How Cycle 1 Shaped Cycle 2

Insight 1  “Situatedness” could not be produced through generalized scenarios.

In Cycle 1, participants struggled to immerse themselves in the fake LLM environment with seven different general scenarios of ethical concerns. Because the scenario reflected researcher's own use cases rather than the participants', the user testing revealed that ethical nudges could not be evaluated meaningfully in a decontextualized, one-size-fits-all task.

Insight 2  Users carry drastically different ethical concerns (including none).

The assumption that all users would bring some ethical concern into AI interactions did not hold. One participant expressed no concerns at all, while another only developed concerns when nudged about environmental impact. Although it was limited user testing, this diversity highlighted the need to understand individuals’ different values before testing nudges.

These two insights required a shift in Cycle 2 toward deeply situated, personally relevant, embodied use rather than externally constructed prompts.  

What Changed Because of Those Insights

The project moved away from fabricating test scenarios and toward exploring ethical nudges within real, lived contexts using self auto-ethnography my self as a design researcher. The method of auto-ethnography was chosen for its capacity to capture reflexive, situated, and affective dimensions of experience within real research contexts.

The scope was narrowed to me as a researcher's own daily academic research web-based tools—five familiar platforms that recently added AI features—to create a truly situated environment.

After the my experience was recorded for each tools, I studied each platforms' AI statement to assess how the companies align with the well-known Responsible AI Principe, The Montreal declaration of Responsible AI. Then later, analyzed the company's stated AI Principles to compare how my real experience was aligned with those principles.  

 

Method Refinement

I chose five everyday web platform tools that I use in ever day design research practice. All services recently adopted AI features. Three pilot studies were conducted to determine the most effective data-capture approach.

1. 1st Pilot Test
Whenever I was using the service platforms, I was intentionally being “aware” of AI feature in each platform and its interfaces. I started screen capturing my screen whenever I was using AI feature. Then after spending some time with the platform using AI Feature, I would stop and collect all screenshots I took and tried to remember which features I clicked and explored. However this process was not accurate because I could not remember all of my behaviors and thoughts. I was not sure which interface I clicked and how long I stayed and lost track of order of all behaviors I have done previously.
        

2. 2nd Pilot Test
I tried the Tobii eye-tracking program, which monitors users’ gaze, attention, and gaze duration. This method provided detailed data on how my eyes focused on certain content or interfaces while I interacted with the platforms. However, scheduling access to the program within a limited time slot was difficult, and the usability of Tobii was not ideal; the requirement to sit still, avoid movement, and maintain a fixed position throughout the session made natural interaction complicated.

3. 3rd Pilot Test

Combined screen recording with post-session notes improved accuracy but lacked real-time affective capture. 

Based on these pilot tests, the final method I chose combines Zoom video/audio recording with think-aloud protocol. Each sessions recordings were then later captured by MacBook screen capture feature and post-session reflective notes, balancing accuracy, emotional depth, and ecological validity.

Revised Assumptions

The first assumption I revised for Cycle 2 was to change the set-up “universal” scenario to "auto-ethnography" which allows the user experience situation in realistic, contextual, and personally relevant setting.

The second assumption I revised to choose general well-known AI Principles Ethical concerns are different in all users and depending on the situations. Ethical nudges may only matter if the user is situated in certain environment of using AI. 

New Design Research Questions

Previous Research Question for Cycle 1 was:
"How might designing ethical nudges affect the usage of AI?"

Refined Research Question for Cycle 2 is:
"How do the AI principles stated by everyday research platforms compare to the felt experience of users interacting with the recent AI features?"

Evolving Research Question for the Next Step is:
"How might ethical generative experiences and interfaces be designed so that users can feel that a platform is acting responsibly?"

What Cycle 2 revealed: 

Many platforms visually emphasize AI with icons and sparkles, yet the experience from auto-ethnographic sessions revealed more friction instead of magic. The “enchantment aesthetic” hides how non-transparent and uncontrollable these tools actually are.

The recent embedded AI features across platforms are default to always-on, reducing user control. The interfaces often obscure what the AI is doing, leading to confusion or loss of agency. The errors in AI features, wait times, and overwritten user content contradict the “enchanted” branding of AI.

Prototype 2

The second prototype I designed was to explore whether ethical GUI/GX patterns—drawn from Responsible AI guidelines—meaningfully influence how users behave, perceive, and reflect while using AI features. The prototype stands in for an everyday tool (e.g., writing platform, search tool) that integrates AI ethically. It demonstrates what responsible UX could look like even before full mechanics are built.

Specifically, this prototype tests:

• Agency: Can the user easily choose AI versus traditional modes, turn AI features on/off, and control what AI does?

• Transparency: Does the interface clearly reveal AI’s capabilities, estimated generation time, limits, and data use?

• Traceability: Can users track their own prompts and understand system actions?

Later, this prototype will be tested to the users to check the experience quality-Do users feel the service is practicing Responsible AI principles?

The prototype functions primarily as a probe to provoke reflection on what “responsible AI experience” actually feels like in everyday tools.

What this prototype offers:

• Allows users to control default AI modes and toggle AI on/off.

• Makes prompt history visible and trackable.

• Presents UI elements intended to communicate responsible, transparent, controllable AI.

• Allows users to easily access AI principles and information on AI features.

What is functional:

• Basic on/off control interactions

• A representational layout for prompt tracking

• Interface elements structured around Responsible AI principles (labels, descriptions, modes)

Next step planned:

• Identifying which existing platform this prototype best maps onto

• Developing alternate nudge modalities (haptic, tangible, sensory) to explore felt nudges

• Design survey and interview to gain insight on the users who use AI daily basis. (Designers)

• Adding behavioral logging to evaluate timing and user response to nudges

• Participatory workshop to test the prototype with the users

Ethical UX design must include predictability, explainability, and boundaries, not just warnings or nudges. Nudges should adapt to the user's situation and values rather than assume uniform concerns. Measuring ethical impact requires access to real behavioral traces, not hypothetical tasks.

Questions that Cycle 2 Raises:

"How do user perceptions of AI shift from the initial enchantment to long-term frustration or adaptation?"

"Which ethical principles matter most to users during real tasks—agency, transparency, environmental impact, or something else?"

"How can nudges be embedded in everyday interfaces without feeling invasive or paternalistic?"

What This Means for the Next Step

The next step is to define what an “effective ethical nudge” looks like within a situated everyday tool and to clarify the relationship between the nudge, the task, and the user’s moment-to-moment behavior. The upcoming prototype needs to incorporate an interface or nudge that responds dynamically to user behavior. These nudges must be tailored to individual situated contexts, and the system must embed transparency and user control without overwhelming the interface. This phase also requires determining whether the focus is on ethical nudges, ethical experiences, ethical interfaces, or a combination of all three within a unified human-centered framework. 

Overall, the project is will move toward defining what a Responsible Generative Experience (RGX) entails and how to design for felt responsibility—not merely stated principles.