39 Principles for Reinventing Human-AI Interaction

A Framework for Human-AI Interaction

In the ever-evolving world of technology, the design of artificial intelligence (AI) interfaces must meet specific requirements to ensure effective and secure interaction between humans and machines. Unlike traditional interfaces, where each action has a predefined function and errors can be anticipated, AI systems are inherently less predictable. They introduce variability where the same input can yield different results, making product quality dependent not only on the model but also on the interface and the instructions surrounding it.

The challenges posed by AI in terms of interaction go beyond aesthetics. They touch on the very ability of users to evaluate results, correct errors, and remain accountable for decisions made. The crucial question is how to help users trust AI appropriately. This article aims to transform complex research into concrete design principles, drawing on fields such as human-AI interaction, mixed-initiative systems, and responsible AI.

Theoretical Foundations of Principles

This design framework is based on established research in several key areas. Among them are boundary model specifications, such as OpenAI's Model Spec and Anthropic's Claude Constitution, which emphasize the importance of honesty about the nature of the system and the necessity of maintaining human oversight. These principles directly influence new design guidelines.

Mixed-initiative interaction, studied by Eric Horvitz, provides a foundation for deciding when an AI system should act or suggest, treating interaction as a shared decision between human and machine. Practical guides like Google's PAIR Guidebook and IBM's generative AI principles provide applied layers for designing systems that meet user needs while accounting for the variability and imperfection inherent in AI.

Microsoft, with its 18 guidelines for human-AI interaction, organizes these principles into four phases, emphasizing error management, which is not a marginal case but a frequent occurrence in AI.

A Practical Design Framework

Interfaces must clarify the role of the system, help users understand and verify results, preserve user control, and support error correction. Here are the key elements of this framework:

• Probabilistic foundation: Design for inference and variability.
• Expectation setting: Clarify the capabilities and limits of AI from the outset.
• Calibrated trust: Align user dependence with system reliability.
• Transparency: Make reasoning and evidence inspectable.
• Control and autonomy: Facilitate acceptance, rejection, editing, and cancellation.
• Graceful failure: Make uncertainty and errors recoverable.
• Co-creation: Treat output as a draft.
• Responsible autonomy: Constrain action by stakes and reversibility.
• Sustained dependence: Manage quality and change over time.

Probabilistic Foundation

AI models behave more like probabilistic services than fixed functions. They produce a distribution of outcomes rather than a single answer. It is crucial to design for this dispersion.

1. Use AI where it excels: AI is particularly effective when inputs are messy or ambiguous, or when large amounts of information need to be synthesized. It is less effective for tasks requiring precision and repeatability.

   For example, Linear uses AI to summarize long discussion threads or identify related work while maintaining a deterministic interface for tasks requiring precision, such as managing statuses or permissions.

2. Design for generative variability: The same prompt can produce multiple acceptable responses. This variation is often an added value rather than a flaw.

   The interface should allow users to work with this variation by offering multiple drafts or regeneration actions. In exploratory tasks like writing or design, showing several options can be more helpful than a single answer.

   Midjourney, for instance, presents a grid of four images per prompt, turning variance into choice rather than error.

3. Adapt the interaction model to the task: Not every AI feature needs to be a chatbot. The interaction model should match the type of input and the consequences of the output.

   For simple tasks, AI can appear as an integrated suggestion. For more complex tasks, the system should include plans and checkpoints.

   Notion, for example, integrates quick rewrites while directing open questions to a conversational AI panel.

Expectation Setting

Users form expectations even before reading the first output from the AI.

1. Indicate the system's capabilities and limits: Users need to know not only what the system can do but also where its limits lie.

   ChatGPT, for example, presents itself as a versatile assistant but must clearly signal when it cannot access certain information.

2. Address the blank canvas problem: An empty prompt box does not guide users. Guides like prompt examples or templates can help express intent more clearly.

   Canva Magic Design, for instance, offers ready-made templates as soon as you describe a project.

3. Frame the output as a starting point: The presentation of the output influences how it is perceived. Terms like "draft" or "suggestion" indicate that the output should be inspected.

   Gmail, with its Gemini feature, presents generated emails as "drafts," encouraging users to revise them before sending.

4. Signal the role of AI: Users should know when content is generated by AI. This avoids issues of attribution and accountability.

   Meta, for example, labels AI-generated content on its platforms to clarify its origin.

5. Tailor the explanation to the user's expertise: Transparency and control should be adapted to the user's skill level.

   GitHub Copilot offers different configuration levels depending on whether the user is a novice or an expert.

6. Honestly represent the nature of AI: Avoid over-humanizing the system to prevent unrealistic expectations.

   Ryanair, for example, presents its AI travel assistant with a human-like personality, which can blur accountability and user expectations.

Calibrated Trust

Calibrated trust means that user dependence should correspond to the actual reliability of the system. The interface should reduce inappropriate forms of dependence by providing clear information about the capabilities and limits of AI.

To ensure that users do not blindly rely on AI, it is essential to calibrate trust. This involves ensuring that user dependence is proportional to the system's reliability in the task at hand. The interface should thus provide clear indications of the system's capabilities and limits, allowing users to adjust their level of trust accordingly.

Transparency

Transparency is a key element in establishing a trusting relationship between the user and the AI system. Users must be able to understand how and why the system makes certain decisions. This requires making the system's reasoning and underlying evidence inspectable when necessary.

Control and Autonomy

Users must have the ability to control the actions of the AI and make final decisions. This means they should be able to easily accept, reject, edit, cancel, or substitute AI suggestions. A good AI system should offer users a level of autonomy that allows them to remain in control of their decisions.

Graceful Failure

AI systems must be designed to handle errors gracefully. This involves making uncertainty and errors recoverable so that users can correct issues without too much difficulty. A well-designed system should allow users to backtrack and correct mistakes without wasting time or effort.

Co-Creation

AI should be viewed as a partner in the creation process rather than as an absolute source of truth. Outputs generated by AI should be treated as drafts or suggestions, allowing users to modify and adapt them to their specific needs. This encourages a collaborative approach where the user and AI work together to achieve a common goal.

Responsible Autonomy

AI autonomy should be constrained by stakes, reversibility, and permission. This means that AI should not make significant decisions without the explicit agreement of the user. Systems should be designed to respect the preferences and limits set by users, ensuring that AI acts responsibly.

Sustained Dependence

AI systems must be designed to manage quality, drift, ownership, and change over time. This involves continuously monitoring and adjusting the system to ensure it remains reliable and relevant for users. A good AI system should be capable of adapting to evolving user needs and expectations, thereby ensuring sustained dependence and ongoing use.

In conclusion, designing effective AI interfaces requires a deep understanding of user needs and the challenges posed by AI systems. By following these principles, designers can create systems that foster trust, transparency, and user control, while enabling a harmonious and productive human-machine interaction.