OpenAI is letting some users try out a new ChatGPT feature that uses its artificial intelligence to use a web browser to book trips, buy groceries, search for bargains and do many other online tasks.
The new tool, called Operator, is an AI agent: it relies on an AI model trained on both text and images to interpret commands and figure out how to use a web browser to figure them out feed. OpenAI claims that it has the potential to automate many day-to-day tasks and workday assignments.
OpenAI’s Operator follows competing releases by both Google and Anthropic, which have demonstrated what is capable of using the web. AI agents are widely seen as the next evolutionary stage for AI after chatbots, and many companies have jumped on the hype train by naming them. In most cases, they are very limited in their capabilities and simply use a language model to automate things that are normally done with regular software.
“AI is evolving from this tool that can answer your questions to one that is also capable of taking action in the world, performing complex, multi-step workflows,” said Peter Welinder, VP of Product at OpenAI. “We will see a lot of impact on people’s productivity – but also the quality of work that people can do.”
OpenAI acknowledges that giving ChatGPT access to a web browser does present new risks, and it says that sometimes the operator misbehaves. It says it has implemented several new safeguards and plans to gradually expand the operator’s capabilities.
Welinder and Yash Kumar, product and engineering lead for OpenAI’s Computer Using Agent, say the plan is to learn from how people use the tool. They acknowledge that the tool can make unwanted bookings or purchases, but add that a lot of work has gone into ensuring it asks before doing anything risky. “It will get back to me and ask for confirmations before taking steps that could be irreversible,” says Kumar.
OpenAI also released a new “system map” today detailing the issues that can crop up with Operator. These include the potential for it to misunderstand commands or deviate from what a user is asking; to be abused by users; or being targeted by cybercriminals.
“It also poses an incredible amount of security challenges,” says Kumar. “Because your attack vector area and your risk vector area increases quite significantly.”
Operator will initially be available as a “research preview” to ChatGPT users with a Pro account, which costs a hefty $200 per month. The company says it plans to expand access as the tool slowly rolls out, because it will inevitably make some mistakes along the way.
In several demonstrations, Operator showed the potential for AI to play a more active role as a web helper. The tool features a remote web browser and a chat window to communicate with a user.
At WIRED’s request, the operator was asked to book an Amtrak train trip from New Haven, Connecticut, to Washington, DC. It went to the right website and entered the necessary information correctly to call up the grid, then asked for further instruction. If a user is logged in to the Amtrak website or to a browser profile with saved credit card information, the operator will be able to go ahead and book a ticket — though it’s designed to ask for permission first.
Kumar asked Operator to reserve a table at Beretta, a restaurant in San Francisco. The program went to the OpenTable website, found the right restaurant and looked up availability before being asked what to do next. OpenAI says it has partnered with a number of popular websites, including OpenTable, to ensure Operator runs smoothly on them.
The new tool is based on OpenAI’s GPT-4o AI model, which can perceive a browser and web page and speak in typed text. The tool includes additional training designed to help it understand how to perform tasks online. OpenAI will also make its compute usage agent available through its API.