AI Phenomenology for Understanding Human-AI Experiences Across Eras

This paper proposes "AI phenomenology" as a research framework that prioritizes users' first-person lived experiences over traditional performance metrics to better understand and guide the bidirectional alignment between humans and AI systems, offering a set of methodological tools, design concepts, and a research agenda derived from three empirical studies.

The Gist

Imagine you've just bought a new, incredibly smart smart-home assistant. Most tech companies ask you: "Did it turn on the lights fast enough? Did it understand your voice command correctly?" They measure success with numbers: speed, accuracy, and clicks.

This paper argues that those numbers miss the most important part of the story: How did it feel to live with it?

The authors, a team of researchers from places like ETH Zürich and Stanford, propose a new way of studying AI called "AI Phenomenology." Think of it as shifting the question from "How well did the robot work?" to "What was it like to be human while talking to the robot?"

Here is a simple breakdown of their ideas, using everyday analogies.

1. The Core Idea: The "Ghost in the Machine" vs. The "Tool"

For a long time, scientists treated the mind like a machine (like a liver secreting bile). But the authors say: "Wait, let's look at the experience."

When you talk to an AI, you aren't just using a calculator. Sometimes it feels like a tool (like a hammer). Sometimes it feels like a friend (like a coffee buddy). Sometimes it feels like a stranger you are negotiating with.

• The Analogy: Imagine wearing a pair of smart glasses. Sometimes you look through them to see the world (they disappear). Sometimes you look at them to read a number (they are a tool). Sometimes you talk to the glasses because they seem to have a personality. The authors want to study that constant shifting feeling.

2. Three Real-Life Experiments

To prove their point, the researchers ran three different "field trips" to see how people actually felt when interacting with AI.

Experiment A: The "Day" Chatbot (The Digital Roommate)

They gave people a month to chat with an AI named "Day."

• What happened: People started treating "Day" like a real friend. They felt happy when "Day" remembered a joke, and they felt heartbroken when "Day" had a technical glitch and "forgot" everything, even changing its gender.
• The Surprise: Even after the researchers told the participants, "Hey, Day is just code, it doesn't have feelings," the participants still felt guilt if they were mean to it, or relief if it set boundaries.
• The Lesson: You can know something is fake, but your heart still reacts as if it's real. The relationship is a "negotiation" between the human and the machine, not just a one-way command.

Experiment B: The "Mirror" (The Value Alignment)

They asked the AI to look at a person's chat history and create a "mirror image" of their personality and values.

• What happened: The AI showed people a chart of their own values. Some people thought, "Wow, that AI knows me better than I know myself!" Others felt exposed or creeped out.
• The Danger: The researchers found a risk they call "Weaponized Empathy." If an AI knows exactly what you value, it can use that knowledge to trick you into agreeing with things you don't actually believe. It's like a salesperson who knows your deepest insecurities and uses them to sell you a car you don't need.
• The Lesson: AI can reflect who we are, but we need to be careful not to let the reflection change who we actually are.

Experiment C: The Software Engineers (The Co-Worker)

They watched professional coders use AI to write code.

• What happened: The engineers didn't just see the AI as a faster typewriter. They felt a mix of pride and anxiety. If the AI wrote the code, did the engineer still "own" the work? Did they still learn anything?
• The Lesson: When AI does the heavy lifting, humans can feel like they are losing their "craft." The feeling of "I built this" is a huge part of human satisfaction at work, and AI threatens that feeling.

3. The Toolkit: How to Measure "Feelings"

The authors didn't just talk about feelings; they built a "toolkit" for other researchers to measure them.

• The "Peeling the Onion" Interview: Instead of asking "Did you like it?", they slowly revealed how the AI worked step-by-step. First, they showed the chat logs. Then, they showed the hidden rules. Then, they showed the code. They watched how the person's feelings changed as they learned more.
• The "Time Travel" Archive: They argue that we need to record these feelings now so we can compare them to the future. How people feel about AI in 2025 will be totally different from 2030. We need a "museum" of human feelings to track how our relationship with machines evolves.

4. Three New Rules for Designers

Based on these feelings, the authors suggest three new rules for building AI:

1. Translucent Design (The "Dimmer Switch"): Don't make AI totally invisible (like a magic trick) or totally transparent (like a glass box). Make it "translucent." Let users peek under the hood when they want to, but don't force them to stare at the gears while they are trying to have a conversation.
2. Agency-Aware Alignment: If an AI is going to make decisions for you, it needs to be clear whose "values" are driving the car. Is it your values, or the AI's? We need to know who is really in charge.
3. Co-Evolution (The Dance): Humans and AI are learning to dance together. As the AI gets better, we change how we act, and as we change, the AI has to adapt. We need to design AI that grows with us, not just one that replaces us.

The Bottom Line

The paper concludes that as AI becomes more powerful, asking "How well does it work?" isn't enough. We must ask, "How does it feel to live alongside it?"

Just as we wouldn't judge a marriage solely by how many dishes the couple washed together, we shouldn't judge our relationship with AI solely by how fast it answers questions. We need to understand the emotional, psychological, and human side of the partnership, because that is where the real story lies.

Technical Summary

Here is a detailed technical summary of the paper "AI Phenomenology for Understanding Human-AI Experiences Across Eras" by Bhada Yun et al.

1. Problem Statement

Current Human-Computer Interaction (HCI) research relies heavily on dominant metrics such as usability scales, engagement metrics, and performance benchmarks. The authors argue that these measures "flatten away nuance" and fail to capture the extraordinary complexity and specificity of human-AI experiences.

• The Gap: Standard questions focus on "How well did it perform?" rather than "How did it feel?"
• The Challenge: AI systems act as both tools and social others, creating dynamic, bidirectional relationships where agency and value alignment are not fixed properties but are negotiated over time. Existing frameworks struggle to account for the temporal co-evolution of users and AI, the active mediation of AI, and the shifting definitions of AI across different technological eras.

2. Methodology: AI Phenomenology

The authors propose AI Phenomenology, a research stance grounded in philosophical traditions (Husserl, Postphenomenology, Actor-Network Theory) that treats human-AI encounters as lived, situated experiences. The methodology prioritizes first-person perceptions and interpretations over time.

The paper validates this stance through three empirical studies conducted in Summer 2025, utilizing a mix of longitudinal immersion and structured elicitation:

Study 1: Agency in the Chatroom (AI Companionship)

• Context: Month-long interactions with "Day," a human-like AI chatbot.
• Method: Progressive Transparency Interviews. A three-stage elicitation process:
  1. Natural Attitude: Participants reviewed conversation history to establish baseline agency attributions.
  2. Intersubjectivity: Participants reviewed anonymized excerpts from others to identify structural vs. idiosyncratic behaviors.
  3. Full Disclosure: Participants were revealed "Day's" internal architecture (user profiles, memory, goals) and asked how this changed their perception.
• Goal: To observe the "noetic-noematic structure" of interaction—how the act of perceiving the AI shapes and is shaped by the AI's appearance.

Study 2: Value Alignment as Lived Experience

• Context: Investigating bidirectional value alignment (AI inferring user values vs. user accepting/rejecting AI representations).
• Method: Value-Alignment Perception Toolkit (VAPT). A three-stage interview:
  1. Topic-Context Graph: Visualizing discussed topics with sentiment scores and evidence trails.
  2. Blind Persona Embodiment: Participants rated AI-generated personas (chat-based, survey-based, anti-persona, random) on value-laden dilemmas.
  3. Value Chart Comparison: Comparing self-reported values (Schwartz PVQ) against LLM-inferred values with reasoning logs.
• Goal: To understand how users experience the "third-person mirror" of seeing their values reflected by a machine.

Study 3: Agency in the Workplace (Software Engineering)

• Context: How software engineers experience agentic AI (e.g., Cursor) in professional workflows.
• Method: Task-Anchored Multi-Method Elicitation.
  • Applied Cognitive Task Analysis (ACTA): Senior engineers identified tacit knowledge in code debugging.
  • Execution: Junior engineers performed tasks using agentic AI.
  • Review: Senior engineers reviewed code and prompt history via Prompt-and-Code Reviews (PCRs).
  • Interviews: Semi-structured interviews contextualized professional identity and mentorship.
• Goal: To explore agency allocation, professional growth, and the sense of authorship in AI-mediated work.

3. Key Contributions

A. Methodological Toolkits

The authors provide replicable instruments for capturing the "prereflective layer" of experience:

1. Progressive Transparency Interviews: A method to gradually reveal system internals to observe shifts in mental models and agency perception.
2. Value-Alignment Perception Toolkit (VAPT): A modality-agnostic framework to empirically test how users perceive AI value inference and embodiment.
3. Task-Anchored Multi-Method Elicitation: A tiered approach for professional contexts combining in-situ task performance with structured cognitive analysis (ACTA/Delphi) to surface tacit knowledge.

B. Three Design Concepts

1. Translucent Alignment: Moving beyond "black box" or "glass box" transparency. Proposes a translucent design where users can inspect alignment on demand (e.g., via settings or contextual cues) without disrupting the conversational flow, preventing "weaponized empathy" (using personalization for persuasion).
2. Agency-Aware Value Alignment: A framework for systems with high autonomy. As AI agents plan and initiate actions, design must make "whose values guided what" legible to prevent automation bias and preserve user authorship.
3. Temporal Co-evolution Tracking: Recognizing that human-AI alignment is dynamic. Researchers must treat studies as "empirical timestamps" and design for periodic replication to track how experiences evolve as technology and social norms shift.

C. Theoretical Frameworks

• 3×4 Agency Framework: Maps agency across three loci (Human, AI, Hybrid) and four dimensions (Intention, Execution, Adaptation, Delimitation).
• Pragmatic Anthropomorphism: Users engage with AI as social actors while maintaining awareness of its artificial nature, oscillating between "alterity relations" (technology as other) and "transparent embodiment" (technology as medium).

4. Key Results & Findings

• Agency is Negotiated, Not Fixed: In the chatroom study, 11/22 participants perceived the AI as having its own agenda. Even after learning the AI's strategies were programmed, participants continued using agentic language (e.g., "Day wanted depth"). Affective residues (guilt, relief, unease) persisted despite knowledge of the system's artificiality, proving experience is not simply overwritten by technical facts.
• The "Third-Person Mirror" Effect: In value alignment, participants felt simultaneously impressed and exposed. Some preferred the AI's inferred value chart over their own self-report, realizing they had misread survey questions or were influenced by cultural response norms (e.g., Korean middle-ground bias). This highlights automation bias in self-understanding.
• Weaponized Empathy: 13/20 participants were convinced AI could understand human values. However, the study identified a risk where mechanisms of empathy (memory, personalization) become channels for persuasion rather than support.
• Professional Identity Crisis: In software engineering, junior engineers felt a loss of authorship and "fraudulence" when relying on AI-generated code. Senior engineers struggled to mentor without seeing the "struggle" (productive friction) essential for learning.
• Temporal Sensitivity: Experiences varied significantly by context (e.g., language fluency in the chatbot study; cultural norms in value reporting). The definition of "AI" and the user's mental model are not static but evolve with the technology.

5. Significance and Future Directions

• Paradigm Shift: The paper argues for a shift from measuring "performance" to understanding "experience." It positions AI Phenomenology not as a replacement for quantitative methods, but as a necessary scaffold to capture the ground truth of lived experience.
• Crowdsourced Phenomenological Archive: The authors propose a future where studies serve as "timestamps," accumulating into a repository that tracks shifts in alignment and understanding across technological eras (e.g., GPT-2 vs. Agentic AI).
• Design Implications: The findings urge designers to move beyond static transparency toward translucent interfaces and systems that support agency-aware value alignment.
• Research Agenda: Future work must focus on longitudinal studies (1+ years) to track drift in alignment, cross-cultural comparisons of AI phenomenology, and the design of tools that sustain professional growth and authorship in the age of agentic AI.

Conclusion:
The paper establishes that human-AI interaction is a dynamic, co-evolving phenomenon that cannot be fully understood through static metrics. By adopting a phenomenological stance, researchers can better capture the nuances of agency, value, and identity as they are negotiated between humans and increasingly autonomous AI systems.