[Paper Review] Interactive AI Alignment: Specification, Process, and Evaluation Alignment

Michael Terry, Chinmay Kulkarni+3 more|arXiv (Cornell University)|Oct 23, 2023

Ethics and Social Impacts of AI16 citations

TL;DR

The paper maps AI alignment to a three-step interaction cycle and defines specification, process, and evaluation alignment for interactive AI, introducing surrogate processes and the Process Gulf to improve user control and understanding.

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.

Motivation & Objective

• Map AI alignment concepts onto a basic user-system interaction cycle (input, processing, output).
• Define three interactive alignment objectives: specification alignment, process alignment, and evaluation support.
• Introduce surrogate processes and the Process Gulf to explain and bridge human-AI differences in production of outputs.
• Illustrate the framework with analyses of image-generation and code-synthesis systems.
• Identify future research directions at the intersection of AI alignment and HCI.

Proposed method

• Propose a three-step interaction model (user input, system processing, user evaluation) as a basis for alignment.
• Define and elaborate specification alignment, process alignment, and evaluation support with subcategories (outcome specification, specification constraints, means alignment, control alignment, surrogate process, verification support, comprehension support).
• Introduce surrogate processes as simplified, controllable representations of the AI’s actual process to aid control and understanding.
• Define the Process Gulf to describe qualitative differences between human and AI processes and challenges this creates for control.
• Review and relate to Horvitz’s mixed-initiative principles and Amershi et al.’s guidelines to ground the framework in existing HCI and AI-alignment literature.
• Apply the framework to analyze real systems in image generation and code synthesis to demonstrate descriptive and prescriptive value.

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

Experimental results

Research questions

• RQ1 How can AI alignment be effectively framed within the three-step interaction cycle of interactive AI?
• RQ2 What mechanisms support specification alignment, process alignment, and evaluation support in interactive AI systems?
• RQ3 What are surrogate processes and the Process Gulf, and how do they aid user control and understanding of AI processes?
• RQ4 How can existing HCI and AI-alignment guidelines be leveraged to design better interactive alignment mechanisms?
• RQ5 What lessons do image-generation and code-synthesis systems offer for interactive alignment design?

Key findings

• A structured framework that maps specification, process, and evaluation alignment to interactive AI improves user ability to specify goals, control AI processes, and verify outputs.
• Surrogate processes enable user-friendly control when accessing the AI’s internal methods may be difficult, by providing an alternative, understandable representation that yields the same end result.
• The Process Gulf highlights the difficulties users face when AI and human production processes diverge, underscoring the need for explicit bridge mechanisms.
• Analysis of real interactive AI systems demonstrates that interfaces with alignment mechanisms produce qualitatively different and improved user experiences.
• The work connects AI alignment with HCI theory and existing guidelines to propose concrete areas for future interactive alignment research.

Figure 2. Surrogate processes . A surrogate process is an alternative way of producing the same result as the AI. To create a surrogate process, the AI produces output as normal ( A ). The original input and the AI’s output is then sent to a system to reverse-engineer a process for producing the sam