💻 computer science

AI usage patterns are shaped by perceived gains in human agency

This ethnographic study of daily AI chatbot users across three countries reveals that sustained human engagement with conversational AI is driven primarily by perceived gains in individual agency rather than technical reliability, challenging traditional trust-based models and highlighting a critical tension between immediate psychological empowerment and long-term structural capacity.

Original authors: Ian Beacock, Rachel Xu, Laura Murray, Patrick Anson, Beth Goldberg, Devika Kumar, Jun Lee, Rebekah Park, Anoop Sinha

Published 2026-07-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ian Beacock, Rachel Xu, Laura Murray, Patrick Anson, Beth Goldberg, Devika Kumar, Jun Lee, Rebekah Park, Anoop Sinha

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Why We Keep Using AI (Even When It Messes Up)

Imagine you have a new, magical sidekick. Sometimes this sidekick is brilliant and helps you build a house in a day. Other times, it accidentally hands you a brick that turns out to be made of jelly, causing the wall to collapse.

According to this study, most people don't fire the sidekick when the wall collapses. Instead, they keep using it because it makes them feel like superheroes.

The researchers (from Google's Jigsaw team and Gemic) interviewed 51 daily AI users in the US, Germany, and Singapore. They wanted to know: Why do people keep talking to AI chatbots every day, even when the AI makes mistakes?

The answer isn't that people trust the AI to be perfect. The answer is that the AI makes people feel more powerful (a concept the paper calls "agency").

Key Finding 1: The "Superpower" Feeling

The study found that people link their daily AI use to a feeling of increased personal power. It's not just about finishing a task faster; it's about feeling like they can shape their future.

The researchers broke this "superpower" feeling down into five areas:

The Tool Belt (Instrumental): "I can fix my door faster."
The Brain Booster (Cognitive): "I can organize my messy thoughts."
The Emotion Coach (Affective): "I can understand why I'm angry and calm down."
The Social Navigator (Relational): "I can figure out how to talk to my boss or date better."
The System Hacker (Structural): "I can finally understand the confusing rules of my son's medical care."

The Analogy: Think of AI like a pair of gym shoes. Even if the shoes are a little scuffed or the laces are untied (errors), the runner keeps wearing them because they feel like they can run faster and further than they could in bare feet. The feeling of potential outweighs the flaw in the shoe.

Key Finding 2: Mistakes Don't Matter as Much as You Think

Traditionally, we think humans only use technology if it is reliable and accurate. If a car brakes fails once, we stop driving it.

But with AI, the study found something different. People treat AI errors like speed bumps rather than roadblocks.

One participant had an AI fail to build a crucial investor presentation right before a meeting. It was a disaster.
However, she didn't stop using the AI. She said, "I can't do this with you right now," but she kept using it for everything else.
Why? Because the AI had helped her escape a cult and improve her life in other ways. The "superpower" feeling was worth the occasional crash.

The Analogy: Imagine a GPS. Sometimes it tells you to drive into a lake. You get annoyed, but you don't throw the phone in the trash. You keep using it because, 99% of the time, it gets you to your destination faster than you could on your own. The benefit of the journey outweighs the occasional wrong turn.

Key Finding 3: The "Trust" Switch Has Been Flipped

For decades, scientists thought we used technology based on Trust in the Machine. (i.e., "Is this machine accurate? Is it safe?")

This study suggests that with conversational AI, the switch has flipped. Now, we judge the technology based on Trust in Ourselves.

We ask: "Does using this tool make me feel capable?"
If the answer is "Yes, I feel smarter, calmer, and more in control," we keep using it, even if the machine is unreliable.

The Analogy: It's like using a magic wand. You don't care if the wand is made of cheap plastic or if it sparks sometimes. You care that you feel like a wizard when you wave it. The value isn't in the wand; it's in the feeling of power it gives you.

Key Finding 4: Who Feels the Most Powerful?

Not everyone feels the same "superpower" boost. The study found two main factors that change how much power people feel:

Life Stability:
- People with chaotic lives (unstable jobs, family issues, loneliness) felt the biggest boost. For them, AI was a lifeline—a "big sister" or "personal trainer" filling gaps in their support system.
- People with stable lives (secure jobs, strong families) felt a smaller boost. They used AI mostly as a simple tool for work, like a calculator, and didn't feel it changed their whole life.
Self-Confidence:
- People who doubted themselves felt a huge boost but also a risk. They sometimes blamed themselves when the AI failed, thinking, "I'm stupid because the AI couldn't help me."
- People who were already confident used AI as a tool. If the AI failed, they just switched to a different one without feeling bad about themselves.

The Big Warning: The "Feeling" vs. The "Reality"

The paper ends with a crucial tension.

The Feeling: People feel more powerful and capable right now.
The Reality: The study does not know if this feeling translates to long-term success or if people are actually getting better at things on their own.

The Analogy: Imagine taking a caffeine pill. You feel energetic, sharp, and ready to conquer the world (perceived agency). But the paper asks: Is this actually building your muscles, or are you just borrowing energy from tomorrow?

There is a fear that people might be getting a short-term "confidence boost" while slowly losing their own ability to think or act without the AI (cognitive atrophy). The study found that while people felt powerful individually, they didn't feel like they had more power to change big systems like the government or the economy.

Summary

What happened: Researchers watched 51 daily AI users.
What they found: People keep using AI not because it's perfect, but because it makes them feel like they have superpowers.
The twist: This feeling of power is so strong that people ignore the AI's mistakes.
The catch: We don't know yet if this feeling of power is real, lasting growth, or just a temporary high that might leave people weaker in the long run.

Technical Summary: AI Usage Patterns Shaped by Perceived Gains in Human Agency

Problem Statement
As conversational AI systems (specifically large language model chatbots) become deeply integrated into daily life, the implications for human agency—the capacity for humans to take meaningful action to shape their world—remain poorly understood. While theoretical frameworks suggest AI could either amplify capability or induce disempowerment through dependency and cognitive outsourcing, there is a scarcity of empirical, ecologically-valid evidence regarding cumulative usage patterns. Current research relies heavily on controlled experiments or large-scale log analysis, leaving critical gaps in understanding: (1) how subjective experiences of agency are affected by real-world AI usage; (2) the contextual and sociotechnical factors influencing these experiences; and (3) the longitudinal effects of AI integration outside of controlled settings. Existing trust-based models, which posit that user adoption relies on reliability and accuracy, may be insufficient to explain sustained human-AI interaction.

Methodology
The authors conducted a deep ethnographic study of daily AI chatbot users ( $n=51$ ) across three distinct cultural contexts: the United States (New York), Germany (Berlin), and Singapore. Data collection occurred between April and June 2025. The study employed a multi-phase, mixed-method approach to capture situated practices:

Phase 1 (Diary Study): Participants submitted video diaries (1–2 minutes) detailing specific AI interactions, accompanied by screen-sharing and voiceover.
Phase 2 (In-Person Interviews): Semi-structured, 120-minute interviews were conducted in participants' homes and offices. These included individual sessions and dyadic interviews (involving a friend or family member) to uncover discrepancies between self-reported behavior and external observation.
Phase 3 (Follow-up): Remote 60-minute interviews with a subset of participants ( $n=12$ ) to refine analytical hypotheses and conduct member-checking.

The analysis utilized constructivist grounded theory, iteratively coding 315 discrete observations to identify themes. The study defined agency not merely as a "sense of control" over actions, but as a holistic phenomenological sensation of capacity mediated by sociocultural and sociotechnical forces.

Key Contributions and Results
The study identifies four primary findings that challenge prevailing assumptions about human-AI interaction:

Link Between Cumulative Usage and Perceived Agency Gains:
Participants consistently linked sustained AI use to an expanded sense of individual agency. This was experienced as both immediate task effectiveness and a holistic increase in long-term potential. The authors categorize these gains into five dimensions:
- Instrumental: Efficiency in completing discrete tasks (e.g., writing emails, fixing physical objects).
- Cognitive: Improved ability to structure thoughts and develop opinions.
- Affective: Enhanced capacity to name, understand, and manage emotional states.
- Relational: Improved ability to navigate and sustain connections with others.
- Structural: Increased ability to navigate complex systems (e.g., healthcare, legal, employment).
Agency Gains Outweigh Errors in Shaping Usage:
Contrary to traditional trust models where reliability is a prerequisite for adoption, participants frequently continued or increased usage despite consequential AI errors (e.g., hallucinations, failed deliverables). Users engaged in a cost-benefit calculus where perceived agency gains outweighed inaccuracies. For many, the technology was viewed as an "unreliable but highly useful instrument," and trust was not a binary state but a risk-balancing heuristic.
Coexistence of Agency Gains and Significant Concerns:
Users simultaneously reported feeling more capable while harboring deep reservations about long-term risks, including cognitive atrophy, dependency, and societal disempowerment. Despite fears of "getting dumb" or losing structural power, the immediate psychological boost to perceived agency prevented users from altering their usage patterns.
Variation Driven by Situational Stability and Internal Conviction:
The intensity of perceived agency gains was not uniform; it was mediated by two sociotechnical factors:
- Situational Stability: Users with low stability (chaotic lives, lack of support) reported more expansive, emotional, and transformative agency gains, often using AI to fill gaps left by missing social support systems. Users with high stability used AI more instrumentally and reported narrower gains.
- Internal Conviction: Users with low self-confidence often tied their self-worth to AI performance, leading to cycles of self-blame. Users with high internal conviction treated AI performance as independent of their own worth, allowing them to disengage from errors more easily.

Significance and Claims
The paper argues that traditional trust frameworks, which focus on the machine's reliability and accuracy, are insufficient for explaining human behavior with conversational AI. Instead, the locus of judgment has shifted: users are evaluating how the technology affects their own sense of capacity and self, rather than evaluating the machine's trustworthiness.

The authors claim that perceived agency acts as a primary driver of sustained usage, often overriding concerns about accuracy. However, they highlight a critical tension: immediate psychological boosts to perceived agency do not necessarily translate into material empowerment, structural change, or long-term skill acquisition. In some cases, aggregate usage may deliver short-term capacity boosts while undermining future agency through cognitive deskilling.

The study concludes that future research and AI benchmarking must move beyond measuring model performance against human reasoning. Instead, frameworks must account for the multidimensional effects of cumulative usage on human agency, considering the dialectical relationship between individual users and the structural contexts in which they operate. The authors call for novel behavioral frameworks and measurement tools that can ensure conversational AI strengthens human agency in substantial, sustained ways, rather than merely providing temporary psychological relief.