After Kumar prompted Operator to use OpenTable to book a table at San Francisco restaurant Beretta around 7 p.m., the tool opened a remote browser window, went to OpenTable’s website and searched for the restaurant — but couldn’t initially find it. As it turned out, OpenTable was set to search for restaurants in Iowa, not California. But Kumar had previously instructed Operator to search within a certain San Francisco zip code for relevant queries, so on its own, the tool switched to searching OpenTable in San Francisco and then shared a reservation for him to approve.
“We see a lot of potential for how this can evolve from small things to medium things to large things,” Kumar told Bloomberg News. He personally has been using Operator to do his grocery shopping and to book reservations at tennis courts.
Operator can also carry out multiple tasks simultaneously. For example, a user might prompt the service to find a hotel in Vancouver that has Peloton bikes in its gym and then, before it has finished, ask Operator to find an American Girl doll bed on Craigslist. As long as a user has confirmed that Operator can carry out a transaction — such as buying a pair of leggings from an online store — and has input any required credentials and payment information, the tool should be able to complete a purchase, Kumar said.
OpenAI plans to eventually release the AI model behind Operator for developers to use to build their own agents.
With those capabilities, however, AI agents like Operator also present new safety and security risks, given the potential for AI to make mistakes or be misused by bad actors. It’s one thing for a chatbot to spit out an inaccurate response to a question about a historical event and quite another for an agent to make mistakes with someone’s credit card.
OpenAI said Operator is meant to turn down some tasks, such as actions involving banking and anything the company considers harmful. There are a number of actions the tool will not complete, and will instead alert a user to carry out, including logging in to websites, providing payment information and filling out CAPTCHAs, the company said. Operator should ask a user for approval before it does things like ordering something online. For some tasks, such as writing emails, the service will require a user to supervise it, OpenAI said. Users can also take control of any tasks that are in process and pause them if need be.
“The user should always feel they’re in control,” Kumar said.
