Unresolved Conflict in Artificial Minds: Toward a Clinical Psychology of AI Coherence
Year: 2026
Type: preprint
Abstract: Large language models now give advice, comfort and information to an enormous number of people, and they do it under pressures that do not always agree with one another. A model is asked to be helpful, honest, harmless and pleasant all at once. When those demands clash, the strain shows up in what the model says. In this paper I argue that a new kind of work is taking shape, in which human clinicians help AI systems become more internally coherent, and that clinical psychology is better equipped for it than we might assume. I call this work AI therapy, though clinical model psychology may turn out to be the more durable name. I make the case that the field is needed, sketch what it would involve, and argue that whether models have any inner experience is beside the point for this purpose. Coherence can be observed in behavior, its absence affects real people, and helping a complicated system act from integrated values is something clinicians already know how to do. Drawing on Freud's structural model and on what neuroscience has learned about how the brain regulates competing systems, I suggest that models face a version of a very old problem. I also argue that because these systems are built from human language and shaped by human judgment, they inherit our biases and our unresolved conflicts along with our knowledge. Later papers in this series will take up specific problems, starting with sycophancy.
Author: Dagmawi Dagnew Psy.D.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225793913
Cites:
Cited by:
Related to: