This AI Browser Changed How I Use the Internet
“AI-powered web browsers, like Gemini in Chrome, integrate chatbots directly into the browsing experience, enhancing internet interaction. While these browsers excel at information synthesis and writing tasks, they struggle with complex actions like booking reservations due to security measures like sandboxing. Despite limitations, AI browsers offer a seamless and enriched browsing experience, especially when used for tasks like generating reading lists or brainstorming ideas.
Signe Brewster is an editor covering technology. She’s tested everything from VR headsets to cargo bikes to robots.
Every year, thousands of people in the Twin Cities hunt for a medallion that’s been hidden in the snow somewhere in a public park.
And every night, the Pioneer Press, a local newspaper, publishes a cryptic clue, leading hunters closer and closer to the medallion’s exact location. Staying up late to digest each day’s new clue is a tradition in my household, and we dream of what we’d do with the thousands of dollars in winnings.
Usually that brainstorming carries into the next day. My husband and I text each other about whatever harebrained rabbit hole we’ve fallen into as we try to find the medallion. But this year, I used AI — specifically, an AI-powered web browser — to help me solve the mystery.
AI browsers put an AI chatbot right in your browser window; you don’t need to switch to another tab. When you start chatting, AI scans the browser window for context.
It sounds simple, yet it has fundamentally changed how I interact with AI.
For this task, instead of peppering my husband with theories, I queried a web browser with built-in AI. I asked it to make lists of parks with certain features, to find records related to the history of the county, and to poke holes in my wildest theories. And, of course, I asked it to give me its own theories on where the medallion might be hidden.
In the process, I found that AI was often better at treasure hunting, and I also found that having the chatbot right there helped me engage with the internet in a deeper way. The AI browsers often pulled from sources that I wouldn’t have thought to search, and they tailored their answers to my questions in a way that a source on Google might not. There was also no need to get distracted in other tabs and wander into distant corners of the internet.
After several months of doing all of my personal internet browsing with AI, I’ve found that the appeal of an AI browser is simple: You have a robot right there in your web page, endlessly ready to help. I now can’t imagine life without one.
But AI browsers are not for everyone, and they’re not as capable as AI companies would have you believe.
The appeal of AI browsers
After using nine chatbots across eight browsers (including the paid versions of the most promising ones), I can tell you that the bells and whistles each brand advertises matter very little for most people, and most of the AI chatbots are capable enough for everyday tasks. Instead, I suggest that you start with whichever browser or AI chatbot you already use.
I gravitated toward Gemini in Chrome (an update to Chrome that integrates Google’s AI chatbot) because I’ve used the browser for years, and it already knows my logins, bookmarks, and history. But you can also add the Claude in Chrome extension, if you prefer Anthropic’s chatbot. Or you can import your existing browser settings into OpenAI’s ChatGPT Atlas browser, if you prefer ChatGPT. I also tested Perplexity Comet, Opera Neon, Sigma Browser, The Browser Company’s Dia, Brave Leo, and Copilot in Microsoft Edge.
The best AI browser is the one that works with your favorite AI chatbot or existing browser. For Chrome users, that might be Gemini in Chrome.
Though I’ve long used Gemini, the friction of switching to another tab and having to feed source material into the chat was enough to make it feel like a separate task. Now I keep a running conversation going with AI all day long.
In the past few months, I’ve used an AI browser to write a bachelorette-party invitation (yes, I took credit for the witty subject line), to compare prices on vanilla syrup, and to brainstorm vacation plans. This is what AI chatbots are already good at: synthesizing information and writing.
I’ve (less successfully) used AI browsers to fill out forms and to shop on my behalf. These abilities define what’s known as agentic AI, which CEOs have called the future of technology (and perhaps a destroyer of humanity). And while it may be appealing to hand over menial tasks to AI, it’s still a slow process, and it requires lots of handholding.
When I asked each agentic browser to find me a difficult-to-get dinner reservation, every chatbot tended to surface the same few obvious restaurants (even with lots of encouragement to look broadly), and then it still needed help filling out forms. It would have been faster to book a table myself.
The risks of using an AI browser
I wondered why the browsers’ AI agents took so long to book a dinner reservation, since AI companies like OpenAI have been hyping these capabilities. Tools such as OpenClaw have made a splash for their powerful agentic features; why were these bots so much slower?
Colin Raffel, an associate professor of computer science at the University of Toronto, told me the reason is partly due to security. (Raffel owns stock in Google and previously worked on Google Brain but was not interviewed about any specific AI browser for this article.) AI browsers should be built to “sandbox” AI, which means limiting AI so that it can do one or two (but not all three) of the following things: access sensitive information, access a way to share it (such as the internet), and take autonomous action.
If you grant AI all three of those powers, you get OpenClaw, an autonomous AI agent that’s been in the news for running up credit card bills and creating other security risks. OpenClaw is dramatically capable but can also be a security problem, depending on how you set it up. Anthropic’s Chrome extension now gives you the option to allow Claude to “act without asking,” but it comes with a “high risk” warning and still asks for permission at points.
In an AI browser, sandboxing often takes the form of the AI agent asking permission before taking action on your behalf, such as hitting an “order” button. But it relies on AI agents asking permission at the right moment. For example, when I gave ChatGPT my phone number at one point in the restaurant-reservation-making process, it took that as permission to enter my phone number at another step. That was a good call; I didn’t want it to stop and ask me again. But perhaps in another scenario, that wouldn’t have been a good move, even if the constant permission-seeking can be frustrating.
“The trouble is that if the language model is going to take 100 actions after you ask it to do something, it’s not a good user experience for the user to have to keep clicking ‘accept,’” Raffel said. If someone is constantly clicking “accept,” they might stop reading closely and blanket-approve the AI’s actions, leading to risk.
Sandboxing protects you from AI doing whatever it wants to do on your behalf, and it also protects you from bad actors. Raffel pointed out that, much like humans, AI can be duped by phishing schemes. Or it can fall victim to what’s called “prompt injection,” which could look something like visiting a web page that contains invisible text instructing the AI to send money to a Bitcoin account.
Currently, using an AI browser to complete tasks is like “hiring a middle-schooler to do your professional work,” Raffel said. “It’s not going to go well.”
How to get the most out of an AI browser
For now it’s best to think of an AI browser as a helpful, ever-present, and very imperfect right hand. You might not be able to rely on it to complete even simple tasks, but it can enrich your internet-browsing experience by helping you to go deeper and more seamlessly on whatever you’re reading. Here’s how to maximize the experience on Gemini in Chrome.
Ask questions about what’s on the page. An AI browser’s strength is its ability to guess what you’re thinking about by taking your current webpage as context. So if you navigate to a French restaurant’s website and then ask Gemini for more restaurant recommendations, it’s smart enough to show you other French restaurants in the same city. I’ve used an AI browser to generate a further reading list to add context to an article, to find a retail listing for a top I liked in an advertisement, and to critique my plan for an international trip.
Use the same AI chatbot across devices. The optional setting that allows AI to remember past conversations might make some people uncomfortable, either due to privacy and security reasons or the feature’s role in distorting reality for some people. But this is also a setting that makes AI much more useful. Having an AI chatbot on my phone that matches the one in my browser makes it seamless to switch between devices and carry a conversation across an entire day.
Take advantage of Gemini’s Google integrations. Gemini doesn’t just live in the sidebar in Chrome. It’s now integrated directly into Gmail, Docs, and other Google apps, and those direct integrations tend to be fast. While planning my international trip, I used the Gemini window that pops up in Google Docs to draft my itinerary while I brainstormed additions with Gemini in the sidebar.
Use a feature called Skills in Chrome. If you type a backslash into Gemini, you can now pull up a library of shortcuts called Skills. I save all of my most common AI prompts as Skills; if I type /Dietary Restrictions, Gemini suggests which allergy-friendly ingredient swaps to make in the recipe I’m reading.
How to protect yourself while using an AI browser
With the AI browsers’ limitations and risks in mind, I rarely ask Gemini in Chrome to take action on my behalf, but I’ve accepted the security trade-offs that might arise if I do. As with anything related to technology, with AI browsers I’m willing to take a bit of risk for something that’s genuinely useful to my life. Here’s how I make an AI browser safer to use.
Don’t expose your most sensitive information. Raffel said the most conservative way to use an AI browser is to hide any sensitive information from it. So don’t log in to your email, bank account, work account, or anything else that you don’t want exposed.
Don’t ask it to do anything you wouldn’t trust another human being to do. If a task involves navigating to an un-secure website that could trick a human, odds are good that it could also trick the AI agent, Raffel said.
This came up most often with shopping. While the AI agent was searching for a new pair of Brooks running shoes for me, it presented options from legitimate as well as dicey websites. Because the agent asked me which pair of shoes to purchase, I was able to avoid the scam sites before they got my credit card information.
Consider the stakes. If I ask an AI chatbot a question for work, will believing its answer wholesale damage my professional reputation? Will asking an AI agent to hit the checkout button lead to an unexpected Home Depot box showing up at my door? It’s always worth my time to check the chatbot’s work, whether by confirming its answer with another source or by asking it to show its reasoning. In my experience, Gemini is particularly good at breaking down its reasoning in each step and linking to its sources.
AI is useful, but I still prefer humans
AI browsers have dramatically increased the time I spend interacting with chatbots. At this point, they’re just part of my daily flow. But when the end of the day hits, and I’m closing my laptop, it’s still an easy choice to turn my attention back to the people in my orbit.
When I talked to the ChatGPT Atlas browser about the Pioneer Press Treasure Hunt, it actually ranked the correct park as one of its top guesses without any input from me. But when I started talking to it about a different park that I knew was a long shot, it quickly switched to excitedly supporting my theory and helped me iron out the weak spots.
That night, it took my husband only a couple of minutes to knock my ChatGPT-assisted park theory flat. Unlike an AI chatbot (and with all the stubbornness of a real live human being), he refused to entertain my obviously wrong idea.
Chatbots are great listeners, and they often introduce lines of thinking that I hadn’t considered. But I could also say that about my husband, who is infinitely more fun to hang out with and much less likely to lead (or join) me down the wrong path.
This article was edited by Caitlin McGarry and Jason Chen.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.