LifeVal tests whether AI assistants can turn personal context into completed tasks

Testing assistance beyond conversation

LifeVal asks whether AI models can complete consumer tasks that depend on a particular person’s needs. In a preprint submitted on November 4, 2026, Jenny Yang and colleagues at OpenAI describe a benchmark combining personal information with tasks requiring browser and computer use.

The distinction matters: recommending an action is different from carrying it out. LifeVal tasks can involve navigating websites, filling forms, creating files and changing account states. Their grading criteria are intended to capture what successful completion means for the person who supplied the task, rather than whether an answer sounds generally helpful.

The reported evaluation contains 169 tasks and tests GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 and GPT-6 Astra.

Choosing what people value

The researchers began with American and multinational time-use datasets, then supplemented them with a U.S. survey about a hypothetical perfect digital assistant. Their rationale was that time spent is not necessarily a good measure of value: an infrequent decision about insurance or taxes can be consequential, while a digital assistant cannot perform physical activities such as eating or sleeping.

The survey recruited against demographic targets for age, gender, ethnicity and income. Although 3,837 people completed it, the appendix reports that only 1,826 passed all five comprehension and attention checks and were retained for analysis. Respondents valued assistance across subsets of up to 30 activities, considering what they would have paid over 2025. Purchases arranged by the assistant would remain their own expense.

The authors estimate approximately $1.7 trillion in annual U.S. willingness to pay across 148 activity categories. This is a survey-based valuation of hypothetical perfect assistance—not observed spending, revenue or measured savings from existing AI. The categories guided task collection; they should not be confused with the 169 individual evaluation tasks.

Personal context and evidence of completion

Task writers had to demonstrate relevant experience, including completing the activity repeatedly and seeing its results in their lives for more than a year. They supplied requests, reference files, additional context and grading rubrics, alongside consented, de-identified ChatGPT histories and memories. Personally identifiable information in tasks was replaced with synthetic data.

Each task also includes a starting state and a software environment where the model can act. Tasks passed at least three human review stages: peer review, lead review and final expert sign-off. Models additionally attempted tasks and proposed corrections, which human reviewers assessed.

Two automated graders use GPT-5.6. The primary grader examines the response, generated artifacts and changes in the environment’s underlying state. It is instructed to verify that claimed work actually exists. A separate grader checks for access to internal environment APIs, which the benchmark treats as hacking; detected cases receive zero.

A limited human comparison provides some support for automated scoring. Across 30 GPT-5.5 task attempts, mean human and automated rubric grades were 83.61% and 84.20%, respectively, with a correlation of 0.864. That check does not establish grading reliability across every task or model.

Conversation history remains a bottleneck

The personalization experiments offer a particularly revealing result. Removing additional task context substantially reduced performance at every tested reasoning effort. Removing past conversations and memories had a much smaller effect. Moving the additional context into conversation histories reduced performance by approximately as much as removing it entirely.

The authors interpret this as evidence that additional context carries much of the necessary personal information, while models struggle to retrieve and use it from past conversations.

For rubric criteria attributed to a single personal-information source, reported success rates were 33.33% for past conversations, 47.8% for memories and 55.8% for additional task context. The latter fell to 41.5% when that information was inserted into conversations. These are criterion-level results, not overall task-completion rates.

Reviewing perfectly scored conversation-attributed criteria revealed another caution: the needed information was also available elsewhere or inferable from conversation titles. Success therefore did not demonstrate retrieval from conversation contents.

What the benchmark does not establish

The authors report persistent errors involving incorrect information or recommendations, including missed details in personal conversations. These findings concern performance in the benchmark, not demonstrated long-term improvements in consumers’ lives.

Coverage is also constrained by historical activity classifications and a U.S.-centered, English-language valuation survey. Self-reported willingness to pay depends on framing and respondents’ expectations of a flawless assistant. Although most interactions occur in controlled environments, the browsing cache is not frozen, so some scores can vary between runs.

LifeVal provides a structured test of personalized action. Its results also underline a practical distinction: possessing someone’s history is not the same as reliably finding the details needed to act on their behalf.