From Intelligent Tutoring to Computerized Psychotherapy

David Servan-SchreiberAAAI 1987

Abstract This paper explores the intersection of artificial intelligence and clinical psychology, proposing that cognitive psychotherapy can be modeled as a specialized form of Intelligent Tutoring System (ITS). By treating the therapeutic process as the identification and correction of a patient's misconceptions, the author introduces a framework for "computerized psychotherapy." This approach is instantiated in Sexpert, a system designed to guide couples through the treatment of sexual dysfunctions using hierarchical dialogue planning and a specialized "bug library" of psychological distortions.

Introduction

In the late 1980s, computerized psychotherapy was in its infancy, largely limited to "bibliotherapy"—electronic textbooks—or simple frame-based systems. These early attempts lacked a sophisticated model of the patient’s internal state, leading to issues with user compliance and engagement. David Servan-Schreiber (1987) argued that therapy is not merely the delivery of facts but a process of "debugging" a patient's mental model. This paper posits that the architectures used to teach students complex subjects like physics or programming could be adapted to the clinical setting.

Therapy as an "Infected" Domain

A key theoretical contribution of the paper is the distinction between "isolated" and "infected" domains. In an isolated domain, such as learning a new programming language, the student typically lacks prior knowledge, allowing the tutor to build a model from scratch. In contrast, therapy deals with "infected" domains—areas like human relationships where patients arrive with deep-seated, prior misconceptions. In these cases, the system cannot simply teach the truth; it must actively "un-teach" errors. This requires a diagnostic module capable of mapping user responses to specific cognitive distortions.

System Architecture: The Sexpert Framework

The author developed Sexpert to demonstrate these principles. The system relies on three primary AI components to manage the complexities of therapeutic interaction:

  1. The Diagnostic Module (The "Bug Library"): Rather than assuming the patient is a blank slate, Sexpert utilizes a library of common cognitive distortions and misinformation. A rule-based inference engine performs "classification problem solving" to identify which specific "bugs" are present in the patient's belief system.
  2. The Dialogue Driver: To maintain a coherent conversation without appearing rigid, the system uses hierarchical planning. It moves from top-level abstract goals (e.g., investigating contributing factors) down to specific questions.
  3. Demons: These are "noticing" mechanisms that monitor the dialogue for sensitive triggers. If a user mentions a critical topic like pregnancy or trauma, these demons can interrupt the planned hierarchy to address the immediate concern.
Hierarchical structure of the therapeutic dialogue plan used in Sexpert.

Methodology and Evaluation

The Sexpert system was built with a knowledge base of 1,200 rules and approximately 170 pages of text. The domain knowledge was manually compiled from clinical literature and expert consultations. To evaluate the prototype, 15 unscreened volunteer couples interacted with the system.

While the system was limited to multiple-choice and yes/no inputs, the qualitative results were promising. All 15 couples reported that the dialogue felt logical and intelligent. Some users even described the system as "smart," suggesting that the hierarchical planning successfully mimicked the flow of a human therapist.

Discussion and Key Insights

A significant finding from the pilot study was the extreme sensitivity of users to linguistic nuance. For instance, users reacted more positively to the word "concern" than "difficulty" when discussing sexual performance. This highlights the high stakes of natural language generation in clinical AI; a single poorly chosen word can derail the therapeutic alliance.

Furthermore, the paper emphasizes the necessity of "discourse knowledge." For therapy to be effective, the AI must understand the structure of a conversation, allowing it to backtrack or pivot when a patient provides inconsistent information.

Limitations and Future Directions

Despite its successes, Sexpert faced several hurdles. The "bug library" approach is labor-intensive, requiring hand-crafted rules for every new psychological issue, which limits scalability. Additionally, while the system could identify what a misconception was, it lacked a model of why the patient held that belief. Future work in this field points toward more systematic studies of belief revision and the integration of free-form natural language processing to move beyond the constraints of menu-based interaction.

In conclusion, by framing therapy as a specialized form of tutoring, this research provided an early blueprint for how AI might one day provide individualized, scalable mental health support.

Summary updated

BibTeX
@inproceedings{servanschreiber1987from,
  author = {David Servan-Schreiber},
  title = {{From Intelligent Tutoring to Computerized Psychotherapy}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure