OpenAI launches LifeVal benchmark to test AI on everyday tasks

OpenAI has released a new benchmark called LifeVal. It measures whether AI models can truly help ordinary people with daily tasks, aiming to make general AI more useful in everyday life.

The tasks were designed using data from surveys like the American Time Use Survey, plus input from 3,995 demographically representative respondents. LifeVal covers areas such as caregiving, health, education, family admin, personal finance, and government services. Typical tasks include choosing a kindergarten for a child or making a care plan for a family member with diabetes.

Unlike traditional benchmarks, LifeVal focuses on whether tasks reach an actionable, deliverable state. Models need to understand chat records, pull together information, and give advice that fits each person's situation. Public examples show finding a kindergarten scored 93.1%, organizing flood volunteers hit 98.0%, and a diabetic meal plan reached 90.4%.

However, LifeVal is still mostly an internal evaluation by OpenAI. The full task set, scoring code, and a reproducible leaderboard are not available to the public, so the results should be read with caution.