Comment on Starke et al.: ‘Computing schizophrenia: ethical challenges for machine learning in psychiatry’: from machine learning to student learning: pedagogical challenges for psychiatry

Christophe Gauld, Jean‐Arthur Micoulaud‐Franchi, Guillaume DumasPsychological 2020 · DOI · OpenAlex

Abstract

This commentary responds to the growing intersection of Artificial Intelligence (AI) and psychiatry, as discussed by Starke et al. While technical hurdles in machine learning are often the focus, Gauld et al. argue that the primary bottleneck is human: the medical community is currently under-equipped to critically handle the "black box" of psychiatric AI. The authors propose a pedagogical framework grounded in medical epistemology to prepare the next generation of physicians.

The Human Bottleneck in Psychiatric AI

As AI begins to reshape psychiatric diagnosis and treatment, ethical "explainability" has become a clinical necessity. Despite a three-fold increase in PubMed listings for AI in psychiatry over just three years, direct clinical applications remain scarce. This gap is not merely technical; it is ethical and educational.

Current psychiatric practice risks falling into "reductionism" or "non-transparency" if clinicians treat AI outputs as infallible or incomprehensible. Traditional ethical pillars—autonomy, beneficence, non-maleficence, and justice—are insufficient for the AI era. Instead, the authors advocate for a framework centered on Explainability, defined through two lenses:

  1. Intelligibility: Understanding the underlying mechanics of the algorithm.
  2. Accountability: Determining responsibility for algorithmic decisions.

The core challenge lies in translating these abstract philosophical concepts into the daily practice of young physicians.

A Three-Tiered Pedagogical Roadmap

The authors synthesize existing ethical frameworks into a proposed three-tiered roadmap for modern psychiatric education. This approach moves beyond simple technical literacy to include medical epistemology—the study of the nature and limits of knowledge.

  • Clinical Level: Training students to interpret AI results critically. This involves distinguishing between "prediction" (identifying what will happen) and "inference" (understanding why it happens).
  • Academic Level: Developing "dual competency" curricula. These programs should blend data science and biomedical engineering with ethics and epistemology, ensuring that future psychiatrists understand the theoretical models behind the code.
  • Collaborative Level: Breaking down institutional silos. The authors point to successful partnerships, such as those between the University of Toronto and the Vector Institute, as models for integrating medical schools with AI research hubs.

Key Insights: Epistemology as a Safeguard

The authors argue that without a foundation in epistemology, AI training in medicine could lead to a "massive misuse of algorithms." Physicians must understand the models and theories that inform AI outputs to avoid blind reliance on "black box" systems.

Furthermore, AI allows for "stratification"—identifying "micro-phenotypes" or specific brain-symptom circuits that move beyond traditional, rigid categorical diagnoses like those found in the DSM. However, for these micro-phenotypes to be clinically useful, doctors must be trained to navigate this new landscape of knowledge. The authors envision a dynamic "to-ing and fro-ing" where scientists adopt philosophical concepts and philosophers adopt scientific rigor.

Proposed Framework for Development

The primary contribution of this commentary is a structured table outlining the pedagogical challenges and goals for the field across clinical, academic, and collaborative levels.

Table 1: Levels of development for integrating AI into psychiatric education, outlining theoretical and practical goals.

By grounding AI training in these principles, the medical community can develop systems that foster—rather than erode—the essential trust between patients and physicians.

Limitations and Future Directions

While the proposal is proactive, the authors acknowledge a current lack of direct clinical applications for AI in psychiatry. A significant hurdle remains the scalability of this model. While elite institutions like Harvard, MIT, and the University of Toronto can implement "dual competency" programs, applying this globally to already overstuffed medical curricula is a massive logistical challenge. Future efforts must focus on designing and testing these interdisciplinary modules in diverse educational settings to ensure that the benefits of AI are accessible and safely managed across the field.

Summary updated

Abstract

Comment on Starke et al.: 'Computing schizophrenia: ethical challenges for machine learning in psychiatry': from machine learning to student learning: pedagogical challenges for psychiatry - Volume 51 Issue 14
BibTeX
@inproceedings{gauld2020comment,
  author = {Christophe Gauld and Jean‐Arthur Micoulaud‐Franchi and Guillaume Dumas},
  title = {{Comment on Starke et al.: ‘Computing schizophrenia: ethical challenges for machine learning in psychiatry’: from machine learning to student learning: pedagogical challenges for psychiatry}},
  booktitle = {Psychological},
  year = {2020},
  doi = {10.1017/s0033291720003906},
}
Figure