Artificial Intelligence in Clinical Decision-Making: Applications, Challenges, and Implications for Patient Safety

Artificial intelligence is increasingly positioned as a means of supporting diagnosis, predicting deterioration, selecting treatment, and organizing clinical information. Its contribution to patient safety, however, depends on the interaction between algorithms, professionals, patients, and healthcare organizations. This narrative review examines applications of artificial intelligence in clinical decision-making and develops a safety-oriented interpretation of the challenges surrounding implementation. Selected clinical studies, methodological publications, reporting guidelines, and international ethical guidance are discussed without claiming a systematic search or pooled estimate of effectiveness. Applications in diagnostic imaging, physiological monitoring, medication management, and language-based assistance illustrate both the opportunities and the limits of current approaches. Major hazards include unrepresentative data, inappropriate outcome proxies, poor calibration, distribution shift, automation bias, alert fatigue, privacy failures, and fabricated generative outputs. A hypothetical graph demonstrates how an apparently accurate alert can have low positive predictive value when the target condition is uncommon. A comparative table links application areas with plausible harms and practical safeguards. The review argues that predictive performance is an intermediate property rather than proof of patient benefit. Safe implementation requires explicit intended use, independent and local validation, prospective assessment of the clinician–system interaction, transparent communication of uncertainty, and continuous monitoring after deployment. Accountability must be distributed clearly across developers and healthcare institutions while preserving meaningful clinical responsibility. Artificial intelligence can strengthen clinical decision-making when its authority remains proportionate to its evidence and when institutions maintain reliable mechanisms for questioning, correcting, and withdrawing unsafe systems.