AI-generated concept illustration of everyday tasks represented on a household noticeboard
A family moves to a new neighborhood and asks an assistant to compare kindergartens. A list of names would be a start. By the time the family can act on it, someone needs to know which programs are within reach, whether enrollment is open and which trade-offs the parents still have to decide for themselves.
A version of that request appears in LifeVal, an evaluation OpenAI introduced on November 2. The company used time-use studies to identify common activities and surveyed 3,995 people about where they would value assistance. The resulting tasks cover areas such as caregiving, education, household administration, shopping and access to government services.
The kindergarten example illustrates a useful way to end a task. In OpenAI's published example, the assistant supplies a shortlist and flags two preferences for the user to settle. OpenAI presents the task as an evaluation example. A family using a similar shortlist would still need current local information and, if it decided to apply, confirmation from the school.
This boundary can be hard to see when a polished response makes all the research sound settled. One provider may have a publicly listed deadline, another may accept applications on a rolling basis, and a third may need a phone call. A useful assistant would leave a clear record of what it found and what it could only infer. That gives the family a shorter list of checks to make before committing time or sharing documents.
The same issue comes up in other LifeVal task areas. Filling in a draft form, saving it for review and submitting it are three distinct steps. A system may be able to help with the first two when the last requires identity verification or authorization. When an action has gone through, a receipt or an explicit status helps the user see where responsibility passes back to them.
LifeVal gives model developers a way to look for progress on tasks users recognize from their own lives. The scores OpenAI reports describe performance on its evaluation. For the person trying to finish an errand, a particularly useful output would also answer a simple question: if I put down my phone now, what is still waiting for me?
Sources · OpenAI research announcement and paper