Reconsidering bioethical principles in the era of artificial intelligence: challenges for autonomy, justice, and beneficence in medicine

Full Text

Introduction

The global market for artificial intelligence (AI)-based medical solutions is growing steadily and exponentially, and contemporary machine and deep learning algorithms have already outperformed humans in several diagnostic tasks. These include histopathological examinations in oncology, equivalent images, medical imaging (e.g., diabetic retinopathy based on fundoscopy images), and disease prevention (e.g., predicting the signs of acute injury 48 hours before symptoms appear). However, the widespread use of AI systems in real-world practice is challenging for technical and fundamental legal and ethical reasons.

Conventional humanistic bioethics views the human—a physician, researcher, or other healthcare professional capable of empathy, self-reflection, and verbal justification of their decisions—as a subject of moral responsibility and ethical judgment. AI systems, particularly deep learning architectures, behave as “black boxes” with no interpretable, traceable, and explainable logical path between the inputs and outputs. This results in a significant gap between technological advancements and established legal and ethical norms, which may become even more evident in the future.

AI systems in medicine mark not just a technological improvement, but a fundamental shift in medical judgment, decision-making, and allocation of responsibilities. In the medium term (5–15 years), when AI systems evolve from pilot projects to routine applications in diagnosis, prognosis, and treatment, their impact on bioethical norms will be systemic and irreversible.

This review aimed to retrospectively assess the impact of AI on the evolution of bioethical standards and practice and the adaptation of conventional bioethical approaches in evidence-based medicine.

Study Objectives

1. To provide specific examples of AI use in the context of modern international bioethical standards (beneficence, autonomy, justice).
2. To identify and discuss legal and technical conflicts resulting from errors made by AI systems, including responsibility, transparency, and accountability issues.
3. To develop and justify recommendations for adapting and modifying existing bioethical principles, guidelines, and regulatory procedures, taking into account the specifics of modern “algorithm-based medicine.”

The study aimed to bridge the current conceptual gap between rapid technological advancements and the slower evolution of legal and ethical norms, providing the theoretical and methodological framework for the responsible and ethically justifiable development of AI systems in healthcare.

This work provides a comprehensive, structured review and conceptual analysis. The study included theme-based synthesis, critical analysis, comparative analysis of international and Russian legal and ethical norms, and case analysis. Bioethical standards are traditionally provided as declarations (Declaration of Helsinki, Declaration of Lisbon, Belmont Report) [1], professional codes, and institutional protocols.

The Regulation OF Artificial Intelligence in the Russian Federation

Russia has legal regulations and guidelines for the use of AI^{1, 2} [2–4]. However, according to numerous studies, AI requires flexible, adaptable, and context-sensitive standards^{3, 4} [4–6], because:

• AI algorithms change constantly during training, preventing static control in this field;
• Clinical applications of AI range from routine diagnosis to emergency care, requiring various levels of control;
• Ethical norms for AI use must be regulated at all stages, from decision-making to output generation, including transparency of actions, understandability of decision-making algorithms, and contestedness of the conclusions made by AI systems.

Ethical principles

The key ethical principles of modern bioethics in medicine—beneficence, autonomy, justice, and confidentiality—form a systematized regulatory framework known as principlism.

These principles in their current form were first proposed by American philosophers Thomas Beauchamp and James Childress. Their fundamental work, Principles of Biomedical Ethics [7], first published in 1979, acted as a catalyst for standardizing the language and methods for addressing moral dilemmas in healthcare. The authors’ initial concept included four basic categories: autonomy, non-maleficence, beneficence, and justice. The principle of confidentiality, while not always regarded as a primary concern, follows logically from the fundamental requirement of respect for autonomy, because protecting the privacy of personal information is an essential condition for self-determination.

Principlism was not an innovative concept, but rather, the product of a critical synthesis and adaptation of key aspects of Western philosophy and ethics.

Utilitarianism (specifically, the ideas of Jeremy Bentham and John Stuart Mill), where the moral value of an action is determined by its consequences, that is, the greatest good for the greatest number, provides the philosophical framework for beneficence and, to some extent, justice.

In contrast, Kantian deontology serves as the foundation for respect for autonomy. Kant’s categorical imperative, which requires that a person be viewed as an end rather than a means, laid the philosophical groundwork for modern concepts such as informed consent and recognition of the patient’s inherent value. The principle of non-maleficence is determined directly by the absolute prohibition of harm, which is implicit in deontology [8].

The theory of justice proposed by John Rawls in his fundamental work A Theory of Justice (1971) influenced the concept of justice in bioethics significantly. Rawls’ difference principle, which states that social and economic inequalities are to be arranged so that they are to the greatest benefit of the least advantaged members of society, provided a powerful regulatory tool for equitable distribution of scarce medical resources and equal access to healthcare [9].

The Belmont Report (1979), a fundamental document designed to protect the rights and welfare of study participants, was a direct practical precursor to Beauchamp and Childress’ work [10]. The report defined three core ethical principles: respect for persons, beneficence, and justice. Therefore, this model expanded the original framework by emphasizing the principle of non-maleficence as an independent, equivalent category to define a professional’s moral responsibility more clearly.

Thus, the principlism concept proposed by Beauchamp and Childress was shaped by integrating key aspects of deontology, utilitarianism, and Rawls’ theory of justice and adapting them to the specific needs of biomedicine as outlined in the Belmont Report. This four-principles model has provided a normative and analytical framework for resolving global bioethical dilemmas, including those related to the development of biomedical sciences and the use of medical technologies, taking into account ethical values and moral norms.

Contribution of Russian researchers

The emergence and evolution of bioethics in Russia, with its unique humanistic, philosophical, and religious perspectives on life and health, has contributed significantly to the critical understanding and adaptation of the principlism-based model. Boris Yudin’s and Pavel Tischenko’s contributions are especially noteworthy.

Boris Yudin, Academician of the Russian Academy of Sciences (1943–2021), was one of the founders of bioethics in Russia, aiming to overcome application-specific interpretations of bioethical concepts. Yudin emphasized that these concepts should be viewed in a broader anthropological and philosophical context. He viewed bioethics not just as a set of rules, but rather, as a means of protecting humanity against enormous technological pressures, particularly those associated with new biomedical technologies. Yudin argued that the principles of autonomy and beneficence needed to be reinforced by the principle of non-maleficence in its existential dimension, i.e., as a defense of the individual’s integrity, identity, and human dignity in the face of science’s manipulative powers. Yudin strongly advocated that ethics committees should be more than just regulatory authorities; they should also be platforms for meaningful interdisciplinary debate [11].

Pavel Tischenko, known for his work in philosophy of biomedicine and anthropology, contributed significantly to the critique of the principle of autonomy’s absolutism in its liberal-individualistic interpretation. Drawing on Russian philosophical traditions, he advocated communicative and dialogic decision-making in medicine. His concept of a “communicative community” suggests that ethical choice emerges from the interaction between the patient, the doctor, the family, and the sociocultural context, rather than an individual’s isolated consciousness. This allows for a critical rethinking of the Western model of informed consent, reinforcing it with shared understanding and responsibility [11]. Specific challenges arising from the shift in bioethical concepts must be assessed and understood using specific examples.

Principle of beneficence

In bioethics, the principles of beneficence (actively doing good) and non-maleficence (primum non nocere, or doing no harm) are complementary ethical principles that require healthcare providers and researchers to act in the best interests of patients and society by preventing or minimizing potential harm. The principle of non-maleficence requires avoiding actions that could cause harm, whereas the principle of beneficence necessitates the active pursuit of improving health and well-being that goes beyond simply not causing harm. This principle faces certain challenges in modern conditions, namely, concepts such as the black box dilemma and clinical validation.

AI is capable of learning/self-learning using the following approaches:

1. Deep learning and neural networks. AI is trained on massive datasets to identify and classify disorders. One example is convolutional neural networks.
2. Machine learning. Algorithms are improved (trained using datasets) for algorithm-based forecasting and decision-making. In medicine, machine learning can be used to classify images, detect abnormalities, and predict treatment outcomes [12].
3. Learning and data processing. Effective training of AI models requires large datasets; however, collecting and labeling such datasets in medicine can be challenging. Transfer learning, data augmentation, and generative adversarial networks are used to overcome these difficulties [13].

Healthcare is one of the most impressive examples of AI applications. AI is used for diagnosis, analysis of medical images (X-ray, magnetic resonance imaging, computed tomography, etc.), prognosis, and treatment. For example, deep learning algorithms can analyze magnetic resonance imaging or X-ray scans to detect the signs of cancer, cardiac diseases, or nervous system disorders. This accelerates diagnosis, reduces the risk of human error, and improves diagnostic accuracy because AI outperforms radiologists in distinguishing between normal and abnormal findings, as it does not get tired or distracted [14–16]. AI can aid in detecting fractures on X-rays, identifying lung or breast tumors during routine check-ups, diagnosing Alzheimer disease, assessing brain injury in hemorrhagic stroke, and monitoring changes during treatment.

Modern AI systems assist physicians in selecting the best treatment strategy based on the entirety of patient data; moreover, they can be used to analyze genetic information to develop personalized treatment approaches and predict responses to different therapies. For example, the IBM Watson system is widely used in medicine to assist in decision-making in the treatment of cancer and other difficult-to-treat conditions [14, 17]. AI systems are currently actively used in oncology, neurology, and cardiology [17, 18]. However, it is vital to remember that AI systems are trained on specific examples; if these examples contain errors, the system will reproduce these errors and inaccuracies when analyzing new images [18].

An algorithm for diagnosing pneumonia based on X-rays showed excellent accuracy when used in the hospital where it was developed. However, its accuracy decreased dramatically when implemented in another hospital. It emerged that the AI system had learned to recognize metadata (a specific artifact produced by the X-ray machine in the first hospital), rather than abnormal changes in the lungs. Physicians were unaware of this because they were not familiar with the algorithm’s logic. Thus, an independent clinical assessment of the algorithm’s outputs is impossible. The physician trusts the outputs implicitly, which can result in a fatal error.

Another challenge is the diffusion of physicians’ responsibility. This issue is related to robotic systems in healthcare. In one example, a surgeon used the da Vinci Surgical System to perform surgery. A failure in the system’s algorithm that overcame hand trembling resulted in a vascular injury. Who was to blame here? Was it the surgeon’s fault for not stopping the procedure? Or was it the manufacturer’s fault for providing subpar software? Or, perhaps, the hospital was to blame for poor equipment maintenance? [19, 20]. As a result, the chain of responsibility became overly long and complex, making it difficult to compensate for damage.

These examples illustrate emerging challenges, highlighting the importance of acting quickly to protect both parties (the patient and the physician) involved in the diagnosis and treatment. These measures should aim to meet the following requirements:

• Ensuring explainable AI (XAI). This refers to developing methods to visualize the algorithm’s decision-making (e.g., heat maps indicating which portions of an image the AI has focused on). As a result, the physician will not trust the AI implicitly to make the diagnosis, but will use AI outputs as additional evidence in their clinical judgment.
• Supporting the physician’s leading role in making medical decisions. The physician must make the final clinical decision, while the AI can only provide recommendations.
• Formalized assignment of responsibilities. This refers to developing clear legal protocols that define the responsibility of the developer, physician, or institution, depending on the nature and causes of the error.

Principle of respect for autonomy

The principle of respect for individual autonomy in bioethics is an essential ethical criterion for recognizing and respecting each individual’s right to self-determination and making their own decisions regarding their health, life, and well-being. This includes the right to informed consent during medical interventions. Patient consent to any medical intervention is now considered standard practice in medicine. The physician does not have the authority to perform examinations, diagnostic workups, and therapeutic or rehabilitation procedures without the patient’s consent (or their representative’s consent, as applicable). This principle can also be influenced by digital technologies.

This includes the need to reconsider informed consent concerning self-learning systems. For example, a patient gives informed consent to use an AI system for ECG monitoring. However, the algorithm is trained continuously on new data, and its decision-making logic can change dramatically after a year. What algorithm did the patient originally consent to? The difficulty here is that conventional consent is static, whereas AI systems are constantly evolving.

Furthermore, there are cases where it may be required to protect the patient’s right to an explanation. For example, an algorithm for analyzing magnetic resonance imaging findings recommends denying a patient complex surgery because the chances of success are extremely low. The patient asks, “Why?” And the physician can only explain, “That’s what the program says.” This violates the patient’s right to understand the factors that influence their treatment and outcome, undermining their autonomy and ability to participate in decision-making.

Potential solutions in such cases are as follows:

• Multilevel, dynamic consent. This refers to the implementation of digital platforms where patients can obtain information about the algorithm version and the data used to train it, and give or withdraw their consent to its use at any time.
• The “right to explanation” must be formalized, as is the case with the General Data Protection Regulation (GDPR) in the EU. This regulation establishes the rules for the collection, processing, storage, and distribution of personal data within the EU, requiring developers to provide interfaces that generate explainable conclusions for physicians and patients [21, 22].

Principle of justice

The principle of justice in bioethics implies an equal and fair distribution of the benefits and risks associated with scientific research, healthcare services, and access to healthcare among all social groups, regardless of their social, economic, or other status. This principle requires equal treatment for individuals in the same circumstances, but unequal treatment for those in different circumstances. Furthermore, it emphasizes the need to remove barriers to healthcare services. This principle also faces various challenges in terms of using AI in medicine. One example is algorithmic bias. Consider the well-known example of the Optum algorithm, which was used in the United States for healthcare management. The algorithm assigned the same risk level to black and white patients only if the black patients’ condition was significantly more severe. This was because the model was trained using historical data on healthcare costs. Historically, black patients have had limited access to healthcare due to systemic health inequities, resulting in lower costs for their treatment. The algorithm interpreted this as “less need for medical care” [23].

Another example is that commercial algorithms for processing skin imaging (dermoscopy) findings are significantly less accurate in detecting melanoma in dark skin types, because they were mostly trained on images of light skin. As a result, AI does not eliminate preconceptions, but rather, codifies and scales them, creating a system that can discriminate against vulnerable populations.

Potential solutions are as follows:

• Mandatory justice audit: before adopting any medical AI system, an independent audit must be conducted using a diverse dataset representative of all social groups.
• Ensuring data diversity: encouraging the creation of open, annotated, and diverse medical datasets.
• Transparency: developers must provide access to demographic information for training and testing data.

In bioethics, confidentiality refers to the obligation of healthcare providers and other professionals to keep patient information confidential when providing healthcare services. This is necessary to protect their personal life, social status, and financial interests.

Principle of confidentiality

The principle of confidentiality is vital for building trust between physicians and patients, encouraging open communication and preventing the potential stigma and discrimination associated with medical data. This principle is closely related to the Hippocratic Oath, which emphasizes respect for the patient and their privacy as a vital component of ethical medical practice.

This is a serious challenge. Large sets of personal medical data are required to train powerful algorithms, but the centralization of such data poses a significant risk of data leakage [24, 25].

Potential solutions are as follows:

• Federated learning: an advanced technique that allows decentralized training of an algorithm. The data remain on the hospital’s servers, the model is trained locally, and only model parameter updates are forwarded to the central server, not the data. This reduces confidentiality risks significantly.
• Differential privacy: adding statistical noise to data, which prevents specific individuals from being identified while preserving general statistical patterns for training algorithms.

This analysis demonstrates that bioethics is on the verge of a paradigm shift. We are shifting from human ethics to system ethics, which necessitates specific actions:

1. Expanding ethical boundaries: Beauchamp’s principles (non-maleficence, beneficence, justice, and respect for autonomy) must now be reinforced by technologically relevant principles:

• Transparency to maximize the explainability of algorithmic decisions

• Accountability to formalize responsibility for each link in the “developer–physician–institution” chain

• Justice to incorporate both technological and social justice into the algorithm’s design.

1. Changing education: future physicians must have not only medical literacy but also algorithmic literacy, which means knowing the fundamentals of AI, its limitations, and potential errors [25].
2. The principle of feedforward control: regulatory authorities (e.g., Roszdravnadzor in Russia) must develop standards and assess both the efficacy and ethical aspects of medical AI systems.

AI is more than just a new tool in medicine; it also has the potential to redefine the ethical foundation of healthcare. Rethinking bioethical principles is not a repudiation of the past, but rather, an important and necessary step forward. The success of AI integration in medicine will be assessed not only by its accuracy and efficacy but also by its ability to strengthen rather than undermine core values such as trust, justice, and respect for autonomy. This calls for the collaboration of developers, physicians, bioethicists, lawyers, and regulators⁵ [26, 27].

Regulation of AI is a complex and multifaceted task. Over-regulation of AI may stifle innovation, while under-regulation may result in major harm to citizens’ rights and a wasted opportunity to build a future civilized society [26].

AI’s main task is the automated execution of similar activities, when machine learning and deep learning are required for computer-aided training [28–30].

Conclusion

This work identified fundamental contradictions between classical principles of medical ethics and the realities of modern medicine when using AI. These include the conflict between the principle of beneficence and the black box dilemma, diffusion of responsibility, flaws in the principle of autonomy, the need to revise the informed consent model for dynamic algorithms, and inadequate principles of justice and confidentiality.

To maintain trust in medicine, traditional principles must be refined rather than abandoned. They must be reinforced with new concepts such as transparency, accountability, and technological justice. This calls for the development of new regulatory standards, mandatory algorithmic audits, and the integration of ethical design into medical AI systems.

Additional information

Author contributions: N.F.T.: conceptualization, supervision, writing—review & editing; G.M.E.: investigation, writing—original draft, writing—review & editing. All the authors approved the version of the manuscript to be published and agreed to be accountable for all aspects of the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Funding sources: No funding.

Disclosure of interests: The authors have no relationships, activities, or interests for the last three years related to for-profit or not-for-profit third parties whose interests may be affected by the content of the article.

Statement of originality: No previously obtained or published material (text, images, or data) was used in this study or article.

Data availability statement: The editorial policy regarding data sharing does not apply to this work, as no new data was collected or created.

Generative AI: No generative artificial intelligence technologies were used to prepare this article.

Provenance and peer-review: This paper was submitted unsolicited and reviewed following the standard procedure. The peer review process involved three external reviewers, a member of the Editorial Board, and the in-house science editor.

¹ National Strategy for Artificial Intelligence Development Through 2030. Approved by Presidential Executive Order No. 490 of October 10, 2019, On Artificial Intelligence Development in the Russian Federation. Available at: http://kremlin.ru/acts/bank/44731. Accessed on: May 1, 2025.

² Order of the Federal Agency for Technical Regulation and Metrology (Rosstandart) No. 1732 of July 25, 2019 (as amended on January 20, 2021), On Establishing the Artificial Intelligence Technical Committee. Available at: https://www.rst.gov.ru/portal/gost/home/activity/documents/orders#/order/104460. Accessed on: May 1, 2025.

³ Artificial Intelligence Alliance. Code of Ethics in AI [Internet]. © AI Alliance Russia. Available at: https://ethics.a-ai.ru/. Accessed on: February 22, 2025.

⁴ Artificial Intelligence Center of the HSE University. Ethics of Artificial Intelligence. Available at: https://cs.hse.ru/aicenter/ethics. Accessed on: February 22, 2025.

⁵ ResearchGate [Internet]. 2008–2025. ResearchGate GmbH. Available at: https://www.researchgate.net/?ref=logo&_sg=Ro8tDhb7EvgtUnJVWpW2kmsOvwHW2LdsHfPvsp104ts4jDuycM63CeWrDnPF_iVAGuoP4KPMLNeGbIY&_tp=eyJjb250ZXh0Ijp7ImZpcnN0UGFnZSI6InB1YmxpY2F0aW9uIiwicGFnZSI6InB1YmxpY2F0aW9uIiwicG9zaXRpb24iOiJnbG9iYWxIZWFkZXIifX0. Accessed on: September 17, 2025.

About the authors

Farida T. Nezhmetdinova

Author for correspondence.
Email: nadgmi@mail.ru
ORCID iD: 0000-0003-2875-128X
SPIN-code: 8441-6943

Cand. Sci. (Philosophy), Assistant Professor

Marina E. Guryleva

Email: meg4478@mail.ru
ORCID iD: 0000-0003-2772-129X
SPIN-code: 6207-9971

MD, Dr. Sci. (Medicine), Professor

References

Supplementary files