- AI can detect latent regularities across diverse data to propose transdiagnostic dimensions, trajectories, and hybrid entities, acting as an ontological engine rather than revealing essences.
- Statistical clusters are provisional; only reproducibility, clinical validity, interpretability, professional acceptability and patient meaningfulness can legitimise them as ontological categories.
- Adopt dynamic, relational, pluralistic ontologies with human validation loops, participatory governance, open standards and continuous revision to mitigate opacity, ossification and ethical risks.
Encephale. 2026 Sep 19:S0013-7006(26)00173-9. doi: 10.1016/j.encep.2026.07.003. Online ahead of print.
ABSTRACT
OBJECTIVES: Psychiatric classifications remain indispensable clinical tools, yet they do not define mental disorders as natural entities with stable boundaries. This article examines whether artificial intelligence may simply refine existing nosologies or contribute to the emergence of a new psychiatric ontology.
METHODS: We conducted a conceptual and integrative analysis combining philosophical ontology, principles of computational ontology engineering, and recent literature on artificial intelligence, digital phenotyping, knowledge representation and psychiatric classification. The aim was not to test an empirical hypothesis but to clarify the epistemic conditions under which psychiatric categories may be constructed, revised and legitimised in the era of computational models.
RESULTS: Artificial intelligence may act as an ontological engine by detecting latent regularities across heterogeneous data sources, including clinical descriptions, questionnaires, digital traces, language, biomarkers, neuroimaging and environmental exposures. These outputs may generate candidate concepts, such as transdiagnostic dimensions, longitudinal trajectories or hybrid clinical entities. However, statistical clusters cannot be considered ontological categories unless they meet conditions of reproducibility, clinical validity, interpretability, professional acceptability and patient meaningfulness. Conversely, explicit ontologies are necessary to make artificial intelligence systems more transparent, interoperable and clinically intelligible.
DISCUSSION: The emergence of artificial intelligence-assisted ontology in psychiatry carries major risks: confusion between correlation and clinical entity, illusion of objectivity, algorithmic opacity, digital ossification of provisional categories, neglect of subjectivity, and political or cultural capture of knowledge infrastructures. A psychiatric ontology should therefore remain dynamic, temporal, relational and pluralistic.
CONCLUSIONS: Artificial intelligence does not reveal the essence of mental disorders. Rather, it offers provisional maps that can reorganise clinical reasoning and research. Its contribution will be clinically and ethically valuable only if embedded within human validation loops, participatory governance, open standards and continuous revision.
PMID:42763252 | DOI:10.1016/j.encep.2026.07.003
Share Evidence Blueprint
Save to Google Notes
Search Google Scholar
Save as PDF
⭐ My Revision List
