Human–AI Collaboration

I designed and conducted a controlled UX research study exploring how people experience collaboration with an AI agent under time pressure. Participants first played a game independently and then advised a reinforcement-learning agent during gameplay. By comparing timed and untimed interactions, we examined how urgency influenced trust, perceived usability, confidence, and the user’s sense of agency. The findings were published and presented at HumanAISE, co-located with FSE 2026.

2025–2026

Live Website->

Overview

As AI systems become more proactive, they increasingly interrupt users to request guidance, approval, or feedback. These interactions may appear efficient, but they can also create pressure and change how people perceive both the AI and their own abilities.

This project investigated the human experience of collaborating with a reinforcement-learning agent in a Pac-Man environment. Participants first completed a human-only task and then worked with an AI agent that periodically asked them for directional advice. In one condition, participants could respond at their own pace. In the other, they had only five seconds before the AI disregarded their input.

The study focused not only on whether users trusted the AI, but also on whether the interaction remained understandable, usable, empowering, and psychologically sustainable. The central UX question was: Can an AI system become more trusted while simultaneously making the human collaborator feel less confident?

The problem

Proactive AI systems are increasingly designed to interrupt users and request decisions in real time. These prompts are often treated as functional interface events, but they can also create cognitive pressure, reduce a person’s sense of control, and influence how users evaluate their own abilities.

Most research on human–AI collaboration focuses on trust in the system, accuracy, explainability, or task performance. Less attention is given to the emotional and psychological experience of being prompted by an AI under urgency.

We wanted to understand whether a timed AI request changes the relationship between the user and the system. Specifically, we examined whether urgency affects:

trust in the AI;
perceived ease of use;
predictability and dependability;
understanding of the AI’s behaviour;
willingness to remain involved in the collaboration;
and confidence in one’s own contribution.

The design challenge was to create an interaction that felt meaningful enough for participants to care about the outcome, while still allowing us to isolate the effect of urgency.

The study was motivated by the gap between research on trust and research on external pressures such as time constraints in collaborative AI systems.

Research and discovery

We conducted an in-person controlled study with 30 senior undergraduate and graduate computing students.

Each session included two main phases:

1. Human-only interaction
Participants played a simple arcade game independently for five minutes. This established a baseline experience and allowed us to measure their general attitudes toward AI before collaboration.

2. Human–AI collaboration
A reinforcement-learning agent took control of the game and periodically asked the participant for directional advice. Participants experienced two modes:

Low Urgency : they could respond without a deadline.
High Urgency: they had five seconds to respond before their advice was ignored.

We counterbalanced the order of these modes to determine whether previous exposure to a calmer interaction changed the experience of urgency.

Before and after the interaction, participants completed questionnaires measuring trust, ease of use, dependability, predictability, goal alignment, transparency, and self-confidence. We also recorded behavioural and performance data and compared responses across the two sequence groups.

The solution

We created a controlled human-AI collaboration experience in which participants could directly contribute to an AI agent’s decisions while experiencing two different levels of urgency.

The final interaction included:

a familiar game environment;
a visible transition from human control to AI control;
periodic AI requests for directional advice;
a no-pressure response mode;
a five-second response mode;
pre- and post-interaction trust measures;
and a counterbalanced sequence to evaluate the effect of onboarding.

The resulting design recommendations were:

Introduce collaboration gradually.
Users should first be allowed to understand the AI’s behaviour without immediate time pressure.

Preserve the user’s sense of agency.
Interfaces should clearly communicate how human input affects the AI’s decisions.

Avoid using urgency as a default interaction pattern.
Timers, countdowns, and disappearing requests may increase pressure even when trust improves.

Measure human confidence, not only trust in the AI.
A system can appear successful while still reducing the user’s belief in their own contribution.

Treat onboarding as part of the user experience.
Familiarity with the AI changed how participants responded to later urgency.

Reflection

This project changed how I think about successful Human-AI interaction.

At the beginning, trust in the AI seemed like the most important outcome. During the analysis, however, it became clear that increased trust does not automatically mean that the experience is healthy or empowering. A user may become more confident in the system while becoming less confident in their own role.

The strongest lesson was the importance of onboarding. Participants who first had time to observe and understand the AI were more resilient when urgency was later introduced. This suggests that onboarding is not simply an instructional screen or tutorial; it shapes confidence, agency, and the long-term relationship between the user and the system.

I also learned the value of measuring human outcomes beyond task performance. Accuracy, completion time, and system trust would not have revealed the full experience. Including self-confidence allowed us to identify a hidden cost of the interaction.

In a future iteration, I would include a validated self-efficacy scale, collect richer qualitative feedback immediately after each condition, and prototype alternative urgency patterns such as adjustable response windows, pause controls, confidence indicators, and explanations of how the AI uses human advice.

This work reinforced my interest in designing AI systems that do not merely perform well, but also help people remain informed, capable, and confident.