OpenAI launches LifeVal test to measure AI on real-life tasks

OpenAI announced on November 2, 2026, LifeVal, an evaluation to measure AI assistants on day-to-day tasks. The tests cover caregiving, household management and government services. It is not a new consumer feature.

The pitch is straightforward: test AI on the work people actually want to hand off. The article says OpenAI used LifeVal internally over the past year. Task selection is based on time-use research and a survey described as demographically representative with 3,995 respondents.

From helpful answer to reliable outcome

The central example is enrolling a child in school. It sounds simple, but it requires checking the correct district, finding the deadline, identifying documents, validating the address and deciding whether the form should be submitted now or reviewed first. A good answer helps, but only a complete process solves the family's problem.

That is where the evaluation gets harder. Traditional AI tests tend to have a closed format, such as a question with an answer, a coding challenge with tests, or a document summary. Administrative routine is messy: it starts with a search, moves through a form, pauses for a family decision and only ends days later, when someone confirms acceptance.

Time, context and limits

To choose what to test, LifeVal draws on sources such as the American Time Use Survey, which measures how people in the United States spend their time, and the Multinational Time Use Study, which harmonizes diaries from several countries. Because the second study includes data from the first, the totals should not be treated as fully separate groups.

These sources show the shape of daily life, but alone they do not say what people want to delegate to AI, whether a task can be delegated safely, or whether the help truly improves life. An hour cooking can be pleasure for one person, a need to adapt recipes and costs for another, and a printed plan due to lack of internet for a third. Time saved does not measure urgency: checking a five-minute deadline can be worth more than an hour of leisure.

What we still do not know about the score

OpenAI's page leaves open the score denominator, the number of attempts and the best-result selection rule. An achievable peak does not show typical performance, variation across attempts, how often the person had to intervene, or what kinds of errors remain. Without a denominator and scoring rule, the reader cannot interpret the percentage precisely.

The article argues that a good report would show the correct-completion rate, partial progress and consequential errors separately. Nine correct steps followed by a payment to the wrong recipient is not a near miss. The full setup also matters: model, instructions, tools, accounts, environment, time allowed and human help, plus permissions, privacy, spending limits and when stopping to ask for approval is the right call.