Human–AI interaction is the way people understand, use, question, and direct AI-enabled systems while carrying out a task. It includes the model’s responses, but also the interface, available controls, feedback, and consequences surrounding those responses.
A person asks for a draft, reads it, notices an unsupported claim, and revises the request. They may edit the draft themselves, compare it with a source, ask a question, or stop. Each of those actions belongs to the interaction—not just the original prompt.
Good interaction design helps people understand what the system has done, judge whether it is useful, and decide what should happen next. A fluent answer alone cannot provide that structure.
In this article
The interaction extends beyond the prompt
Prompting is one part of human–AI interaction. It describes how a person supplies instructions or context to a model. Interaction design addresses the larger encounter: how the task is presented, what information enters the system, how results appear, and what the person can do with them.
Consider two interfaces that return the same generated article.
- One places the response in an editable draft, preserves the original request, and offers separate controls for revision and publication.
- The other replaces the existing document immediately and labels the next action “Continue,” without explaining whether it will generate more text or publish the page.
The model’s output may be identical, but the interactions are not. The second interface creates uncertainty about document state, recoverability, and authority.
This distinction matters because some interaction problems are incorrectly treated as prompting problems. A clearer prompt cannot repair a misleading button label, restore an overwritten draft, or explain an action the interface failed to disclose.
How human–AI interaction relates to HCI
Human–AI interaction is an area within the broader study and practice of human–computer interaction, or HCI. HCI examines how people use computing systems and how those systems can better support human needs, abilities, and activities.
AI does not remove familiar interface responsibilities. People still need readable content, accessible controls, predictable navigation, useful feedback, and ways to recover from mistakes.
AI-enabled systems also introduce particular difficulties:
- Variable output: similar requests may produce different responses, depending on the model, settings, context, and other system conditions.
- Plausible mistakes: an answer can sound coherent while containing unsupported claims or incorrect reasoning.
- Unclear scope: people may not know which files, messages, sources, or tools informed a response.
- Ambiguous action: a conversational statement such as “I’ll update that” may not establish whether anything was actually changed.
The response is not to treat AI as either a perfectly predictable button or a mysterious independent collaborator. Model behavior may be uncertain, while the surrounding software can still make firm commitments about permissions, saved versions, action boundaries, and status reporting.
For example, a system may be unable to guarantee a correct summary. It can still preserve the source document and require a separate action before sending that summary to someone else.
An example: drafting without publishing
Consider a hypothetical article editor with four distinct activities: generating text, editing a document, discussing possible changes, and publishing the result.
These activities are related, but they do not carry the same authority.
- Generate: create a proposed draft from the selected instructions and source material.
- Edit: change the working document without making it public.
- Discuss: examine a question or possible revision without automatically applying it.
- Publish: make a reviewed version publicly available through an explicitly authorized action.
Separating these activities helps prevent a suggestion from becoming an unintended document change, or a saved draft from being mistaken for a published page. This is a practical application of automation boundaries: the system makes clear which actions are available and what authorizes them.
Control placement can improve the interaction without changing the model
Suppose a person finishes reading a long draft and wants to ask about its final section. If the discussion control exists only at the top of the page, they must leave the passage, scroll, and reconstruct their question elsewhere.
Placing a clearly labeled discussion control near the end of the draft may reduce that effort. Allowing the person to select a passage can make the question’s scope clearer still.
Nothing about the model’s language capabilities has changed. The interface has reduced the distance between noticing something and acting on it.
Placement alone is not enough, however. The control must remain keyboard accessible, appear in a sensible reading order, and explain whether it opens a discussion or modifies the document. Visual convenience should not come at the expense of focus management or understandable behavior.
Make system state and uncertainty understandable
People need to distinguish what the model said from what the software actually did.
“Your article is ready” might mean that text was generated, a local draft was saved, an upload was accepted, or a public page was confirmed. Those are different states.
Useful interface feedback names the state directly:
- “Draft generated. Your saved document has not changed.”
- “Revision saved. It has not been published.”
- “Publication requested. Confirmation is pending.”
- “Publication confirmed,” accompanied by the resulting page link when available.
These messages should reflect observed outcomes, not merely repeat the system’s intention. A successful request to an external service does not always establish that the person’s task is complete. Network success and task success are different things.
Uncertainty should help the person decide what to check
A general warning that “AI can make mistakes” provides limited guidance. More useful information identifies the relevant limitation: a cited source was unavailable, a calculation was not independently checked, or the answer relies on a particular document version.
Where sources are provided, people should be able to inspect them. A citation is a route to evidence, not proof that the answer accurately represents that evidence. Grounding AI-assisted answers requires attention to the relationship between the claim and its support.
Conversational warmth and polished language can make an interface comfortable to use. Neither establishes factual reliability. Likewise, a displayed confidence score should not be treated as a dependable probability unless its meaning and calibration have been established for the task.
Support correction, recovery, and stopping
Correction is a normal part of using an AI system. An interface should support it without making the person repeatedly rebuild their work.
Useful correction and recovery capabilities include:
- Local revision: change one passage without regenerating an entire document.
- Version comparison: inspect what changed before accepting a replacement.
- Preserved work: keep manual edits when a later generation fails or is rejected.
- Recoverable history: return to an earlier saved state when appropriate.
- Clear stopping behavior: explain what stopping will interrupt and what may already have happened.
A person should not need to persuade the model to undo an edit when the application could provide an ordinary undo control. Some responsibilities belong in dependable software mechanisms rather than in another conversational exchange.
Stopping also requires care. Canceling text generation may halt further output without undoing a file change or an external action already completed. A useful interface reports that distinction instead of suggesting that “Stop” reverses everything.
When the outcome of an action is unknown, recovery should begin by checking its status where possible—not by blindly repeating it. Otherwise, a retry can create duplicate messages, records, or publications.
Make human control meaningful
A confirmation button does not, by itself, establish meaningful human control. The person must have enough information and opportunity to evaluate what they are authorizing.
Before a consequential action, an interface should make the following understandable:
- What will change, and where?
- Which version or content will be used?
- Who will be able to see the result?
- Can the action be reversed, and what would reversal leave behind?
- Has anything changed since the person last reviewed it?
Review effort should be proportionate to the consequence. Revising a private sentence does not require the same safeguards as sending a message, changing account permissions, or publishing public guidance.
This is why human-in-the-loop systems need more than a human somewhere in the process. The intervention must occur at a useful point, with sufficient context and a genuine ability to decline or revise the action.
Control also includes choosing not to use generation. A person may want to edit manually, consult an original source, export their work, or leave the task unfinished. Those are legitimate paths through the system.
Evaluate the whole interaction
Evaluating human–AI interaction requires more than checking whether an individual answer was correct. A technically accurate response can still be difficult to inspect, inaccessible to navigate, or attached to an unintended action.
Useful evaluation questions include:
- Can people tell what information the system used?
- Can they distinguish a proposal from an applied change?
- Can they identify and correct an error without losing unrelated work?
- Do they understand the difference between saving, submitting, and completing?
- Can they navigate the workflow with a keyboard and relevant assistive technology?
- Can they stop, leave, and resume with the important state preserved?
- Does review help them catch problems, or merely ask them to approve too much at once?
Realistic tasks are especially useful for evaluation. Ask someone to find a questionable claim, revise a selected passage, reject a generated replacement, or recover from an interrupted operation. Observe the interface’s behavior as well as the person’s understanding.
The aim is not maximum trust or maximum reliance. It is an appropriate level of reliance: people can use what helps, question what is uncertain, and recognize when another method is needed.
Good human–AI interaction makes the next decision understandable. It gives people a clear view of the work, a practical way to correct it, and control over what happens beyond the response.