AI Glossary / AI and machine learning / Beginner

Human - AI interaction

Definition

Human - AI interaction is the study and design of how people communicate, collaborate and make decisions with AI systems. It includes the interface, the surrounding workflow and the way people respond to the system's capabilities, uncertainty and mistakes.

What it means

Interaction can involve a conversation, a recommendation, an automatically sorted list or a robot's behavior. It does not require a chatbot. Any situation where people act on, guide or respond to AI output raises interaction questions.

AI adds complications to ordinary interface design because its outputs can vary, be uncertain and change as the system or context changes. Users need workable expectations about what the system can do and when its output needs checking.

The field examines both immediate actions and longer-term relationships. A single correction matters, but so does whether people learn to rely on the system appropriately after repeated successes and failures.

Good interaction is not simply pleasant wording or a humanlike personality. It should help people accomplish their task, understand consequential actions and recover from problems without unnecessary effort or confusion.

How it works

Identify the user's task and the points where AI participates. Specify what information the person provides, what the system returns and what action follows. Include handovers to other people or software.

Design the exchange around real capabilities. Make important limitations understandable in context, and distinguish suggestions from completed actions. If an input is ambiguous, choose an appropriate way to clarify or limit the response.

Provide effective ways to inspect, correct, dismiss or redirect the system. A correction should have a clear consequence. Users should not have to guess whether feedback edits one result, changes future behavior or triggers an external action.

Evaluate with representative tasks and users. Observe errors, recovery time, interruptions and inappropriate reliance as well as satisfaction. Continue testing after changes because altered model behavior can invalidate expectations established by the earlier design.

Three practical examples

1. An email assistant distinguishes a suggested draft from a message already sent. The user can edit the draft and sees clearly when an action will reach a recipient.

2. A recommendation interface lets shoppers explain that an item is irrelevant for a specific reason. The design shows whether that feedback affects only the current suggestion or future recommendations.

3. A scheduling assistant presents an ambiguous date for clarification before acting. The interaction resolves the uncertainty at the point where a mistaken interpretation would affect the calendar.

Think of it this way

Think of working with a new colleague whose abilities you are still learning. Clear requests, understandable responses and useful corrections help coordination. The analogy is limited because AI does not necessarily share human understanding, intentions or accountability, and a friendly conversational style should not be mistaken for those qualities.

Worked example

Consider two ways to correct an AI-generated address. Interface A requires opening a separate settings page, finding the record and reentering the entire address. Interface B allows editing the incorrect field beside the source document and explains that the change affects this record only.

In an invented usability exercise, five users each correct two addresses. That creates ten correction attempts. If eight attempts succeed without help in A and all ten succeed in B, the observed unassisted completion rates are 80 percent and 100 percent.

The small exercise does not prove a general performance advantage. It identifies an interaction difference worth investigating with more representative tasks, including accessibility needs and cases where the underlying source is itself unclear.

Illustrative case study

Fictional case: Willow Desk builds an assistant that summarizes customer requests. Staff initially assume that correcting a summary teaches the system a permanent preference, although the correction only changes the current record.

The product team clarifies the feedback controls and tests whether users can predict their consequences. It also provides a way to inspect source text when a summary seems incomplete.

The team evaluates fewer misunderstandings and easier recovery as interaction outcomes, separately from the summarization model's accuracy. This fictional case illustrates the importance of shared expectations, not a real company's measured result.

Common misconceptions and mistakes

Mistake: Equating human - AI interaction with prompt writing. Prompts are one input mechanism within a larger process involving outputs, actions, feedback and recovery.

Mistake: Assuming humanlike language guarantees usability. It can make an interface approachable while also encouraging inaccurate assumptions about capability or understanding.

Mistake: Measuring only user satisfaction. People may enjoy a system while misunderstanding it or relying on incorrect outputs, so task outcomes and error behavior also matter.

Business use

Human - AI interaction affects whether an AI feature can be used effectively in everyday work. Poorly designed handovers or correction controls can consume the time that the underlying model was expected to save.

Business teams should evaluate the complete task, including review and recovery. Define which actions require explicit user intent and make the system's state visible. Treat support questions and repeated workarounds as evidence about the interaction design rather than automatically blaming users for not understanding the technology.

Check your understanding

A feedback button changes future recommendations, but users believe it only dismisses one item. What interaction problem exists?

Answer: The consequences of the action are unclear, so users cannot reliably control the system as intended.

Related terms

Human-centered AI
Feedback
Automation bias
Explainability
Usability testing
user experience
mixed-initiative interaction

Learn next

Learn usability testing and mental models. Trace one AI-assisted task and ask a user to predict what each control will do, then compare those expectations with the system's actual behavior.

Usability testing

Sources

Amershi and colleagues, Guidelines for Human-AI Interaction, primary research overview: https://www.microsoft.com/en-us/research/blog/guidelines-for-human-ai-interaction-design/. Original paper DOI: https://doi.org/10.1145/3290605.3300233. Author overview consulted; interface examples, arithmetic and fictional case are original.

Previous: Huffman encoding Next: Human-based evolutionary computation