Artificial Intelligence
6 min read
ChatGPT just came out with its own web browser. Use it with caution.
OpenAI’s Atlas promises AI-powered convenience. The price? Letting ChatGPT track and store “memories” of what you do online.
Column by Geoffrey A. Fowler
The maker of the world’s most popular chatbot, ChatGPT, launched a web browser this week that promises to make surfing the internet smarter. In exchange, ChatGPT Atlas wants permission to watch — and remember — everything you do online.
The browser from OpenAI out-surveils even Google Chrome, and that’s saying something. It doesn’t just log which websites you visit; it also stores “memories” of what you look at and do on those sites. It can even grab control of your mouse and browse for you.
It’s too early to evaluate whether Atlas’s new artificial intelligence capabilities are useful enough to make it worth all the data gathering. But the implications for privacy are vast, and the controls for managing what Atlas remembers are confusing at best.
There’s a lot at stake in which browser you choose. It’s your daily portal to the internet — a spigot of lucrative information that companies can use to target you with ads, steer you toward certain sites, and train AI on your behaviors and interests.
OpenAI isn’t alone in trying to reinvent browsers with AI. Search engine Perplexity makes a browser called Comet, and Google has in recent months added its Gemini bot to Chrome — and says it will soon also add agent capabilities that let AI do tasks for you. (The Washington Post has partnerships with OpenAI and Perplexity.)
So what does Atlas actually do? It replaces Google with ChatGPT as the main source to find websites and information. An “Ask ChatGPT” button in the upper right lets you chat with the bot about pages you’re viewing. For instance, you could ask it to summarize an article or analyze data. And it puts ChatGPT one click away for tasks such as revising email drafts.
“The goal was to make it easier for ChatGPT to work alongside you as you browse the web,” said Adam Fry, OpenAI’s Atlas product lead.
But reviewing Atlas’s privacy practices and controls revealed something darker. Behind the scenes, it is working to learn much more about you. If you grant permission during setup, the browser builds a trove of memories about sites you visit and surfaces them “when you need” them. You could tell Atlas: “Open the Halloween decorations I was looking at last week in some tabs,” and it could do it.
Atlas remembers not just website addresses, but “facts and insights” from the sites themselves based on summaries of the content OpenAI makes on its own servers. It might remember you have a trip coming up, prefer Delta Air Lines and use Google Calendar, Fry said.
These memories shape your experience across the browser. ChatGPT tailors its responses to your Atlas memories in future chats. And the browser’s home screen offers personalized suggestions of things you should do next, such as “find a vegetarian recipe.”
(By contrast, Google’s Chrome doesn’t let Gemini store memories about the contents of webpages. But Gemini can — if you ask it to — answer questions about your web browsing history.)
That level of personalization brings privacy risks that are hard to understand, much less control. There are things you might want an AI to remember and bring up again in the future — and things you definitely wouldn’t, such as relationship troubles or that embarrassing medical condition you researched at 2 a.m.
The details of what Atlas will or won’t remember get confusing fast. OpenAI says memories could include tasks you’re working on and your preferences, but not entire page contents. Atlas isn’t supposed to remember government IDs, bank account numbers, addresses, passwords, medical records and financial information. It also isn’t supposed to attempt to remember the content of adult websites.
However, a test found that Atlas kept memories about registering for “sexual and reproductive health services via Planned Parenthood Direct,” according to Lena Cohen, a staff technologist at the Electronic Frontier Foundation. It also kept a memory about the name of a real doctor. “The extensive data collection in the Atlas browser could be a privacy nightmare for users,” she said. (OpenAI said it improved its systems after I reported this to the company.)
You can control Atlas’s memories, but it takes effort. In the address bar, you can tell Atlas not to remember certain websites. Clearing your browsing history is supposed to clear memories from that period. In settings, Atlas shows a list of its memories, which you can delete individually or all at once. But this memory file is separate from the one ChatGPT already keeps about you.
Atlas offers an “incognito” mode that doesn’t add to your history or memories. But like Chrome’s misleadingly named mode, it doesn’t actually hide you from other websites or even from ChatGPT itself.
Companies love to claim that they give users control over data collection. But offering scattered privacy settings isn’t the same as meaningful control. A 747 aircraft has lots of controls, too. But that doesn’t mean just anyone can fly it.
“We are using memory specifically to enhance the product features of Atlas — not for anything else,” Fry said. “The way we use data here is pretty different from social media companies that build interest profiles.”
OpenAI also does not have an advertising business, he said. Yet.
Online surveillance can have serious consequences: Will governments be able to ask OpenAI to hand over people’s browsing data and memories? What if they’re researching activities that are illegal in certain states, such as abortion? After I initially published this story, OpenAI told me it would only disclose user data through “valid” legal processes or in an “emergency situation.” It also said browser memories are held on its servers for 30 days and then deleted.
In one area, Atlas takes a more conservative approach. By default, it doesn’t use the contents of your browser for training its AI — though there’s a setting to turn that on if you want. (This is separate from whether your ChatGPT conversations can be used for training, which does happen by default.)
Atlas also introduces new risks by letting AI operate your browser for you. These agents can go on shopping hunts or, as I wrote recently, help you quit subscriptions. But AI agents are still prone to mistakes — and when an agent has access to a browser with your log-in credentials and payment info, that’s a lot of power to hand over.
Fry said Atlas lets people use agent mode in a completely cleared-out version of the browser. And certain high-risk tasks involving money happen only in a mode where the user is supposed to monitor what the AI is doing.
Browsers have evolved from simple windows to the web into sophisticated data-collection tools. Atlas might make browsing more convenient — but only if you’re comfortable letting AI deep into your life.
Advertisement
