Interactive AI is software that works with a person through repeated exchanges, where the person can start, correct, or stop what the system does. Design research gives it concrete rules: a 1999 paper by Eric Horvitz set out 12 principles for mixed-initiative interfaces, and a 2019 Microsoft study by Saleema Amershi and colleagues condensed more than 150 recommendations into 18 guidelines. Both treat the quality of the interaction, and not only the accuracy of the model, as part of whether an AI tool works.
Key takeaways
- The 1999 Horvitz paper frames the core question as when a system should act, when it should ask, and when it should stay quiet.
- The 2019 guidelines for human-AI interaction are grouped by phase: at the start, during use, when the system is wrong, and over time.
- Guidelines on correction, dismissal, and explanations matter most when the model errs, since every model does.
- The guidelines were developed on products with graphical interfaces, and the authors say voice and high-stakes uses may need extensions.
What counts as interactive AI
A spam filter that runs silently is AI, but it is hardly interactive. A writing assistant that proposes a sentence, lets the writer edit it, and adjusts to the edit is interactive. The difference lies in the loop: the person provides input, the system responds, and the person can redirect it. Researchers call the design problem mixed initiative, because either side may take the next step. The label covers chatbots, drafting tools, recommendation panels, and scheduling assistants alike, so the useful questions are about control and timing, not about the technology underneath.
Act, ask, or wait: the 1999 framing
Horvitz presented his paper at the CHI conference in Pittsburgh in 1999, using a Microsoft email tool called LookOut that spotted meeting requests and filled in a calendar entry for the user to edit. He listed key problems with agents in interfaces: poor guesses about user goals, weak accounting of the costs and benefits of acting, poor timing, and too little room for users to guide or refine results. His principles include considering uncertainty about the user's goals, weighing the user's attention before interrupting, and letting users invoke or terminate a service directly.
The paper also gives a decision rule. A system estimates the probability that the user wants help and compares the expected value of acting with the expected value of doing nothing. Adding a third option, asking a short question, produces two thresholds: below the lower one the system stays idle, between them it asks, and above the higher one it acts. The thresholds move with context, for instance when the user is deeply focused on another task and an unwanted action would cost more.
The 18 guidelines, by phase
Microsoft Research announced the guidelines on February 1, 2019, ahead of their presentation at CHI 2019 in Glasgow. The team gathered more than 150 recommendations from papers, blog posts, and internal documents, grouped them, and refined the set through several rounds of review by experts in user experience and human-computer interaction. A related Microsoft Research page reports a study in which 49 design practitioners checked the guidelines against 20 popular AI-infused products.
| Phase | Guidelines (numbers as published) |
|---|---|
| Initially | 01 Make clear what the system can do. 02 Make clear how well the system can do what it can do. |
| During interaction | 03 Time services based on context. 04 Show contextually relevant information. 05 Match relevant social norms. 06 Mitigate social biases. |
| When wrong | 07 Support efficient invocation. 08 Support efficient dismissal. 09 Support efficient correction. 10 Scope services when in doubt. 11 Make clear why the system did what it did. |
| Over time | 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 authors state that the guidelines were developed and tested on products with graphical interfaces, and that voice interaction and high-stakes settings such as semi-autonomous vehicles may need extensions. They are design guidance from one research group, not a standard or a regulation.
Turning the guidelines into a review for a school or nonprofit chatbot
The checklist below is an illustrative adaptation by this site, not part of the published guidelines. It shows how the phases translate into questions that a program manager could ask before putting a chat tool in front of students or clients.
- Initially: does the first screen say what the tool can and cannot help with, and how often it may be wrong?
- During use: does it respond at sensible moments, and does its wording suit the audience, including age and language level?
- When wrong: can a user dismiss a suggestion, edit it, or ask why it appeared, and does the tool narrow its answer when unsure?
- Over time: does it keep context within a session, are changes to its behavior announced, and can a user switch off personalization?
Why the 1999 and 2019 work still matter
The guidelines on stating limits and explaining outputs apply directly to current chat tools. The Horvitz thresholds also describe a problem anyone who has used an eager assistant recognizes: acting too often costs trust, and asking too often costs patience. Neither paper settles how much autonomy a given product deserves. That depends on the stakes, the audience, and the cost of an unwanted action, and the papers give a vocabulary for discussing those trade-offs instead of a formula.
FAQ
What is the difference between interactive AI and an ordinary AI model?
An ordinary model maps inputs to outputs. Interactive AI adds a loop in which a person can respond, correct, or redirect, so design choices such as timing, explanations, and undo become part of the system.
Who wrote the 18 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, presented at CHI 2019.
Are the guidelines legally required?
No. They are research-based design guidance. Legal duties on transparency or automated decisions come from separate laws and policies.
Do the guidelines cover voice assistants and self-driving cars?
The authors developed them on graphical-interface products and say voice and high-stakes uses may need extensions.
Related reading
- GPT-3 as a chat bot
- Teaching AI to kids through interactive activities
- A review of AI in education
- AI art and creative tools
More in AI Basics.
Last verified: October 2026
Sources
- Principles of Mixed-Initiative User Interfaces, Eric Horvitz, CHI 1999, Microsoft Research
- Guidelines for human-AI interaction design, Microsoft Research blog, February 1, 2019
- Guidelines for Human-AI Interaction, Amershi et al., CHI 2019, Microsoft Research publication page
- HAX Toolkit design library, Microsoft