Guidelines for Human-AI Interaction
Eighteen best practices for human-centered AI design
By Mihaela Vorvoreanu, Saleema Amershi, and Penny Collisson
Today we’re excited to share a set of Guidelines for Human-AI Interaction. These 18 guidelines can help you design AI systems and features that are more human-centered. Based on more than two decades of thinking and research, they have been validated through a rigorous study published in CHI 2019.
Why do we need guidelines for human-AI interaction?
While classic interaction guidelines hold with AI systems, attributes of AI services, including their accuracy, failure modalities, and understandability raise new challenges and opportunities. Consistency, for example, is a classic design guideline that advocates for predictable behaviors and minimizing unexpected changes. AI components, however, can be inconsistent because they may learn and adapt over time.
We need updated guidance on designing interactions with AI services that provide meaningful experiences, keeping the user in control and respecting users’ values, goals, and attention.
Why these guidelines?
AI-focused design guidance is blooming across UX conferences, the tech press, and within individual design teams. That’s exciting, but it can be hard to know where to start. We wanted to help with that, so…
- We didn’t just make these up! They come from more than 20 years of work. We read numerous research papers, magazine articles, and blog posts. We synthesized a great deal of knowledge acquired across the design community into a set of guidelines that apply to a wide range of AI products, are specific, and are observable at the UI level.
- We validated the guidelines through rigorous research. We tested the guidelines through three rounds of validation with UX and HCI experts. Based on their feedback, we iterated the guidelines until experts confirmed that they were clear and specific.
Let’s dive into the guidelines!
The guidelines are grouped into four categories that indicate when during a user’s interactions they apply: upon initial engagement with the system, during interaction, when the AI service guesses wrong, and over time.
1. Make clear what the system can do.
2. Make clear how well the system can do what it can do.
The guidelines in the first group are about setting expectations: What are the AI’s capabilities? What level of quality or error can a user expect? Over-promising can hurt perceptions of the AI service.
PowerPoint’s QuickStarter illustrates Guideline 1, Make clear what the system can do. QuickStarter is a feature that helps you build an outline. Notice how QuickStarter provides explanatory text and suggested topics that help you understand the feature’s capabilities.
3. Time services based on context.
4. Show contextually relevant information.
5. Match relevant social norms.
6. Mitigate social biases.
This subset of guidelines is about context. Whether it’s the larger social and cultural context or the local context of a user’s setting, current task, and attention, AI systems should take context into consideration.
AI systems make inferences about people and their needs, and those depend on context. When AI systems take proactive action, it’s important for them to behave in socially acceptable ways. To apply Guidelines 5 and 6 effectively, ensure your team has enough diversity to cover each other’s blind spots.
Acronyms in Word highlights Guideline 4, Show contextually relevant information. It displays the meaning of abbreviations employed in your own work environment relative to the current open document.
7. Support efficient invocation.
8. Support efficient dismissal.
9. Support efficient correction.
10. Scope services when in doubt.
11. Make clear why the system did what it did.
Most AI services have some rate of failure. The guidelines in this group recommend how an AI system should behave when it is wrong or uncertain, which will inevitably happen.
The system might not trigger when expected, or might trigger at the wrong time, so it should be easy to invoke (Guideline 7) and dismiss (Guideline 8). When the system is wrong, it should be easy to correct it (Guideline 9), and when it is uncertain, Guideline 10 suggests building in techniques for helping the user complete the task on their own. For example, the AI system can gracefully fade out, or ask the user for clarification.
Auto Alt Text automatically generates alt text for photographs by using intelligent services in the cloud. It illustrates Guideline 9, Support efficient correction, because automatic descriptions can be easily modified by clicking the Alt Text button in the ribbon.
12. Remember recent interactions.
13. Learn from user behavior.
14. Update and adapt cautiously.
15. Encourage granular feedback.
16. Convey the consequences of user actions.
17. Provide global controls.
18. Notify users about changes.
The guidelines in this group remind us that AI systems are like getting a new puppy: they are long-term investments and need careful planning so they can learn and improve over time. Learning (Guideline 13) also means that AI systems change over time. Changes need to be managed cautiously so the system doesn’t become unpredictable (Guideline 14). You can help users manage inherent consistencies in system behavior by notifying them about changes (Guideline 18).
Ideas in Excel empowers users to understand their data through high-level visual summaries, trends, and patterns. It encourages granular feedback (Guideline 15) on each suggestion by asking, “Is this helpful?”
If you’d like some more ideas, stay tuned for another post on this work where we share some of the uses we’ve been working with at Microsoft. We’d love to hear about your experiences with the guidelines. Please share them in comments.
Mihaela Vorvoreanu is a program manager working on human-AI interaction at Microsoft Research.
Saleema Amershi is a researcher working on human-AI interaction at Microsoft Research AI.
Penny Collisson is a user research manager working on AI in Office.
With thanks to our team who developed The Guidelines for Human-AI Interaction: Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Paul Bennett, Kori Inkpen, Jaime Teevan, Ruth Kikin-Gil, and Eric Horvitz.
Thanks also to Ruth Kikin-Gil for her thoughtful collaboration, and for curating examples for this post.