A Diagnosis From an Algorithm: Who Is Responsible When AI Gets Medicine Wrong

A quiet revolution is under way in operating rooms and diagnostic suites around the world. Machine learning algorithms read X-rays, analyze biopsy results, detect arrhythmias on electrocardiograms and predict the likelihood of cancer. According to the World Health Organization, by the start of 2026 more than 500 AI systems had received regulatory approval for clinical use in the United States alone. The technology works, and often works impressively. But behind the façade of technological optimism sits a question that neither physicians, nor legislators, nor developers have yet answered convincingly: who is responsible when the algorithm makes a mistake?

The problem is not theoretical. In 2025, several high-profile cases in Europe and Asia showed that AI assistants can miss critical pathologies or, conversely, generate false-positive results that lead to unnecessary invasive procedures. In each case, responsibility was diluted among several players in the chain: the developer said the system was merely a decision-support tool, the hospital pointed to the product's certification, and the doctor was caught between their own expertise and the recommendation of a machine that had been more accurate than they were in thousands of previous cases.

The root of the problem lies in the architecture of modern medical AI. Most clinical systems are built on a "black box" principle: deep neural networks make decisions whose logic cannot be fully explained even by their creators. When a radiologist looks at a scan and issues a conclusion, they can justify every step of their reasoning. When an algorithm does the same, it produces a probability and an attention heat map, but not a clinical rationale. This creates a fundamental gap: the doctor must either trust the system blindly or carry out a full independent assessment every time, which wipes out the efficiency gains.

Another facet of the problem is the so-called automation effect. Decades of research in aviation and industry have shown that people tend to over-rely on automated systems and gradually lose their own critical-assessment skills. In medicine, this effect is potentially more dangerous than in any other field. Young doctors who work with AI assistants from their first years of training risk never developing the depth of clinical thinking that allowed their predecessors to catch rare pathologies that do not fit statistical patterns. An algorithm is trained on millions of typical cases, but medicine is in large part the art of handling exceptions.

The regulatory landscape still looks like a patchwork quilt. The European AI Act, which is taking effect in phases, classifies medical AI systems as high-risk and requires transparency, but the specific mechanisms for allocating responsibility remain vague. In Russia, the Ministry of Health is actively promoting the digitalization of healthcare, but the regulatory framework for AI diagnostics is developing more slowly than the technology itself. The question of legal liability runs into a classic dilemma: an AI system is not a legal entity. It cannot bear responsibility, lose its license or be taken to court.

The most mature approach is emerging in the United Kingdom, where the National Health Service has developed a framework that splits responsibility into three levels. The developer is responsible for validating the algorithm and ensuring its safety. The medical institution is responsible for correct deployment and monitoring. The doctor retains final clinical responsibility, provided they have adequate tools to critically assess AI recommendations. It is not a perfect solution, but it is at least a workable framework that lets the field move forward.

The industry will have to accept an uncomfortable truth: the technological maturity of medical AI has far outrun institutional readiness to use it. Algorithms already outperform average specialists in a number of narrow diagnostic tasks, but a healthcare system is not only about accuracy in recognizing pathologies. It is also patient trust, legal protection, ethical standards and human empathy, which cannot be digitized. The future of medical AI lies not in replacing the doctor but in building a new model of partnership, in which the boundaries of responsibility are defined as clearly as treatment protocols. Until that happens, every algorithmic diagnosis remains an experiment: technically brilliant, but legally and ethically vulnerable.