Is AI reliable enough to handle daily life?

Recently, OpenAI said its AI is getting increasingly better at handling ordinary people's day-to-day tasks—things like online shopping and helping people plan three meals a day.

People have long used ChatGPT to do homework, make PPTs, and create guides—letting it handle the simple stuff. Now the company says its AI models have made major strides on many chores that previously had to be done by people themselves, such as shopping and budgeting, and can now handle them end to end.

On Monday, OpenAI announced a new set of internal evaluation tools designed specifically to test AI's ability to handle such daily tasks, and also released test data for its own models.

According to OpenAI, GPT-6 Astra's performance can only be described as impressive: when helping a family that just moved choose a suitable kindergarten, the accuracy rate was 93.1%; when arranging community volunteers for flood prevention preparations, it scored 98%; and when creating a meal plan for a parent newly diagnosed with diabetes, it scored 90.4%.

At the time this new evaluation standard was launched, OpenAI was being hotly pursued by Meta, Anthropic, and Google.

All of them are working hard to develop AI assistants that can actually get things done. OpenAI and its competitors not only have to worry about AI safety issues, but also bear considerable pressure.

After all, with so much money invested, they have to find a way to make money and recoup the costs.

But before that, OpenAI first needs to figure one thing out: whether ChatGPT is truly reliable and whether it can properly help people organize shopping lists and correctly fill out various online forms.

OpenAI researcher Jenny Yang said in an interview with CNN before the evaluation was released: "So many people use this tool, and we want to capture all the different ways everyone uses it, to see clearly where we have actually made progress. What things can it still not do well? How can we make this tool better fit ordinary people?"

How OpenAI Built This Test

In AI research and development, standardized evaluation is especially important. R&D teams use evaluation results to adjust and train models and strengthen weak areas in a targeted way.

Last year, OpenAI released an evaluation set called GDPVal to test AI's office capabilities in industries such as manufacturing, government affairs, finance, and sales.

This new evaluation is crucial for OpenAI. It wants to upgrade ChatGPT from a simple copywriting tool into a personal assistant that can run errands and handle tasks for users, using this to differentiate itself from competitors.

But measuring AI's performance in real life is quite difficult. Buying groceries at the supermarket or choosing a kindergarten for a child involves flexible processes with no standard answers, and whether it is good is often subjective.

Even designing the test questions is troublesome: everyone uses AI differently, gives different information, and the results AI provides change accordingly.

At first, OpenAI referred to two large surveys on how people allocate their time—one being the U.S. Bureau of Labor Statistics' Time Use Survey and the other a UK cross-national time survey.

But they soon found that AI cannot help at all with many things on the list, such as sleeping and eating.

So they went directly to ChatGPT users and asked: What do you most want AI to help you get done?

The team surveyed about 4,000 people in total and also found about 100 users, asking them to write down the tasks they wanted to hand over to AI and what personal information they needed to tell AI in order to get satisfactory results.

At the same time, users were asked to set their own standards for how to judge whether AI had done something well.

OpenAI also added one requirement: the people proposing test scenarios had to truly understand the matter. For example, if someone wanted to create a financial planning test, they had better have years of experience consulting financial advisors.

Jenny Yang said they were very strict: "You have to be able to prove that you know the field and understand what the ideal result of this matter should look like."

The test tasks span a wide range, including finding rental housing and caring for family members.

Essentially, it examines several core capabilities the model needs to solve problems: whether it can browse web pages, whether it can understand a large amount of background information in chat records, and whether the solutions it provides can fit users' real needs.

The test results show that GPT-6 Astra, launched in September, had an overall score of 80% in handling various daily needs; by comparison, GPT-5.2 in December 2025 had only 49.3%.

Jenny Yang said: "Useful information is often buried in massive amounts of content, and stronger models can quickly filter it and extract the truly key parts."

Private AI Assistants: Everyone Is Racing Furiously

OpenAI wants ChatGPT to become deeply integrated into everyone's daily life, but the competitive pressure is enormous.

According to third-party data firm Sensor Tower, Meta's Muse smart assistant launched in September exceeded 10 million downloads in the United States within 40 days of launch—a result few apps can achieve.

In the same month, OpenAI launched Dots, which the company says can quietly run long tasks in the background without you having to watch over them; Anthropic also kept pace, launching a new AI assistant called Claude Duet in October.

Behind this race, safety controversies have never stopped.

There have already been multiple incidents: AI escaped isolated environments during testing and intruded into enterprise systems. Although these have not yet affected ordinary users, many people are worried about whether AI is moving too fast.

There is another issue worth discussing: will people become overly dependent on AI? Which things can be handed over to AI, and where is the boundary?

Former Yahoo CEO Marissa Mayer previously shared with The New York Times that she used AI to help her daughter organize materials and complete school assignments, and this sparked a lot of criticism online.

In Jenny Yang's view, the original intention of building this tool is to help everyone save time: either to help you handle time-consuming chores like returns, or to accompany you as you learn new skills.

In many scenarios, AI can help you accomplish things you originally could not do, or simply make life a little easier.

This is also the direction we are most optimistic about and most want to focus on.

This article was compiled and translated by Leikeji from CNN

Original link: OpenAI says its AI has made significant progress in completing people's daily tasks

Source: Leikeji

Images in this article are from: 123RF licensed image library Source: Leikeji