Md Atik Enam

Md Atik Enam

Human-Centered Computing Ph.D. Candidate studying accessible mobility and human-AI interaction

Making AI Roles Clear to Users

AI Roles / User Understanding / Experiment Design

I studied whether people could recognize an AI system's intended role as observer, collaborator, or leader, and why different users interpreted the same interface in different ways.

RoleLead researcher
Method3 x 3 experiment and measurement audit
Participants498 valid responses
ToolsQualtrics, Prolific, R, TypeScript
Diagram showing role assignment, interface behavior, and user recognition.

Project Objective

Designers often describe AI as an assistant, collaborator, or leader. A label alone does not establish that role. The system's actions, decision rights, and communication must support the same relationship, and users must be able to recognize it.

The project separated three questions: what role the designer assigned, what role the interface behavior communicated, and what role participants recognized.

Research Timeline

IRB (exempt)~1 month Prototype~2 weeks Recruit + collect~2 weeks Analysis~1 week Report~2 weeks

* Approximate active project time from IRB preparation through reporting.

Study Design

1. Assign rolesCreated observer, collaborator, and leader conditions.
2. Vary communicationCrossed the roles with suggestive, facilitative, and directive wording.
3. Compare recognitionAsked participants to rate all three possible roles.
4. Examine differencesUsed profile and continuous-score analyses to understand variation between users.

Method Alternatives and Choice

Alternative: Rate one assigned role

Asking only whether an interface seemed collaborative would be easy, but a high score would not show whether users saw it as more collaborative than observer-like or leader-like.

Alternative: Treat the label as success

Using the assigned role label as the outcome would confuse design intent with user recognition. It would not test whether interface behavior actually made the role clear.

Alternative: Profile analysis only

Profiles can describe differences between participants, but they cannot by themselves establish that the intended role was recognized across realistic alternatives.

Chosen: Comparative role audit

This was the best fit because each participant rated all three roles. Direct contrasts, sensitivity checks, and profile analysis separated role recognition from individual variation in interpretation.

What Role Legibility Requires

Clear actions

The AI must act in a way that matches the role being claimed.

Matching decision rights

The interface must make clear who can decide, act, approve, or override.

Comparative recognition

Users should rate the intended role above realistic alternatives.

Recourse

Users need a way to respond when the AI's actions do not match their expectations.

Study Figures

Conceptual model connecting AI role, communication, authority, and user recognition.
The conceptual role-legibility model used to separate designed role from perceived role.
Examples of interface messages used to represent different AI roles.
Message and interface examples used to operationalize observer, collaborator, and leader roles.
Results showing how participants rated the different AI role conditions.
Role-rating results used for direct recognition tests and sensitivity analysis.

Evidence

9messages across role and communication conditions.
498valid participant responses.
3roles rated for every message.
1message clearly recognized as its assigned role after correction.

The results showed that differences in user ratings do not automatically mean the intended role was established. A successful role claim needs direct evidence that users recognized the intended role over the alternatives.

Key Insights

Role clarity

Role labels were not enough.

Some messages contained actions or decision rights that did not match the assigned role.

Recognition

Comparative ratings gave stronger evidence.

Testing the intended role against alternatives was more useful than asking about one role alone.

User differences

People interpreted the same interface differently.

Perception profiles helped describe variation, but they did not prove that fixed user types existed.

Measurement

Design claims and user evidence must stay separate.

An assigned condition, a perceived difference, and successful role recognition are different findings.

Design Recommendations

Show the role through behavior

Use actions, timing, permissions, and control options that support the intended role.

Test realistic alternatives

Ask whether users recognize the intended role more strongly than other plausible roles.

Check role and authority together

Make sure communication style and decision rights tell the same story.

Provide recovery options

Let users question, correct, or override the AI when its role becomes unclear.

Broader Product Relevance

Although the experiment used an autonomous-vehicle setting, the same questions apply to AI copilots, assistants, recommendation tools, and decision-support systems. Product teams should test whether users understand what the AI is responsible for and what remains under human control.

Dissertation Study 2 of 3. Part of Understanding Human-AI Relationships in Autonomous Vehicles.