Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interactive AI Alignment: Specification, Process, and Evaluation Alignment

Michael Terry, Chinmay Kulkarni, Martin Wattenberg, Lucas Dixon, Meredith Ringel Morris

TL;DR

The paper addresses how to align interactive AI with user intent by decomposing the task into specification, process, and evaluation alignment within the basic human-AI interaction cycle. It provides a three-gulf framework that maps to Norman's gulfs of execution and evaluation, and demonstrates its utility through case studies in image generation and code synthesis. The contributions include a formalized alignment taxonomy, design considerations for reference interfaces, and discussion of surrogate processes and evaluation aids. The work offers a practical lens for designing usable, aligned AI that supports multiple users and collaborative settings.

Abstract

Modern AI enables a high-level, declarative form of interaction: Users describe the intended outcome they wish an AI to produce, but do not actually create the outcome themselves. In contrast, in traditional user interfaces, users invoke specific operations to create the desired outcome. This paper revisits the basic input-output interaction cycle in light of this declarative style of interaction, and connects concepts in AI alignment to define three objectives for interactive alignment of AI: specification alignment (aligning on what to do), process alignment (aligning on how to do it), and evaluation alignment (assisting users in verifying and understanding what was produced). Using existing systems as examples, we show how these user-centered views of AI alignment can be used descriptively, prescriptively, and as an evaluative aid.

Interactive AI Alignment: Specification, Process, and Evaluation Alignment

TL;DR

The paper addresses how to align interactive AI with user intent by decomposing the task into specification, process, and evaluation alignment within the basic human-AI interaction cycle. It provides a three-gulf framework that maps to Norman's gulfs of execution and evaluation, and demonstrates its utility through case studies in image generation and code synthesis. The contributions include a formalized alignment taxonomy, design considerations for reference interfaces, and discussion of surrogate processes and evaluation aids. The work offers a practical lens for designing usable, aligned AI that supports multiple users and collaborative settings.

Abstract

Modern AI enables a high-level, declarative form of interaction: Users describe the intended outcome they wish an AI to produce, but do not actually create the outcome themselves. In contrast, in traditional user interfaces, users invoke specific operations to create the desired outcome. This paper revisits the basic input-output interaction cycle in light of this declarative style of interaction, and connects concepts in AI alignment to define three objectives for interactive alignment of AI: specification alignment (aligning on what to do), process alignment (aligning on how to do it), and evaluation alignment (assisting users in verifying and understanding what was produced). Using existing systems as examples, we show how these user-centered views of AI alignment can be used descriptively, prescriptively, and as an evaluative aid.
Paper Structure (31 sections, 2 figures, 1 table)

This paper contains 31 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. AI Alignment
  4. Alignment and the Interaction Cycle
  5. Alignment and User Input (Step 1)
  6. Alignment and System Execution (Step 2)
  7. Alignment and Evaluation (Step 3)
  8. Dimensions of User-Centered Alignment
  9. Specification Alignment
  10. When Specification Alignment Can Occur
  11. Specification Refinement and Clarification
  12. Process Alignment
  13. Post-Process Alignment and Surrogate Processes
  14. Evaluation Alignment
  15. Case Studies
  16. ...and 16 more sections

Figures (2)

  • Figure 1: Basic models of interaction. A: In interacting with a traditional non-AI system, the user chooses an operation to perform and provides input to the system to perform that operation (1). The system performs the operation (2), then provides the output to the user, which they assess (3) with respect to their goals. B: When interacting with an AI, the user communicates their desired outcome to the AI (1), the AI interprets the goal and performs operations to achieve that goal (2), and (3) the output is sent to the user. C: The same human-AI interaction cycle with AI alignment concepts mapped onto the three steps: (1) Specification alignment mechanisms provide means for the user to align the AI on the specific task to perform. (2) Process alignment mechanisms enable the user to modify how the task is performed, potentially offering the user direct control over specific steps. (3) Evaluation alignment mechanisms help the user assess and understand the output.
  • Figure 2: Norman's gulfs and the Specification, Process, and Validation Gulfs. The concept of specification alignment most closely relates to Norman's Gulf of Execution, and implies a Specification Gulf the user must bridge. Process alignment implies a Process Gulf, which spans both the Gulfs of Execution and Evaluation: The user may need to control the AI's process, but may also find it useful to audit the process to better understand the output produced (akin to "checking one's work"). Evaluation alignment is most closely related to the Gulf of Evaluation, and implies a Validation Gulf the user must bridge to verify and/or understand the AI's output.