What question did this study set out to answer?

The aim is to create a framework for non-directive conversational AI that promotes human agency and responsibility.

June 22, 2026Open Access

Design Principles for Non-Directive Conversational AI — A Structural Approach to Preserving Human Agency and Responsibility in Dialogue Systems —

Key Points

The aim is to create a framework for non-directive conversational AI that promotes human agency and responsibility.
Proposes a three-layer architecture separating observation, translation, and human decision responsibility.
Introduces the Human Agency Ratio (HAR) to analyze the distribution of decision agency in conversations.
Develops a Safety Floor model for crisis-state response to ensure low-pressure support.
Establishes HAR as a tool for understanding how conversation turns affect user agency.
Differentiates non-directive support from passive non-intervention to provide safety in critical situations.
Includes a coding-based pilot dataset for future evaluations and inter-rater reliability testing.

Abstract

Abstract This work presents a conceptual and methodological design framework for non-directive conversational AI, with a focus on preserving human agency, user autonomy, and responsibility in dialogue systems. It proposes a three-layer architecture that separates observation, translation and structuring, and human decision responsibility. Within this structure, the AI system is limited to organizing information, clarifying context, and presenting perspectives without asserting conclusions, ranking choices, or substituting its judgment for the user’s decision-making. The framework introduces the Human Agency Ratio (HAR) as an operational indicator for analyzing the distribution of decision agency in conversation. HAR is not presented as a psychological measure of free will, but as a coding-based tool for examining how dialogue turns preserve, defer, or displace user agency. This version includes a blind-coding-oriented pilot dataset based on AI-generated simulated conversations. The dataset is not real-user empirical data and is not intended as statistical validation. Instead, it is provided as publicly inspectable pilot coding material for future independent annotation, critique, replication, and inter-rater reliability testing, including possible use of Cohen’s Kappa. The document also develops a Safety Floor model for crisis-state response. This model distinguishes non-directive support from passive non-intervention by allowing transparent, low-pressure safety support in cases involving life, bodily harm, or serious danger, while avoiding hidden profiling, coercive guidance, or decision substitution. Related concepts include Safety Mode with Soft Transparency, Presence-First Safety, Low-Pressure Direct Safety Check, Message Drafting Support, and External Resource Option. The paper situates the framework in relation to AI ethics, self-determination theory, person-centered support, nudge theory, dark patterns, automation bias, and safety planning intervention. It also includes bilingual Japanese-English examples of user-facing crisis-state response phrases. Keywords: non-directive conversational AI; human agency; user autonomy; AI ethics; Human Agency Ratio; HAR; safety floor; crisis-state response; decision support; automation bias; blind coding dataset; inter-rater reliability; Cohen’s Kappa; AI-generated simulated conversations

Read Full Paperexternally

Ask AI

Mark Helpful

Bookmark

Relay

View Full Paper