Mixed-initiative systems AI
In a mixed-initiative system, useful actions can begin with either a person or the software. The person need not specify every step, but automation does not take over the entire activity. The design problem includes when a system should offer help, ask a question, act, or leave the person to continue.
Eric Horvitz’s 1999 work illustrates this through Lookout, a scheduling assistant that inferred possible intentions from messages. Initiative here concerns how assistance enters an ongoing task, rather than simply whether a task can be automated.
In a research environment, a person might introduce a question, a model suggest a comparison, and the person revise the proposed method before conventional software runs it. This is an illustrative application of the idea. Calling a project mixed-initiative does not identify a particular software framework or establish that its division of responsibilities works well.
Later work such as AI Chains makes intermediate language-model steps and results editable. It offers a concrete example of keeping participation possible within a multistep process, rather than restricting the user to an initial prompt and a final answer.
Further reading
- Eric Horvitz, Principles of Mixed-Initiative User Interfaces (1999), including the Lookout scheduling example.
- Tongshuang Wu, Michael Terry, and Carrie J. Cai, AI Chains (2022), on editable steps and intermediate results in language-model interactions.