Personal data anonymization in the context of artificial intelligence: from de-Identification to anonymization

Background . The processing of medical data by artificial intelligence (AI) systems requires mandatory de-identification, yet simple removal of direct identifiers does not ensure anonymity. Modern AI models can re-identify individuals from indirect attributes, leaving a measurable residual risk. Objective . The aim of this study was to define the boundaries of de-identification in the AI era, systematize protection methods, and assess the actual attitudes of future physicians toward patient data handling. Methods . We conducted an anonymous cross-sectional survey of 1,283 medical university students and residents (me dian age 21 years). A self-administered questionnaire was used. Data were analyzed in Python (pandas) using chi-squared tests and Cramér’s V to compare differences across training levels. Results . Entering de-identified clinical text into cloud-based AI services was considered acceptable by 61.5 % of respon dents; 12.4 % had already uploaded real patient data, with 1.2 % doing so in identifiable form. Awareness of risks did not deter risky behavior: those who uploaded data had higher self-assessed risk knowledge (3.53 vs. 3.36, p < 0.0001), and 22 % admitted their actions were unacceptable yet continued them. The proportion of uploaders increased from 7.5 % in junior years to 21.0 % in senior years (χ² = 43.3, Cramér’s V = 0.18, p < 0.0001), while willingness to inform patients about AI use declined from 43.4 % to 24.4 %. Conclusions . Anonymity is not a state to be achieved but a continuous process of managing residual re-identification risk. Effective protection requires institutional bans on sharing de-identified data with public cloud services, adoption of methods with quantifiable residual risk (differential privacy, federated learning, homomorphic encryption), and strengthening education al curricula on ethics and information security.