Giving the command is not owning the outcome
Two experiments separated a person's sense of control over a computer's first move from their judgment about its final effect. A person may feel they started the work without seeing where it ended. The paper does not test safety in real AI systems. Show the final result before approval.
What the source reports
Johanna Didion, Diego Garaialde and David Coyle ran two experiments about instructing a computer. Participants felt a direct sense of control over its first action. Their considered judgment about causing a later, meaningful result was different. In some cases, people still judged themselves responsible for that result. Feeling in control at the start is not the same as seeing and approving the end. The digest suggests a draft-only check: compare the computer's first action with its final work and name who may approve it. That is an editorial use of the finding. The paper's abstract does not show that such a review step makes a live AI system safer.
Original source
- Title
- Judgement and the Feeling of Agency When Instructing a Computer to Act
- Author
- Johanna Didion, Diego Garaialde, David Coyle
- Publication
- arXiv
- Date
- Tuesday, October 6, 2026
Also covered by the other newsletter.