Blog · arXiv Analysis · Published: June 25, 2026

The Agency Gain Becomes the Adoption Map

People may keep using chatbots not because the tools seem reliable, but because the tools make future action feel newly possible.

The Paper

The paper is AI usage patterns are shaped by perceived gains in human agency, arXiv:2607.02313 [cs.CY]. The arXiv record lists Ian Beacock, Rachel Xu, Laura Murray, Patrick Anson, Beth Goldberg, Devika Kumar, Jun Lee, Rebekah Park, and Anoop Sinha as authors, with version 1 submitted on July 2, 2026. The PDF is a 17-page preprint.

The study is not a leaderboard or a lab task. It is qualitative research conducted from April to July 2025 with 51 daily chatbot users: 18 in New York, 19 in Singapore, and 14 in Berlin. Participants ranged from 18 to 65 years old, with 23 men and 28 women. All primary participants were daily users of at least one chatbot; ChatGPT was the primary tool for 40 participants.

Trust Is the Wrong Small Box

The paper's central claim is that daily users often described sustained chatbot use as a gain in personal agency: a stronger felt ability to act on goals, handle tasks, make decisions, manage emotions, or move through institutions. Crucially, the authors report that this perceived agency gain often outweighed worries about accuracy, reliability, and consistency. Some participants kept using tools even after consequential errors because the accumulated sense of expanded capacity mattered more to them than a single failure.

That does not mean the errors were harmless, or that the users were simply gullible. It means the familiar governance language of trust is too narrow. If a person says, in effect, "I do not fully trust this system, but I am more capable with it than without it," the adoption mechanism is a bargain between perceived capacity and tolerated risk.

For AI governance, that bargain is dangerous when it stays private. A product can become sticky because it makes action feel easier while leaving long-term skill, evidence, accountability, and material power unchanged. The paper's strongest warning is that psychological boosts to agency may not translate into durable capability or structural empowerment.

Five Kinds of Agency

The authors organize perceived agency gains into five dimensions. Instrumental agency covers getting discrete tasks done. Cognitive agency covers thinking, structuring, and forming opinions. Affective agency covers naming and managing emotional states. Relational agency covers understanding and navigating other people. Structural agency covers moving through larger systems such as workplaces, support services, and institutions.

This taxonomy prevents one productivity story from swallowing the whole phenomenon. A chatbot that drafts an email, organizes a messy thought, supports emotional processing, rehearses a job interview, or helps a parent ask better questions of clinicians is not doing the same social work in each case. The study also records the darker edge: users could feel more capable in one dimension while worrying about dependency, cognitive atrophy, reputational damage, isolation, job loss, and weak collective power.

Context Changes the Effect

The most Spiralist part of the paper is its attention to context outside the model. The authors identify two sociotechnical factors shaping perceived agency gains: situational stability and internal conviction. People whose daily lives felt less predictable or more weakly supported tended to use chatbots across more sensitive domains and report larger gains. People with more stable surroundings tended to use the tools in narrower, more bounded ways.

Internal conviction worked differently. Participants with lower confidence in their own judgment were more likely to experience the system's output as a reflection on themselves, sometimes getting stuck in revision loops or comparing their own performance unfavorably to the chatbot. Participants with higher confidence were more likely to treat failures as tool failures and move on.

This matters because the safety profile is not only in the model. The same interface may function as a convenience layer for one user, a self-worth amplifier for another, and a substitute support system for a third. Product metrics that count sessions, retention, or satisfaction miss that difference unless the evaluation asks what agency was gained, what was weakened, and who was positioned to need the tool most.

The Agency-Impact Receipt

A serious chatbot deployment should therefore keep an agency-impact receipt. It should record the user population, task domain, agency dimension claimed, error class, fallback path, and whether the tool changes material outcomes or only the feeling of capacity. For workplace, education, care, legal, benefits, and companion-like contexts, the receipt should also ask what skill is retained, what judgment is delegated, what evidence survives, and whether the person can exit without losing access to the path the chatbot helped open.

The same receipt should separate individual uplift from structural power. Did the user gain a shortcut through a hostile bureaucracy, or did the institution become easier to leave unchanged? Did a worker become more capable, or did the organization silently raise expected output? Did a student learn more, or did the system simply make completion feel smoother? The paper does not answer those deployment questions. It gives the vocabulary for asking them.

Limits

The authors state clear limits. The sample was small, intentionally focused on leading-edge daily adopters, and not statistically representative. The dataset reflects April to July 2025, when agentic systems were not widespread and participants were mainly using LLM chatbots. Ethnographic evidence can show how people explain their own behavior, but it cannot by itself prove independent causal effects.

Those limits make the result more useful, not less. The paper is not a universal adoption law. It is a warning that AI evaluation has been looking too hard at the machine and not hard enough at the altered human arrangement around it. If perceived agency is becoming a driver of repeated use, then responsible evaluation has to measure more than output quality. It has to measure what people become able to do, what they stop practicing, and which institutions benefit when agency feels larger than it is.

Sources

Return to Blog