Ownership in AI-Assisted Everyday Tasks
Megan Wei, Melanie Subbiah, Audrey Lee, Annya Dahmani, Dave Edwards, Helen Edwards, Ellie Pavlick
cs.AI, cs.HC
2026-09-18
Brown survey of 64 heavy AI users: people disown work they merely approve but keep ownership when they lead, iterate, or rewrite; tasks impossible without AI feel most their own.
When AI helps finish a task, does the result still feel like yours? As AI spreads through everyday work, the question goes beyond bylines. A sense of pride and ownership over one's decisions is a pillar of purpose and well-being, and slowly diluting it carries a cost. Prior research focused on cognitive offloading, the delegation of cognitive effort to a tool, and found comprehension and learning suffer while efficiency gains stay modest. A team from Brown, CMU, UC Berkeley, and the Artificiality Institute pointed the camera at the emotional layer: when does AI-assisted output still feel like one's own, and when does that feeling evaporate?
An exploratory qualitative survey built around self-comparison. Each participant recalled two recent AI-assisted tasks, one whose result felt like their own and one that did not, presented in randomized order, then answered free-text questions on identity, independent contribution, pride, and willingness to disclose AI use, and finally explained what differed between the two.
| Measure | High-ownership tasks | Low-ownership tasks |
| “Definitely could do it without AI” | 42% | 63% |
| Proud of the result | 74% | 25% |
Four themes:
Disclosure came apart from pride. Roughly 80% in both groups were comfortable revealing AI use, while 10% were explicitly uncomfortable on low-ownership tasks. What governed the decision were community norms and fear of credit erasure, not pride. Most preferred explaining how they had used AI over a binary “AI-generated” label.
For AI product teams the survey offers three testable directions: process visibility (does recording and showing edits, rejections, and prompt history strengthen felt ownership?), personalization that actually differentiates (is the scaffolding users believe unique really effective?), and disclosure formats (narrative versus categorical labels). For heavy users the takeaway is blunt: if you want the work to feel like yours, do not retreat into the approver's seat. This is exploratory qualitative work; it produces hypotheses, not conclusions.
The authors concede a badly skewed sample: mostly mid-to-late career, highly educated, white (79.7%), US-based (71.7%) heavy AI users recruited through professional networks. Fifty-two complete responses cannot support statistical inference, every theme comes from qualitative coding, and the authors position the study as a first step before structured research. One confound goes unaddressed: high-ownership tasks skewed toward creative projects, so creative pride and process-driven ownership cannot be separated in this data. Pride and disclosure answers were coded by keyword rules with denominators shifting by question, so read 74% versus 25% as an order of magnitude.