For decades, human-computer interaction has focused on how people interact with software: clicking buttons, navigating interfaces, entering information, and completing tasks through systems with relatively predictable behavior.
Artificial intelligence is changing that relationship.
An AI system can interpret what a person means, generate content, make recommendations, adapt to context, retrieve information, and increasingly take actions on a user’s behalf. For HCI, the important change is that system behavior itself has become part of the interaction to design.
This is where Human-AI Interaction (HAI) becomes important.
HAI builds on the foundations of human-computer interaction (HCI), while addressing interaction challenges that become particularly important when software can generate, predict, adapt, and act.
The central question is no longer simply:
How should people use an interface?
It is increasingly:
How should people and AI systems work together?
This is at the heart of Human-AI Interaction.
From Human-Computer Interaction to Human-AI Interaction
Traditional software generally follows a relatively explicit interaction model.
A user performs an action, the system processes it, and the interface produces a response.
For example:
User clicks “Save” → application saves the document → interface confirms the action.
AI systems can behave differently.
Consider asking an AI assistant:
“Summarize this document and tell me what I should pay attention to.”
The system has to interpret the request, determine what information may be relevant, and generate a response. The result may vary depending on the input, context, model, and information available to the system.
The user is no longer simply operating a predefined function.
The user is communicating an intention to a system that interprets and generates a response.
That difference creates new UX questions.
What does the system think the user means? How should it communicate uncertainty? What happens when its answer is wrong? How can the user correct it? When should the AI make a suggestion, and when should it wait for the user’s decision?
These are HAI questions.
AI Changes the Interaction, Not Just the Interface
It can be tempting to think of AI as another feature that can simply be added to an existing product.
A product might add an AI assistant, an automatic recommendation, or a generative feature.
But AI can change the underlying interaction model.
Imagine a writing application.
Without AI, a user might select a paragraph and choose an editing command. With generative AI, the user might instead write:
“Make this clearer and more concise for a technical audience.”
The user is communicating an intention, rather than selecting a predefined operation.
This shift affects the entire experience.
The interface is therefore only one part of the interaction design problem.
The behavior of the AI becomes part of the user experience.
Four Characteristics That Make HAI Different
Although AI systems vary considerably, several characteristics create recurring HAI challenges.
- Uncertainty
An AI system may produce an answer that sounds convincing but is incorrect.
This means a polished interface does not necessarily produce a reliable user experience.
Designers may need to help users understand uncertainty.
The goal is not necessarily to make users distrust AI. It is to help them develop a well-calibrated understanding of what the system can and cannot do.
- Adaptability
Traditional interfaces often provide relatively stable responses to the same actions.
AI systems can be more context-dependent.
The response may depend on the user’s wording, previous conversation, available information, model behavior, or other system conditions.
This adaptability can make AI more useful, but it also means that users may need clearer feedback about what the system is doing and why.
- Intention-based interaction
AI makes it increasingly possible for people to describe what they want rather than specify every step required to achieve it.
Instead of navigating a sequence of controls, a user might describe a goal and let the system determine how to help.
This can reduce interaction effort, but it also changes what the designer needs to control:
How do we design an interaction when the user and the system may interpret an intention differently?
- Increasing AI initiative
AI systems can move from answering questions to recommending actions, using tools, and performing multi-step tasks.
An assistant might suggest an action, while a more autonomous system might prepare or execute it.
As AI takes more initiative, questions about permission, oversight, control, and recovery become increasingly important.
Who Should Be in Control?
One of the central questions in HAI is the appropriate balance between human control and AI autonomy.
Imagine an AI assistant that can schedule meetings.
It could range from suggesting a suitable time to preparing or automatically booking a meeting.
These are different interaction models, even if they use similar underlying AI capabilities.
The right level of autonomy depends on the task and its consequences.
For a low-risk task, automation may be convenient.
For a consequential action, users may expect explicit confirmation.
The objective is therefore not simply to maximize automation.
It is to create appropriate human control.
Trust Is More Than Making AI Look Reliable
Trust is another important HAI issue.
A system that users completely distrust may not be useful. But a system that encourages people to accept every AI output without verification can create different problems.
The exact approach depends on the context.
An AI brainstorming tool does not necessarily require the same safeguards as an AI system involved in a high-consequence decision.
This is why HAI cannot be reduced to a collection of universal interface patterns. The interaction needs to reflect the capabilities, limitations, and consequences of the particular AI system.
Why AI Technology Matters to HCI
Understanding AI interaction increasingly requires some understanding of the technology behind it.
An interface may look like a simple chat window, but the system behind it could involve a language model, retrieval system, external tools, application data, memory, and other components.
Each can affect the user’s experience.
For example:
-
- A model’s capabilities affect what the system can generate.
- Context affects what information the AI can consider.
- Retrieval can determine what external information is available.
- Tool use can allow an AI system to perform actions.
- Model inference affects responsiveness and latency.
- System design affects how failures are handled.
This does not mean every HCI professional needs to become a machine-learning engineer.
It means that technology and interaction can no longer be treated as completely separate layers.
A UX decision may depend on how the AI works.
And a technical decision may ultimately change what the user experiences.
That connection is one of the most interesting aspects of modern HAI.
From Interfaces to Human-AI Relationships
As AI becomes more capable, the object being designed is increasingly larger than the interface itself.
A designer may need to consider questions that concern the relationship between the user and the AI system.
For example, consider an AI research assistant.
It might identify themes in interview transcripts and suggest potential insights. But a researcher may need to inspect the evidence behind those suggestions, correct an incorrect interpretation, and make the final judgment.
The AI contributes capabilities such as pattern finding and synthesis.
The human contributes context, domain knowledge, judgment, and responsibility.
The goal is not necessarily to make one replace the other.
HAI Is Both a Design and Research Problem
This shift also changes UX research.
Traditional usability research might ask:
- Can users complete the task?
- How long does it take?
- Where do they encounter difficulties?
- Do they understand the interface?
These questions remain important.
But AI introduces additional questions.
Researchers therefore need to consider both human behavior and system behavior.
An AI system may produce different outputs for different inputs, encounter unexpected situations, or change as the underlying system evolves.
This makes HAI research particularly interesting: the experience is shaped by the interaction between a person and a system that may itself be probabilistic and adaptive.
Where Human-AI Interaction Is Going
The familiar chatbot is only one form of AI interaction.
AI systems are increasingly working with multiple types of information and interaction, including text, images, audio, and other modalities. They can retrieve information, use external tools, maintain context, and perform increasingly complex tasks.
This creates opportunities for new interaction models:
- Conversational interfaces
- Multimodal interaction
- AI-assisted workflows
- Adaptive interfaces
- Generative interfaces
- AI recommendations
- Agentic systems
- Human-AI collaboration
As these systems develop, the boundaries between UX design, HCI research, AI technology, and product engineering become increasingly connected.
That is why understanding AI interaction requires more than designing the visible interface.
It requires understanding the relationship between people, interfaces, AI behavior, and the technology underneath.
Conclusion
Human-AI Interaction is not simply traditional UX with an AI feature added.
AI changes the nature of interaction itself.
People may communicate intentions instead of commands. Systems may generate information instead of simply displaying it. AI may recommend, adapt, retrieve, collaborate, and increasingly take action.
These capabilities create new questions about trust, uncertainty, mental models, control, autonomy, and collaboration.
The most useful question may therefore not be:
“How do we design interfaces for AI?”
but:
“How should we design the relationship between people and intelligent systems?”
That is the perspective of Human-AI Interaction—and it is where the next generation of HCI increasingly begins.