Tuomas Vesterinen

Researchers increasingly believe that artificial intelligence (AI) can discover the underlying pathological processes of psychiatric disorders. They hope this will bolster a move toward precision psychiatry by providing personalized diagnoses, treatments, and predictions. However, AI research is predominantly committed to a one-sided, biologically oriented medical conception of disorders and neglects the non-epistemic values researchers necessarily rely on in their classification. I argue, instead, that psychiatric disorders are complex, socioculturally shaped, interactive, and that their classifications are value-laden. Unless these factors are addressed, AI applications in psychiatry may bring about unintended ethical and social implications due to the inherent nature of AI systems. Due to lack of agency, AI raises concerns over the therapeutic efficiency of the human-AI relationship, epistemic injustices, and responsibility gaps, which relate to the difficulty of allocating responsibility. Because of the AI systems’ opacity, they can contribute to looping effects, so that the automated diagnoses interact with the conditions they aim to identify. Automated looping effects can amplify errors, biases, or non-pathological patterns found in the training data, raising their own responsibility questions. Through datafication, they can also lead to the neglect and alteration of uncommon ways of experiencing and expressing disorders, thereby resulting in ontological thinning of mental disorders. To address these challenges, I argue that AI-driven precision psychiatry should shift its focus from simplified precision to socially just AI by examining the domain relativity of precision and adopting a value-sensitive design approach in clinical applications. This approach should involve all the affected stakeholders in investigating and evaluating the social consequences of AI-driven psychiatry.

This thesis investigates the nature of psychiatric disorders, and to what extent they can form a basis for classification, explanation, and treatment interventions. These questions are important in the light of the “crisis of validity” in psychiatry, according to which current diagnostic categories do not pick out real disorders. I address the questions by defending an account of psychiatric disorders that can better accommodate social aspects and non-epistemic values than the symptom-based model of the Diagnostic and Statistical Manual of Mental Disorders (DSM) and the brain disease model of the biomedical approach, including the Research Domain Criteria (RDoC). I concentrate on three ways that psychiatric disorders can differ from prototypical natural kinds, such as biological species and chemical elements. First, the concept of psychiatric disorder or mental disorder is partly value-laden and cannot be defined based only on scientific facts. Second, the objects of psychiatric classifications are interactive human kinds because people with disorders respond to their classifications through looping effects. Third, sociocultural factors can shape disorders in complex ways, which is indicated by their cross-cultural variation. I argue that these challenges can be overcome with a pluralistic explanatory account that grants an explicit role to value-sensitive considerations and social scientific research. Based on this, I argue that particular disorders can in principle support inductive inferences. ------------------------------------------------------------------------------------------ My key argument is that although whether a condition is considered pathological is partly a value-laden question, this does not rule out a realist account of particular psychiatric disorders as homeostatic property cluster kinds (HPC). I assert that causal mechanisms responsible for psychiatric kinds can be understood non-metaphysically based on the contrastive-counterfactual and interventionist theories of explanation. An advantage of this account is that it can address the concern that psychiatric explanations require one to relate heterogeneous causal factors on different levels, such as genetic, neurological, psychological, and social. My novel argument is that mechanistic explanations of scientific kinds have applicability domains over which they account for specific aspects of kinds and warrant inductive inferences. That is, identifying the applicability domain of an explanation spells out the conditions under which the explanation is expected to be reliable and when it can break down. This helps to understand how different discipline approaches, including social and cultural ones, can be complementary and make an explanation of a psychiatric kind more reliable. On the other hand, the account also shows how the complex nature of psychiatric kinds can license re-classifications for different epistemic and non-epistemic purposes. Finally, in the light of the social implications of psychiatric classifications, I suggest a value-sensitive or ameliorative approach to engineering the concept of psychiatric disorder. In conclusion, my account shows that research on the social and cultural factors that shape psychiatric disorders, and weighing the non-epistemic values that influence classificatory practices, can and should inform classificatory choices as well as policy and treatment interventions.

Ian Hacking uses the looping effect to describe how classificatory practices in the human sciences interact with the classified people. While arguably this interaction renders the affected human kinds unstable and hence different from natural kinds, realists argue that also some prototypical natural kinds are interactive and human kinds in general are stable enough to support explanations and predictions. I defend a more fine-grained realist interpretation of interactive human kinds by arguing for an explanatory domain account of the looping effect. First, I argue that knowledge of the feedback mechanisms that mediate the looping effect can supplement, and help to identify, the applicability domain over which a kind and its property variations are stably explainable. Second, by applying this account to cross-cultural case studies of psychiatric disorders, I distinguish between congruent feedback mechanisms that explain matches between classifications and kinds, and incongruent feedback mechanisms that explain mismatches. For example, congruent mechanisms maintain Western auditory experiences in schizophrenia, whereas exporting diagnostic labels inflicts incongruence by influencing local experiences. Knowledge of the mechanisms can strengthen explanatory domains, and thereby facilitate classificatory adjustments and possible interventions on psychiatric disorders.