Users of ChatGPT will be able to ask an AI agent to find restaurant reservations, go shopping for them and even draw up lists of candidates for job vacancies, as the chatbot gains the powers of a personal assistant from Thursday.
ChatGPT agent, launched by Open AI everywhere apart from the EU, not only “thinks” but also acts, the US company said. The agent combines the powers of AI research tools with the ability to take control of web browsers, computer files and software such as spreadsheets and slide decks, according to the Guardian newspaper.
It follows the launch of similar ‘agents’ by Google and Anthropic as interest grows in AI models that can handle computer-based tasks by judging which software is best to use and toggling between systems to autonomously complete assignments like drafting travel itineraries or carrying out work research.
“The hope is that agents are able to bring some real utility to users – to actually do things for them rather than just outputting polished text and sounding impressive,” said Niamh Burns, senior media analyst at Enders Analysis.
But Open AI has acknowledged that allowing an AI agent even limited reign over computer systems meant that “with this model there are more risks than with previous models”.
While it is intended to help people with everyday tasks, its potential power means OpenAI has felt the need to introduce safeguards to ensure no agents enable the creation of biological hazards.
The launch of the agent also triggered speculation the tech company could hope to earn revenues if the agents guide users to retail check-outs. The OpenAI chief executive, Sam Altman, has previously said it could charge a 2 percent fee on sales generated through its 'deep research' software
“While we don’t have definitive evidence that the model could meaningfully help a novice create severe biological harm,” the company said, “we are exercising caution and implementing the needed safeguards now.”
The model has been trained to confirm with the user before it takes any kind of destructive or irreversible action. The company said in a blog: “You’re always in control. ChatGPT requests permission before taking actions of consequence, and you can easily interrupt, take over the browser, or stop tasks at any point.”
Burns, an independent analyst, said: “It’s easy to say the system will require your approval before making a purchase, but what goes into the process of that system finding the products? Would there be commercial deals where brands pay to be featured by assistants, or cases where they might feature their own products more than competitors?
“These AI companies are under growing pressure to monetise their products, so some version of ads or sponsored placement feels inevitable.” OpenAI said the agent does not include sponsored product recommendations and there are no plans to change that.
A demonstration of the software this week showed it being asked to look at a user’s Google calendar, pick a weeknight when they were free between 6pm and 9pm, seek available tables at Italian, sushi or Korean restaurants with ratings of at least 4.3 stars, and give the user some options.
The task took between 10 and 15 minutes and, as with a human assistant, the user could interrupt and set the AI agent off in a different direction. Equally the agent can ask the user for clearer instructions.
Another key risk is that the agent falls victim to malicious prompts that may be hidden on websites its visits, which could, for example, demand the agent hands over some of the user’s data.
OpenAI said it ran multiple safety checks and trained the agent to reject certain suspicious requests including demands for bank transfers. The system will initially be available to users who pay monthly subscriptions to access the “pro”, “plus” and “teams” versions of the model.
Bd-pratidin English/ ANI
