OpenAI launches ChatGPT-powered web browser. What to know before downloading. – USA Today
OpenAI has launched its first internet browser powered by ChatGPT, but while it may make surfing the web easier, some are worried about privacy risks that everyday users may overlook.
OpenAI announced the internet browser, ChatGPT Atlas, on Tuesday, Oct. 21. The free browser functions similarly to Google Chrome, Firefox or Safari but has ChatGPT built into it. This artificial intelligence core can remember more browser history, allowing users to retrieve past searches or information.
Atlas also works with ChatGPT agent, a feature OpenAI launched earlier this year that allows the AI to take over a user’s computer and follow specific tasks. For example, a user can open a retailer’s website and ask ChatGPT to add certain items to a cart, purchase them and have the items delivered to their home.
While the new capabilities can be exciting for seasoned AI users, Lena Cohen, staff technologist at the Electronic Frontier Foundation, told USA TODAY that users should think twice before using Atlas as a full-time browser.
“Even though users might already be sharing sensitive data with OpenAI when they use ChatGPT, they likely have a better sense of the information they’re handing over. When you type out a question to ChatGPT, you’re in control of the information you’re sharing with them,” Cohen said. “On the other hand, when you’re using the Atlas browser for day-to-day tasks, it’s easier to forget that all of your activity could be sent to OpenAI.”
Here’s a more in-depth look at what ChatGPT Atlas offers and what precautions new users should take.
In many ways, ChatGPT Atlas functions similarly to a traditional internet browser. Users can type a URL, a question or a term into the main search bar for search results. But Atlas also offers new, unique features:
OpenAI paid Plus, Pro and Business subscribers may also try ChatGPT Atlas’s agent mode. When using agent mode, users give ChatGPT unique access to their browser.
“ChatGPT can complete end to end tasks for you like researching a meal plan, making a list of ingredients and adding the groceries to a shopping cart ready for delivery,” an OpenAI blog post states. Users are “always in control” of ChatGPT, even when it’s in agent mode, OpenAI says.
When using agent mode, the AI cannot run code in the browser, download files, install extensions, or access browsing history, saved autofills for passwords and other sensitive information or a computer’s apps or files.
For launch, ChatGPT Atlas is only accessible on Mac systems running macOS 12 Monterey or later, as per OpenAI’s blog post.
Users must download Atlas from the OpenAI website at chatgpt.com/atlas/.
To bring saved bookmarks, passwords and browser history from another browser, such as Google Chrome, users must import data from that browser to ChatGPT Atlas. OpenAI provides step-by-step instructions for this process on its blog.
One of the main concerns technology experts have with ChatGPT Atlas is its promoted browser memories, which can save and remember information from visited websites. According to OpenAI, this feature must be activated by users and can be toggled off.
OpenAI claims ChatGPT Atlas does not store “personally identifiable information,” such as medical and financial information, government IDs, Social Security numbers, online credentials and addresses. But Cohen isn’t so sure.
Testing ChatGPT Atlas herself, Cohen took several quizzes on The Washington Post website, using, of course, fake personal and financial information. When she checked Atlas’s browser memory after completing the quizzes, Atlas didn’t regurgitate the “sensitive” information she provided but did acknowledge how many of the quizzes she finished.
“It’s not clear to me if OpenAI received my answers to the quiz, but based on the fact they say, ‘They’ve completed most quiz questions,’ I know that Atlas is sending more than just the blank quiz.”
Cohen tried a similar test, scheduling a fake appointment on a Planned Parenthood website on the ChatGPT Atlas browser. Upon checking the browser’s memories, she was once again informed in-depth about her actions on the website.
“Part of what users should be cautious about the Atlas browser’s privacy promises are not backed up with many details or evidence,” she said. “We have to take OpenAI at their word that they’re filtering out sensitive data. Really, really basic, quick tests on medical and financial websites show that more information is being sent to OpenAI’s servers than many users would probably be comfortable with.”
Ultimately, Cohen recommended users instill critical thinking before launching the browser and enabling certain features.
“I think that users should think twice before handing over their entire browsing activity to OpenAI or any other company for that matter,” Cohen said.
Greta Cross is a national trending reporter at USA TODAY. Story idea? Email her at gcross@usatoday.com.
