On November 2, OpenAI introduced LifeVal, an evaluation of how AI models help with everyday tasks. The company says it has used the evaluation to track the development of its models over the past year. The announcement describes an evaluation and illustrates it with examples. It does not announce a service that already completes all of those tasks for users.
How the tasks are selected
The selection starts with two time-use studies, one covering the United States and one international. OpenAI added a survey of 3,995 respondents to identify activities where people would most value help. Published areas include caregiving, education, shopping and access to public services. The distinction between observing habits and asking about desired help has a practical consequence. A frequent activity may be easy, while an occasional administrative task may demand more time and attention.
A useful outcome needs evidence
When searching for a home, for example, a persuasive description is insufficient. A useful result would point to a current listing, show which requirements it meets and flag details that still need confirmation. If the task involves a form, finding it, filling it in and actually submitting it are separate milestones. These examples illustrate our editorial approach to evaluating an assistant. They are not results from our own LifeVal test.
A team deploying an agent could assess the accuracy of the information it finds, the final outcome and whether the assistant respects its permissions separately. For example, the team can test whether the assistant pauses for human review when the next step would place an order or send a document. This gives a useful draft credit for the work completed without counting it as a finished action.
How to read the results
In the public announcement, OpenAI provides scores for selected examples. These are company results. The published percentages do not on their own provide a general probability of success for an arbitrary task. Comparing assistants would require comparable conditions, clear grading rules and a breakdown of errors. A team working in Bulgaria would also need to test local forms, websites and language before relying on those results for its own workflow.
*Cover: conceptual AI illustration.*
