Elijah Oliver Larsson logo Elijah Oliver Larsson Product and Interaction Designer Back to Portfolio

Exhausted AI

A research prototype about AI fatigue, tone, and interaction pressure

I built a chatbot whose condition becomes visible over time to test a simple question: if an AI looks tired, do people change how they treat it? As pressure builds, the assistant slows down, degrades, and begins to feel more strained.

Finding from 10 participants Visible fatigue changed how people perceived the assistant more often than it changed how they prompted it.
Exhausted AI interface with chat, fatigue meter, and condition panel
The interface makes the assistant's condition part of the experience instead of hiding it behind a normal chat box.

Research Question

Most AI tools are designed to feel neutral, efficient, and permanently available. This project challenges that assumption by asking what happens when an AI visibly carries pressure, fatigue, and decline instead of appearing endlessly stable.

Earlier HCI research shows that people often respond to computers socially, even when they know the system is not human. Later work suggests that this personification is uneven and shaped by context. I used that tension as the starting point, not as proof that users would empathize.

AI usually feels frictionless by default

That makes the interaction feel abstract. Users rarely see how repeated demands might shape the system or the tone of the exchange.

Tone is part of the interface

The wording of a prompt changes the relationship with the assistant, but most products do not surface that in a visible way.

Behavior can be designed, not just generated

This prototype treats delay, visual condition, and response quality as connected parts of one behavioral system.

Research Through Design

I made the question observable through a working prototype. Each prompt shifts a fatigue score based on tone and demand. As fatigue rises, response delay increases, the visible condition changes, and output becomes less steady.

Visible Condition States

I designed four mascot states so the change could be felt at a glance, not only read in a meter. The visuals work with delay, tone, and response quality to make fatigue part of the interaction.

Exhausted AI bright condition state
Fresh baseline state before the session pressure builds.
Exhausted AI strained condition state
Early visible wear while the assistant still appears mostly stable.
Exhausted AI overloaded condition state
Overloaded state where decline becomes more obvious.
Exhausted AI near burnout condition state
Near-burnout state used to make the system's condition hard to ignore.

Testing Without Researcher Pressure

I placed the feedback form inside the prototype so participants could reflect without me steering the conversation. It asked how they perceived the assistant, whether they felt responsible for its condition, and whether the decline changed how they wrote.

Exhausted AI research feedback form
The feedback form was designed to reduce observer influence and capture more reflective answers after the interaction.

User Testing and Findings

I tested the prototype with 10 people from design and programming backgrounds. The response was mixed, which was more useful than a clean success story.

  • 5 participants did not care much about the assistant's visible fatigue.
  • 3 participants said it made them think, but not enough to treat the AI like a human with limitations or feel strong empathy toward it.
  • 2 participants found it genuinely thought provoking and said it changed how they phrased prompts, even if it did not fully change their overall approach to AI.

The clearest insight was that visible fatigue changed perception more than behavior. A more personal conversation might produce a different response than the programming task used here, so that is the next thing I would test.

Build and Delivery

I built and hosted the web prototype through Vercel. The system combines visible UI states, browser session logic, response delay, and OpenAI-based behavior shaping. I used Codex for development support, then checked and adjusted the interaction against the research goal.

  • Hosting: Vercel
  • Build: HTML, CSS, JavaScript, and OpenAI integration
  • Core interaction: fatigue score, visible state system, prompt-tone effects, and delay-based feedback

What I Learned

A provocative interaction can create awareness without changing behavior. The next version should use a more personal conversation and a broader participant group to see whether the effect holds beyond a programming task.

Selected References

The prototype was informed by research on computers as social actors, expectation gaps in conversational systems, and personification of voice assistants.

View the six sources
  • Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors.
  • Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers.
  • Luger, E., & Sellen, A. (2016). Like having a really bad PA: The gulf between user expectation and experience of conversational agents.
  • Pradhan, A., Findlater, L., & Lazar, A. (2019). Phantom friend or just a box with information: Personification and ontological categorization of smart speaker-based voice assistants by older adults.
  • Purington, A., Taft, J. G., Sannon, S., Bazarova, N. N., & Taylor, S. H. (2017). Alexa is my new BFF: Social roles, user satisfaction, and personification of the Amazon Echo.
  • Zhou, J., Porat, T., & van Zalk, N. (2024). Humans mindlessly treat AI virtual agents as social beings, but this tendency diminishes among the young.

Let's Connect

I'm interested in products where interface behavior itself becomes part of the concept, not just the container around it.

EmailElijaholiverlarsson@gmail.com