Loading page…
Users helped supply tasks, personal context and grading rules. The internal evaluation measures a harder-to-define goal than answering correctly: getting a useful result for a particular person.
Listen to this story
LifeVal turns OpenAI’s personal-assistant ambitions into an internal yardstick: people with relevant experience write tasks, provide personal context, and define what counts as success. OpenAI says it has used the benchmark for a year to guide model development in areas including browser use, computer use, and service connections; its November 2, 2026 introduction makes the effort public. The benchmark covers household and personal responsibilities, while examples show that some choices still require user review.
OpenAI surveyed 3,995 demographically representative respondents to identify where people want help across areas including caregiving, health, finances, and government services.
About 100 users wrote tasks and grading criteria; authors had to show relevant experience and understand what a satisfactory result would look like.
Unlike GDPVal’s work-focused assignments, LifeVal tests assistance with personal responsibilities such as shopping, budgeting, and finding childcare.
OpenAI introduced LifeVal on November 2, 2026, bringing household errands and personal decisions into its evaluation of AI progress. The internal benchmark covers tasks such as shopping, budgeting and finding childcare. Its focus is assistance that fits someone’s circumstances—not simply an answer to a question—and OpenAI says it has already used the evaluation to guide a year of model development.
The starting problem was deciding what useful help should look like. OpenAI wanted to identify everyday responsibilities people would actually want to hand to an assistant. Researcher Jenny Yang told CNN that the evaluation is meant to reveal both improvement and areas where the company’s models still fall short.
OpenAI began with the American Time Use Survey and the Multinational Time Use Study. Those datasets catalog how people spend their time across demographic groups, times of day and seasons. They supplied a starting inventory of activities, rather than a ready-made list of jobs for an AI assistant.
Time spent was not enough to determine demand. As CNN described the process, activities such as sleeping and eating ranked prominently but offered little for an assistant to do. OpenAI therefore asked people where they wanted help. The company says its new survey included 3,995 respondents and was demographically representative.
The resulting evaluation spans caregiving, health, education, household administration, personal finances, shopping and government services. That gives LifeVal a different remit from GDPVal, OpenAI’s earlier evaluation of work-related skills across industries. Here, the assignments concern managing a household and navigating personal responsibilities rather than producing work for an employer.
A smaller group of about 100 users then wrote down tasks they wanted AI to perform, according to CNN’s reporting. They supplied the personal context an assistant would need and helped develop grading rules for judging success. The request and the definition of a satisfactory result came together, rather than treating the prompt alone as the whole test.
OpenAI also screened task authors for relevant experience. Someone proposing a financial-planning assignment, for example, needed experience working with a financial adviser for multiple years. Yang told CNN that participants had to demonstrate that they understood their proposed task and knew what the right answer would look like for them.
The skills under examination include navigating websites, finding relevant information in a user’s chat history and reaching conclusions that match that person’s needs. CNN notes that everyday-task results can be subjective and can change with the information supplied. LifeVal’s user-written context and grading rules address that difficulty by specifying what success means for each request.
To be a truly helpful personal assistant, ChatGPT should be able to take actions that actually get things done.
OpenAI, LifeVal research publication
LifeVal’s public introduction follows its use inside OpenAI to track and improve model performance over the past year. The company connects that effort to training in browser use, computer use and connectors—connections to other services. It says those capabilities are helping move ChatGPT from a conversational tool toward a personal assistant that can act.
The published examples show how several requirements can sit inside one seemingly simple errand:
The examples also leave room for human decisions. In the kindergarten workflow, the assistant presents a shortlist but says two preferences still need the user’s call. In the diabetes-care example, it offers a first-week plan for review before sharing it with the parent. Those examples show assistance that prepares choices and asks for review, alongside the company’s ambition to complete tasks.
Loading discussion...
Join the conversation
Explain where personal preferences should outweigh a common grading standard.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.