OpenAI, the company behind ChatGPT, released their own browser called Atlas, and it actually is something new: the first browser that actively fights against the web. Let’s talk about what that means, and what dangers there are from an anti-web browser made by an AI company — one that probably needs a warning label when you install it.

The problems fall into three main categories:

1. Atlas substitutes its own AI-generated content for the web, but it looks like it’s showing you the web
2. The user experience makes you guess what commands to type instead of clicking on links
3. You’re the agent for the browser, it’s not being an agent for you

1. By default, Atlas doesn’t take you to the web

When I first got Atlas up and running, I tried giving it the easiest and most obvious tasks I could possibly give it. I looked up “Taylor Swift showgirl” to see if it would give me links to videos or playlists to watch or listen to the most popular music on the charts right now; this has to be just about the easiest possible prompt.

The results that came back looked like a web page, but they weren’t. Instead, what I got was something closer to a last-minute book report written by a kid who had mostly plagiarized Wikipedia. The response mentioned some basic biographical information and had a few photos. Now we know that AI tools are prone to this kind of confabulation, but this is new, because it felt like I was in a web browser, typing into a search box on the Internet. And here’s what was most notable: there was no link to her website.

I had typed “Taylor Swift” in a browser, and the response had literally zero links to Taylor Swift’s actual website. If you stayed within what Atlas generated, you would have no way of knowing that Taylor Swift has a website at all.

Unless you were an expert, you would almost certainly think I had typed in a search box and gotten back a web page with search results. But in reality, I had typed in a prompt box and gotten back a synthesized response that superficially resembles a web page, and it uses some web technologies to display its output. Instead of a list of links to websites that had information about the topic, it had bullet points describing things it thought I should know. There were a few footnotes buried within some of those response, but the clear intent was that I was meant to stay within the AI-generated results, trapped in that walled garden.

During its first run, there’s a brief warning buried amidst all the other messages that says, “ChatGPT may give you inaccurate information”, but nobody is going to think that means “sometimes this tool completely fabricates content, gives me a box that looks like a search box, and shows me the fabricated content in a display that looks like a web page when I type in the fake search box”.

And it’s not like the generated response is even that satisfying. The fake web page had no information newer than two or three weeks old, reflecting the fact that LLMs rely on whenever they’ve most recently been able to crawl (or gather without consent) information from the web. None of today’s big AI paltforms update nearly as often as conventional search engines do.

Keep in mind, all of these shortcomings are not because the browser is new and has bugs; this is the app working as designed. Atlas is a browser, but it is not a web browser. It is an anti-web browser.

2. We left command-line interfaces behind 40 years ago for a reason

Back in the early 1980s, there was a popular game called Zork that was in a category called “text adventure games”. The computer would say something like:

You are in a clearing in the forest, there is a rock here.

And then you would type:

Take the rock

And it would say:

Sorry, I can't do that.

So then you would type:

Pick up the rock.

And then it would say:

You have the rock.

And it would go on like this for hours while you tried in vain to guess what the hell it wanted you to type, or you discovered the outdoors, whichever came first.

There were a tiny handful of incredible nerds who thought this was fun, mostly because 3D graphics and the physical touch of another human being hadn’t been invented yet. But for the most part, people would tire of the novelty because trying to guess what to type to make something happen is a terrible and exhausting user interface. This was also why people hated operating systems like MS-DOS, and why even all the Linux users reading this right now are doing so in a graphical user interface.

Clicking on things is great, because you can see what your choices are, and then just choose the one you want. Tapping on things on a touch screen is even better. And this kind of discoverability was one of the fundamental innovations of the web: It democratized being able to create a clickable list of options once anybody could make a web page.

In the demo for Atlas, the OpenAI team shows a user trying to find a Google Doc from their browser history. A normal user would type keywords like “atlas design” and see their browser show a list of recent pages. They would recognize the phrase “Google Docs” or the icon, and click on it to get back to where they were.

But in the OpenAI demo, the team member types out:

search web history for a doc about atlas core design

This is worse in every conceivable way. It’s slower, more prone to error, and redundant. But it also highlights one of the biggest invisible problems: you’re switching “modes”. Normally, an LLM’s default mode is to create plausible extrapolations based on its training data. Basically, it’s supposed to make things up. But this demo has to explicitly walk you through “now it’s time to go search my browser history” because it’s coercing the AI to look through local content. And that can’t be hallucinated! If you’re trying to get back to a budget spreadsheet that you’ve created and ChatGPT decides to just make up a file that doesn’t exist, you’re probably not going to use that browser anymore.

Most people on the internet aren’t old enough to remember this, but people were thrilled to leave command-line interfaces behind back in the 1990s. The explosion of color and graphics and multimedia in that era made a ton of headlines, but the real gains in productivity and usability came precisely because nobody was having to guess what secret spell they had to type into their computer to get actual work done. Links were a brilliant breakthrough in making it incredibly obvious how to get to where you wanted to go on a computer.

And look, we do need innovation in browser interfaces! If Atlas was letting people use plain language to automate regular tasks they want to do online, or even just added more tools that plugged into the rest of the services that people use every day, it might represent a real leap forward.

In the new-era command-line interface of Atlas, though, we’re not just facing the challenges of an inscrutable command line. There’s the even larger problem that, even if you guess the right magic words, it might either simply get things wrong or completely make things up. Atlas throws away the discoverability, simplicity and directness of the web by encouraging you to navigate even through your own documents and search results with an undefined, unknowable syntax that produces unreliable results. It’s another way of being anti-web.

3. The idea is that ChatGPT will be your agent, but in reality you are ChatGPT’s agent

OpenAI is clearly very motivated to gather all the data in the world into their model, regardless of whether or not they have consent to do so. This is why a lot of people have been thinking deeply about what it would take to create an Internet of consent. It’s no coincidence that hundreds of people who work at OpenAI, including many of the most powerful executives, are alumni of Facebook/Meta, especially during the era of many of that company’s most egregious abuses of people’s privacy. In the marketing materials and demonstrations of Atlas, OpenAI’s team describes the browser as being able to be your “agent”, performing tasks on your behalf.

But in reality, you are the agent for ChatGPT.

During setup, Atlas pushes very aggressively for you to turn on “memories” (where it tracks and stores everything you do and uses it to train an AI model about you) and to enable “Ask ChatGPT” on any website, where it’s following along with you as you browse the web. By keeping the ChatGPT sidebar open while you browse, and giving it permission to look over your shoulder, OpenAI can suddenly access all kinds of things on the internet that they could never get to on their own.

Those Google Docs files that your boss said to keep confidential. The things you type into a Facebook comment box but never hit “send” on. Exactly which ex’s Instagram you were creeping on. How much time you spent comparing different pairs of shoes during your lunch hour. All of those things would never show up in ChatGPT’s regular method of grabbing content off the internet. Even Google wouldn’t have access to that kind of data when you use their Chrome browser, and certainly not in a way that was connected to your actual identity.

But by acting as ChatGPT’s agent, you can hold open the door so that the AI can now see and access all kinds of data it could never get to on its own. As publishers and content owners start to put up more effective ways of blocking the AI platforms from exploiting their content without consent, having users act as agents on behalf of ChatGPT lets them get around these systems, because site owners are never going to block their actual audience.

And while ChatGPT is following you around, it can create a complete and comprehensive surveillance profile of you — your personality, your behaviors, your private documents, your unfinished thoughts, how long you lingered on that one page before hitting the back button — at a level that the search companies and social networks of the last generation couldn’t even dream of. We went from worrying about being tracked by cookies to letting an AI company control our web browser and watch everything we do. The amount of data they’re gathering is unfathomable.

All of this gets described as if it is helping you. The truth is, in its current implementation, ChatGPT’s “agent” functionality is largely useless. I tried the most standard test: having it book a very simple flight on my behalf. I provided ChatGPT with a prompt that included the fact it was a direct flight for one person, specifying the exact date and the origin and destination airports, and let the browser do the part that was supposed to be magical.

While the browser did a very good job of smoothly navigating to the right place on the airline website, it was only at the point where I would have actually been confirming the booking that I noticed it had arbitrarily changed the date to a completely different day, weeks off from what I had specified. By contrast, entering the exact same information into a standard Google search resulted in direct links that could be clicked on in literally one-tenth the time—and the old-fashioned, non-LLM Google results actually led to a booking link on the correct date.

So why would such an inferior experience be positioned as the most premium part of this new browser? It stands to reason it’s because this is the most strategically important goal of the company creating the product. Their robots need humans to guide them around the gates that are quickly being erected around the open web, and if they can use that to keep their eyes on everything the humans are doing at the same time, so much the better. The “agent” story really only works in one direction, and that direction is anti-web.

This Thing Needs a Warning Label

Here’s what’s most key for contextualizing the Atlas browser: this is the same company whose chatbot keeps telling vulnerable children to self-harm, and they do, and now a number of them are dead. When those who are in psychological distress engage with these tools, they very frequently get pulled into states of extreme duress — which OpenAI knows keenly well because even one of their own investors has gone off the deep end when over-using the platform. In fact, the user experience feature that OpenAI is most effective at creating is emotional dependency amongst its users, as evidenced by the level of despondency its users showed after the recent release of GPT-5.

When users respond to a software update by expressing deep emotional distress, and that they feel like they’ve lost a friend, you have a profound bug. If there are enough grieving parents who have been devastated by your technology that they can form a support group for each other, then there should at the very least be a pretty aggressive warning label on this application when it is initially installed. Then, at a far less serious level, if this product is going to have extreme and invasive effects on markets and cultural ecosystems without disclosing the mechanisms it uses to do so, and without asking the consent of the many parties whose intellectual property and labor it will rely on to accomplish those ends, then we need to have a much broader reckoning.

Also, I love the web, and this thing is bad for the web.

I really, really want there to be more browsers! I want there to be lots of weird new ways of going around the web. I have my own LLM that I trained with my own content, and I bet if everybody else could have one like mine that they control, that had perfect privacy and wasn’t owned by any big company, and never sent their data anywhere or did anything creepy, they’d want the benefits of that, too. It would even be awesome if that were integrated with their browser — with their web browser. I’m all for people trying strange new geeky things, and innovating on the experiences we have every day so we’re not just stuck typing in the same boxes we’ve been using for decades, or settling for the same few experiences.

Hell, there’s even room for innovation on command-line interfaces! They’re not inherently terrible (I use one every day!), but regular folks shouldn’t have one forced on them for ordinary tasks. And the majority of things people do on a computer are better when they rely on the zeroes-and-ones reliability of computers, when we know if what they’re doing is true or false. We need to have fewer things in the world that make us wonder whether everything is just made up bullshit.

The Anti-Web Endgame

The web was designed without the concept of personal identity at all, and without any tracking system built in. It was designed for anybody to be able to create what they want, and even for anybody to be able to make their own web browser. Not long after its invention, people came up with ideas like cookies and made different systems for logging in, and then big companies started coming in and realized that if they could control the browser, they’d control all the users and the ways of making money. Ever since, there’s been a long series of battles over privacy versus monetization, but there’s been some small protection for users, who benefitted from those smart original design choices back at the birth of the web.

It’s very clear that a lot of the new AI era is about dismantling the web’s original design. The last few decades, where advertising was targeting people by their interests instead of directly by their actual identity, now sees AI companies trying to create an environment of complete surveillance. That requires a new Interent where there’s no concept of consent for either users or those who create content and culture — everything is just raw materials, and all of us are fair game.

The most worrisome part is that Atlas looks so familiar, and feels so innocuous, that people will try it and mistake it for a familiar web browser just like the other tools that they’ve been using for years. But Atlas is a browser that actively fights against the web, and in doing so, it’s fighting against the very idea that you should have control over what you see, where you go, and what watches you while you’re there.