What this video shows
The video presents LifeVal as a way to measure assistance with everyday work such as remembering, planning, comparing and following up. One example asks an agent to combine health information, dietary preferences and a budget, then act on a grocery site for a parent with diabetes.
OpenAI says it began with the American Time Use Survey and Multinational Time Use Study, then surveyed 3,995 respondents about where they wanted help. The resulting areas include caregiving, health, education, household administration, personal finance, shopping and government services.
Read OpenAI’s LifeVal methodology and examples before using the published task scores. Inspect the underlying American Time Use Survey and Multinational Time Use Study to understand the task-selection inputs.
What you will learn
- Choose evaluation tasks from work that people actually want help completing.
- Grade the finished state and the actions taken, rather than rewarding a plausible written plan alone.
- Include private-data handling, confirmation and recovery requirements in each task rubric.
- Treat vendor-published example scores as evidence about those examples, not as a promise for your users.
How to apply this safely
- Interview users and collect ten recurring tasks with observable outcomes.
- Write the starting state, allowed tools, private data, confirmation points and expected final state for each task.
- Run each task several times because browser state and model behavior can vary.
- Grade completion, factual support, unsafe actions, unnecessary disclosure, latency and recovery from tool failure.
- Require a person to confirm medical, financial, legal, purchase and irreversible actions.
Important limitations
- OpenAI built LifeVal and trained the models it evaluates, so independent replication is needed.
- The public article shows selected high-scoring examples but does not publish a complete task set, model-by-model table or enough detail to reproduce every score.
- A one-minute video summarizes the motivation and omits rubric design, scorer reliability and failure distributions.
- Everyday tasks can involve health, financial and family data, which raises privacy and consent requirements beyond task completion.
Sources to check
- LifeVal: Measuring AI progress on the tasks that matter in everyday life OpenAI video published November 6, 2026.
- LifeVal OpenAI research page published November 2, 2026 with survey background and selected tasks.
- American Time Use Survey U.S. Bureau of Labor Statistics time-use study used as a task-selection input.
- Multinational Time Use Study Cross-national time-use dataset used as a task-selection input.
Continue learning on Learnetto
Build an AI agent test set
Turn real workflows into cases with known outcomes and review rules.
Backtest AI agents
Evaluate historical decisions without future-data leakage.
AI evals guide
Choose frameworks, graders and production monitoring for agent systems.