Legal and ethical challenges of artificial intelligence in obstetrics


Reproductive healthcare carries uniquely high stakes for patients, families, and clinical teams. In 2023, an estimated 260,000 maternal deaths occurred across the globe. Alongside this severe clinical burden sits immense medico-legal vulnerability. In England, obstetric claims represent only 11% of clinical negligence claims by volume, yet they make up 55% of the total financial value of all claims. As software developers introduce artificial intelligence into ultrasound analysis, risk prediction models, decision support, robotic surgery, remote monitoring, and direct patient communication, software failures, bias, or unexplained outputs carry severe clinical and legal liabilities.

How the evidence was synthesised

This study is a narrative review with a thematic synthesis designed to map the primary medico-legal and bioethical challenges of artificial intelligence in obstetrics and gynaecology, and to set out a lifecycle governance framework. The authors searched PubMed, Scopus, and Google Scholar for material published between January 2020 and August 2026, supplemented by citation tracing and searches of official legal databases. Two reviewers evaluated 131 records against predefined criteria. Ultimately, 41 sources met the inclusion threshold, comprising 36 academic publications and five formal legal instruments. The selection path was tracked using a PRISMA-style flow diagram.

As a narrative review, this work offers conceptual clarity rather than direct empirical testing. It synthesises regulations, legal principles, and professional obligations, but it does not evaluate patient outcomes or run clinical validation trials. The conclusions should be interpreted as an informed governance blueprint rather than proof of how specific software performs in practice.

Six risk domains and the myth of displaced liability

From their thematic analysis, the authors identified six interlinked areas of concern:

  • Validation and generalisability across diverse populations.
  • Professional competence and the risk of automation bias, where clinicians might defer excessively to automated outputs.
  • Distributed liability and evidentiary opacity, where black-box models obscure why an erroneous output was generated.
  • Informed consent and patient autonomy.
  • Privacy and cybersecurity vulnerabilities surrounding sensitive reproductive health records.
  • Systemic discrimination, unequal access, and reproductive justice.

The review also examined generative tools, identifying acute operational dangers including algorithmic hallucinations, non-reproducible answers, persuasive framing that can mislead users, and unmonitored, undocumented staff usage. Importantly, the review dispels the assumption that using automated systems shifts legal liability away from the bedside. While legal exposure is shared across manufacturers, health services, and staff, legal frameworks hold that clinical judgment and documented supervisory oversight remain central. Under European Union law, the Medical Device Regulation and General Data Protection Regulation apply immediately. Dedicated AI Act requirements for AI-enabled medical devices take effect in August 2028, and the updated Product Liability Directive applies to products released after 9 December 2026. However, proving causation and establishing professional fault remain matters of individual national law.

Translating findings to health service governance

For health services, clinical educators, and informatics teams, integrating these tools safely demands structured institutional processes rather than simple reliance on vendor accuracy claims. The authors translate their findings into seven operational governance requirements:

  • Clear, accountable institutional ownership for every deployed system.
  • Baseline local validation to ensure algorithms perform reliably on local patient cohorts before release.
  • Role-specific training for staff to prevent blind deference and identify system faults.
  • Auditable clinical documentation that records when software advice was reviewed, accepted, or rejected.
  • Ongoing equity monitoring to detect algorithmic bias against demographic subgroups.
  • Active post-deployment surveillance to detect real-world performance drift.
  • Standardised incident reporting mechanisms dedicated to software-related near misses and errors.

In low-resource environments, where the burden of maternal mortality remains highest, introducing automated screening or triage without local validation or reporting structures risks compounding disparities. The authors stress that these technologies should augment rather than displace accountable, direct clinician-patient relationships.

Study limits and open questions

The study relies on a narrative synthesis of 41 items drawn from an initial pool of 131 screened records. It does not measure the actual error rates of obstetric software in clinical trials, nor can it establish the legal standard of care in every international jurisdiction. Regulatory details analysed in the paper focus heavily on European legislation, which limits their direct application in jurisdictions operating under different statutory or common-law frameworks. Further work is required to test how local validation rules and mandatory incident logs function under the everyday pressures of busy labour wards and outpatient clinics.

Reference

Santangelo G, De Angelis G, Muzii L, Di Donato V, Perniola G, Napoletano G, et al.. Artificial Intelligence in Gynecology and Obstetrics: Medico-Legal, Ethical, and Governance Challenges-A Narrative Review. Healthcare (Basel, Switzerland). 2026. doi:10.3390/healthcare14213760. PMID: 43028780.

This research digest summarises a recently published article. It was drafted with AI assistance from the article's abstract and reviewed by the author before publication.

Comments