Interactive AI Alignment: Specification, Process, and Evaluation Alignment

1 Introduction

• Users increasingly provide high-level natural-language descriptions of desired outcomes, while the AI determines how to achieve them rather than invoking a specific operation.This declarative interaction contrasts with traditional interfaces, where users select an operation and supply its input.Source: §1 · pp. 1, 3

Specification alignment

• Specification alignment concerns what the AI should do, including the possibility of aligning the user on what the AI can do.Relevant challenges include ambiguous or underspecified input and communicating system capabilities.Source: §1 · p. 3, Specification alignment · p. 4

Process alignment

• Process alignment concerns how the AI produces the outcome, giving users some degree of control and transparency over its process.The paper describes it as aligning the user and AI on how to achieve the desired outcome.Source: §1 · p. 3, Process alignment · p. 4

1 Introduction

• Case studies use the alignment objectives to assess existing interactive AI systems and suggest opportunities to improve their interfaces.The paper also connects AI alignment and HCI concepts to highlight common ground and joint opportunities.Source: §1 · p. 3

Evaluation alignment

• Evaluation alignment helps users assess whether the AI met their objective and, more strongly, understand its output.Mechanisms may include sandboxed code execution or appropriately commented code.Source: §1 · p. 3, Evaluation alignment · p. 4

2 Related Work

• The review organizes user interaction into input, system execution, and output assessment, with the cycle repeating when users take further action.It draws on AI alignment and HCI work to examine each step.Source: §2 · pp. 4–5, §2.2 · p. 5

2.1 AI Alignment

• The paper focuses on end-users pursuing specific goals and limits its scope to alignment needs within the basic interaction cycle.It excludes broader topics such as AI safety, specification gaming, and malicious or deceptive AI.Source: §2.1 · p. 5

2.2.1 Alignment and User Input (Step 1).

• Users may struggle to provide complete, correct specifications because natural-language requests can be ambiguous, underspecified, or incorrect.This specification problem motivates interactive mechanisms that help users align AI systems to their goals.Source: §2.2.1 · p. 6
• The paper emphasizes interactive alignment during application use rather than only noninteractive model-alignment approaches such as instruction tuning and RLHF.It also notes that users can still find it challenging to achieve desired outcomes with natural-language input.Source: §2.2.1 · p. 6

2.2.2 Alignment and System Execution (Step 2).

• Users may want more control over AI execution, transparency into its process, and agency over outcomes.The section connects these concerns to process alignment and the balance between automation and augmentation.Source: §2.2.2 · pp. 6–7

2.2.3 Alignment and Evaluation (Step 3).

• Output evaluation includes verifying that an AI fulfills user intent and understanding its output, both of which can be difficult or time-consuming.Verification may be infeasible or require extensive work, while large output volumes can overwhelm users; explainability addresses why a system produced an output, a related but distinct concern.Source: §2.2.3 · pp. 7–8

3.1 Specification Alignment

• Specification alignment seeks sufficient agreement that the AI has correctly interpreted the user’s intent.The user may provide a specification or the system may infer one, which the user can validate and refine.Source: §3.1 · p. 8

3.1.1 When Specification Alignment Can Occur.

• Specification alignment can occur before, during, or after the AI acts, but costly, safety-relevant, or irreversible actions call for prospective verification.Later alignment can be efficient when the interpretation is correct or editing the output is as efficient as editing the specification.Source: §3.1.1 · pp. 8–9

3.1.2 Specification Refinement and Clarification.

• When specifications are ambiguous or underspecified, systems can request clarification before or after partially or fully acting on a request.One example is generating code completions and then asking which best matches the user’s intent.Source: §3.1.2 · p. 9

3.2 Process Alignment

• Process alignment gives users direct control or influence over how the AI achieves an outcome, and makes process observation, auditing, or understanding valuable.Examples include changing prompts during generation or specifying corpora, models, APIs, resources, or where computation occurs.Source: §3.2 · p. 9

3.2.1 Post-Process Alignment and Surrogate Processes.

• A surrogate process is a simplified, controllable representation reverse-engineered from the AI’s actual process.In the code-synthesis example, users edit pseudocode that is then used to regenerate code.Source: §3.2.1 · pp. 9–10

3.3 Evaluation Alignment

• Evaluation alignment helps users efficiently and effectively assess AI output against their goals.Support may be personalized to users’ needs and capabilities.Source: §3.3 · p. 10
• Verification support helps users check whether output meets their objectives, while comprehension support helps them understand the output.Comprehension is a stronger requirement, particularly for systems that rival or exceed users’ capabilities or expertise.Source: §3.3 · p. 10

4 Case Studies

• The case studies use the alignment dimensions to describe existing interfaces and suggest design or research opportunities.They also compare the user experience of systems with and without alignment mechanisms.Source: §4 · p. 10

4.1.1 Midjourney.

• Midjourney supports post hoc specification refinement through image variations and prompt-shortening, while its “chaos” parameter offers noninteractive process control without feedforward about its effects.The authors suggest linking prompt words to image regions as a possible evaluation aid.Source: §4.1.1 · p. 11

4.1.2 PromptPaint.

• PromptPaint blends multiple prompts for specification alignment and lets users swap prompts during image generation for process alignment.The authors suggest comparing blended output with images from each unblended prompt to support verification and debugging.Source: §4.1.2 · p. 11

4.2.1 CoPilot.

• CoPilot’s specification can be explicit in comments, documentation, or pseudocode, and implicit in context from which suggestions are inferred.It also offers blocking of suggestions matching public code and can support evaluation through comments or unit tests.Source: §4.2.1 · p. 12

4.2.2 Liu et al.’s Spreadsheet Code Synthesis.

• Liu et al.’s system transforms generated code into editable pseudocode, which can reflect the request, expose a controllable process, and help users understand the output.In a user study, participants found the pseudocode useful for understanding the code and helpful for debugging.Source: §4.2.2 · p. 12

5 Discussion

• The discussion relates the alignment dimensions to Norman’s original Gulfs of Execution and Evaluation.Source: §5 · p. 12
• It considers challenges and opportunities for making use of the alignment dimensions.Source: §5 · p. 12
• It also discusses promising areas for future work.Source: §5 · p. 12

5.1.1 Specification Alignment →Specification Gulf.

• Specification alignment covers both communicating user intent to AI and communicating AI capabilities and limitations to users.Unlike traditional input that invokes an operation, users describe the desired outcome.Source: §5.1.1 · p. 13

5.1.2 Process Alignment →Process Gulf.

• Process alignment concerns differences in how humans and AI achieve outcomes and in the representations each uses, creating a need to understand, modify, or control the AI’s process.A diffusion model’s transformation of statistical noise into an image illustrates a process unfamiliar to most people.Source: §5.1.2 · p. 13
• The Process Gulf spans execution and evaluation: controlling the AI’s process can be necessary, and auditing it can help users understand its output.Designers can use this observation to address user needs.Source: Figure 2 · p. 14, §5.1.2 · p. 14

5.1.3 Evaluation Alignment →Validation Gulf.

• The Validation Gulf is the challenge of verifying and understanding AI output, especially when it exceeds user expertise or is poorly suited to users’ perceptual or cognitive capabilities.An AI may produce hundreds of lines of code in seconds, even when the user can review code but another output format could better support evaluation.Source: §5.1.3 · p. 14

5.1.4 Strategies for Bridging Alignment Gulfs.

• Gulfs can be addressed through model improvements or interaction design, including instruction tuning and systems that ask users clarifying questions.Instruction tuning addressed instruction-following limitations in early LLMs; interactive clarification is another approach.Source: §5.1.4 · p. 15
• The three alignment gulfs may be impossible to bridge fully for sufficiently advanced AI, motivating research on bridging them and practical work to reduce them.The paper identifies capability disclosure, interpretability, and validation as possible barriers.Source: §5.1.4 · p. 15

5.2 Appropriate Levels of Alignment: Helping, Not Hindering Task Progression

• When an incorrect action has negligible cost, reaching complete alignment before acting may be cumbersome or impractical.For code synthesis, generating code and then refining the request, code, or pseudocode may be more efficient.Source: §5.2 · p. 15
• CoPilot lets users dismiss or ignore irrelevant or unneeded code suggestions as part of an iterative refinement strategy.Users can refine the initial request, resulting code, or a representation such as pseudocode.Source: §5.2 · p. 15
• Interfaces should let users reach appropriate alignment without negatively affecting user experience or introducing potential negative side effects.CoPilot and Liu et al.’s code synthesis system balance progress with facilities for corrections.Source: §5.2 · p. 15

5.3 The Challenge of Supporting Evaluation

• As AI capabilities increase, users’ need to verify and understand AI output also increases.This need has also been a focus in AI alignment research on training aligned AI.Source: §5.3 · p. 15
• Prior approaches include self-debate, recursive reward modeling that trains agents to help evaluate facets of outputs, and process supervision for chain-of-thought reasoning.Recursive reward modeling can ask people to assess summaries of character arcs rather than an entire novel [27]; process supervision can lead to better output [29].Source: §5.3 · p. 16
• Training methods can inform interactive evaluation support by producing outputs that are easier to assess or exposing reasoning steps to users.Self-debate and recursive reward modeling can yield more evaluable outputs, while reliable chain-of-thought can be shown step by step.Source: §5.3 · p. 16
• Further research is needed to help people reliably and efficiently verify and understand AI outputs.Source: §5.3 · p. 16

5.4 Alignment with Multiple Parties: CSCW and AI Alignment

• Multi-party AI interaction adds the need to attain alignment across all involved parties while retaining the same alignment goals.A study of AI introduced into two-person music creation illustrates how collaboration can change interaction requirements.Source: §5.4 · p. 16
• Groups could develop specifications jointly with an AI, while the AI could help individuals understand outputs through representations suited to their expertise.Specification alignment may occur among users as well as between users and the AI.Source: §5.4 · p. 16
• Interactive alignment for groups focuses on interface support for specification, process, and evaluation, distinct from pluralistic alignment of model values [15].Model advances for pluralistic alignment and interface advances for multi-party alignment may complement one another.Source: §5.4 · p. 16

• Specification, process, and evaluation alignment respectively concern what outcome to produce, how to produce it, and helping users verify and understand what was produced.Source: §6 · p. 17
• The alignment perspectives can describe existing interactive AI interfaces and suggest ways those interfaces could be enhanced.Source: §6 · p. 17
• The paper relates its alignment dimensions to Norman’s original Gulfs of Execution and Evaluation and considers a corresponding three-gulf model.Source: §6 · p. 17
• Future work could extend interactive alignment to scenarios with multiple simultaneous users and/or AIs.Source: §6 · p. 17