Introduction

Artificial intelligence (AI), particularly through machine learning (ML), represents one of the most significant advances in modern medicine. This technology relies on algorithms that learn from vast amounts of data to make decisions or predictions. Unlike traditional computer programs where every step is explicitly defined by programmers, ML uses data, loss functions, and neural network architectures to autonomously discover solutions. The training process resembles mathematical optimization: the algorithm searches for optimal parameters (weights) to minimize error, much like finding the lowest point in a complex mountain landscape. This enables systems to process information in ways that surpass human speed and scale, yet it also introduces risks due to the lack of direct human control over how decisions are reached.

Despite the excitement, ML has fundamental limitations that must be addressed for safe medical application. One key issue is the delegation of responsibility: in classical software development, humans understand and test every detail, whereas in ML, the solution emerges automatically through training, making it difficult to predict behavior on unseen data. This can lead to errors, as illustrated by the famous 'Husky vs. Wolf' experiment, where an image classification model learned to rely on snowy backgrounds rather than actual animal features (Ribeiro et al., 2016). In medicine, similar spurious correlations can be dangerous—for instance, a model linking a disease to demographic factors due to biased data rather than true causal relationships. As Judea Pearl emphasizes, data contain only correlations, not causation, which is essential for informed medical decisions (Pearl, 2018).

Another limitation concerns problem solvability. Mathematically, ML represents an implicit definition: it is possible that no solution satisfies all conditions, analogous to seeking a number that is simultaneously greater than 4 and less than 3. In medicine, this manifests when data are insufficient to capture the full complexity of the human body or when ethical concepts like fairness cannot be adequately quantified through loss functions. Datasets often carry societal biases—for example, medical databases predominantly based on Western populations may lead to discriminatory outcomes for other ethnic groups. Furthermore, many models function as 'black boxes': their internal workings remain opaque, similar to how we understand individual neurons but not the entire human brain. This opacity undermines trust in medical applications, where explainability is crucial for physicians and patients.

Algorithms for image processing can detect tumors on X-rays with accuracy surpassing experienced radiologists, or predict disease outbreaks by analyzing social media and health data. In oncology, ML enables personalized therapies based on genetic profiles, while in psychiatry, AI-powered chatbots provide mental health support. In occupational medicine, the focus lies on prevention: ML can analyze data from wearable devices to predict workplace injury risks, such as respiratory issues in industrial workers or stress in office environments. These applications hold promise for improving worker health, but they require rigorous validation to avoid errors that could affect safety and rights.

Looking to the future, predictions for AI in medicine are optimistic yet cautious. By 2030, we anticipate widespread integration of causal models, inspired by Pearl's work, enabling a shift from mere correlations to genuine causal understanding and thereby improving diagnostic accuracy. Advances in explainable AI (XAI) will make models more transparent, allowing physicians to 'interrogate' decisions as they would with a colleague. In our vision, hybrid systems—combining AI with human expertise—will dominate, such as real-time virtual assistants analyzing symptoms with final decisions resting with doctors. In occupational health, this could include predictive analytics for corporate wellness, integrating IoT sensors and electronic health records to forecast work-related illnesses. Achieving this vision requires global regulation enforcing ethical standards, algorithm transparency, and professional education. Challenges like data privacy and potential misuse must be tackled to prevent dystopian scenarios where AI supplants empathy.

Conclusion

In conclusion, artificial intelligence is not a magical solution but a powerful tool with inherent logical boundaries. In medicine, it has the potential to revolutionize healthcare through predictive and personalized approaches—but only if transparency, ethics, and integration with the human element remain central. With responsible development, AI can help create a healthier future where technology serves humanity, rather than the other way around.

References:

• Ribeiro MT, Singh S, Guestrin C. 'Why should I trust you?' Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016: 1135-1144.

• Pearl J. The Book of Why: The New Science of Cause and Effect. New York: Basic Books, 2018.