to gamete and embryo selection and outcome prediction. Despite the initial enthusiasm, its real-world clinical value remains uncertain. This review critically reassesses the current evidence to clarify where artificial intelligence meaningfully contributes and where expectations exceed demonstrated impact.

Most published studies show relevant methodological weaknesses, including limited reproducibility, poor external validation, scarce explainability, and weak comparison with standard clinical practice. Research efforts have disproportionately focused on embryo selection, an area with intrinsically constrained potential to improve treatment efficacy, while other clinically relevant domains remain underexplored. As a result, reported improvements often concern surrogate or intermediate endpoints rather than robust clinical outcomes.

Artificial intelligence holds greater promise in domains such as gamete assessment, automated data extraction, and personalized outcome prediction, where it may enhance treatment management, counselling, and decision-making for both clinicians and patients. Realizing this potential requires a strategic shift in research priorities and rigorous adherence to shared standards, including model transparency, uniformity, external validation, and benchmarking against established clinical workflows. Without such recalibration, artificial intelligence risks becoming a hyped technology with limited clinical relevance in assisted reproduction.