People own AI-assisted work when they lead it, not when they approve it
People feel they own the work they do with AI when they lead it, and lose that feeling when they only approve what the machine produces. That’s the main finding of “Ownership in AI-Assisted Everyday Tasks,” a paper from our summer research fellowship. It has been accepted to the AI and the Self workshop at NeurIPS, one of the biggest AI research conferences, in Paris on Dec. 12.
The fellows asked 64 people, mostly heavy AI users, to describe two recent tasks they’d done with AI: one that felt like theirs and one that didn’t. People who directed, iterated, and rewrote kept ownership. People who signed off did not. As one respondent put it: “I reviewed, I looked, I was in the loop, but I wasn’t in charge.” Being in the loop isn’t the same as being the author.
The most surprising result is that people often felt most ownership over work they couldn’t have done alone. Only 42% said they definitely could have done their high-ownership task without AI, compared with 63% for tasks that didn’t feel like theirs. Two respondents had long wanted to make illustrated books but couldn’t draw. With AI as their illustrator, they owned the vision, if not the drawing.
Two things wore ownership away: AI’s bland, generic voice, and output people couldn’t understand. When respondents couldn’t follow or judge what the AI had made, they also felt less responsible for it. Whether they told others they’d used AI had little to do with pride. It depended on the norms around them, and on a fear that an “AI-generated” label would erase their own contribution.
The sample is small and skews toward heavy users, so these are leads, not laws. The paper turns them into questions for designers: Would showing people a record of their own edits and decisions strengthen their sense of ownership? What would a disclosure look like that tells the story of the work instead of stamping a label on it?
This is the fellows’ paper. Megan Wei and Melanie Subbiah of Brown University, Audrey Lee of Carnegie Mellon University, and Annya Dahmani of UC Berkeley led the work. Ellie Pavlick, who heads ARIA, a US National Science Foundation institute that studies how people interact with AI assistants, joined Helen and Dave Edwards as co-authors.
We want to make the fellowship permanent. Academic research rarely produces ideas anyone can build, and corporate labs mostly chase more capable models. The fellowship sits in between: It starts with people’s stories and ends with designs that could change how AI gets made. In August the fellows spent a week together in Bend, Oregon, turning these findings into experiments and prototypes.
We’re raising money from foundations and individual donors to run the fellowship as an ongoing program of the Artificiality Institute, with ARIA as our academic partner. The money would pay fellows’ stipends, bring them to Bend, and fund the design work that follows. If you’d like to help, email us at hello@artificialityinstitute.org.
