OpenAI puts household chores on AI’s test paper

A test built around the jobs people want help with

OpenAI says it drew on American and international time-use studies, then surveyed 3,995 people about where they would value a personal assistant. The resulting tasks cover caregiving, health, education, finances and household administration. The company says it has used LifeVal internally over the past year to measure and improve its models.

The published examples make the ambition concrete. A school-search demonstration compares nearby kindergartens and checks enrollment requirements; its final message presents a shortlist and leaves two preferences for the parent. Another example finds a home and adds a deed to a cart for review. These are different stopping points in a job, with human decisions still visible.

What a score can tell you

The chart notes say LifeVal averages attempts within each of 169 tasks, then gives each task equal weight in the overall result. Its percentage labels express an average reward, the evaluation’s scoring measure. They should not be read as the percentage of families whose real errands ChatGPT will successfully finish.

The announcement shows the direction OpenAI wants to measure: useful actions across websites, computer interfaces and connected services, beyond a good written answer. But the inspected material does not provide the complete grading rubrics and execution records needed to independently reproduce the headline results. We have not rerun LifeVal, and the announcement does not establish availability for each demonstrated workflow in Belgium or another particular market.

A test of everyday assistance asks a worthwhile question. The next useful evidence is how the same jobs hold up outside the developer’s evaluation, including where the assistant stops and how much checking the person must still do.