AI-Powered Browsers Are Falling Short
But here’s the core issue you might not want to overlook: AI-enabled browsers promised to redefine how we explore the web, yet they’re failing to live up to that hype. The AI industry has sold a future where intelligent assistants automate tasks across the browser, much like they’ve touted autonomous AI agents for computer work. And while breakthroughs keep being announced, the day-to-day reality remains underwhelming for most users.
A recent round of tests from The Verge put several chatbot-integrated browsers under the microscope, and the verdict was clear: they’re often clunky, surprisingly slow, and require more tinkering than they’re worth. The overarching pattern was frustration and wasted effort, even though automation is the selling point.
As The Verge’s Victoria Song summarized, the problem isn’t just the tool’s intelligence—it’s how users must structure prompts to get results. “No matter the browser, I kept encountering the same fundamental hurdle: you have to think extra hard about how to craft the right prompt,” she wrote. “Linking an AI assistant to a browser doesn’t automatically transform how you interact with a chatbot.”
OpenAI unveiled Atlas, a new AI browser centered on ChatGPT, in October. This move thrust AI browsers into the mainstream spotlight and suggested they’ll become a major battleground in tech. Competitors include Perplexity’s Comet, which debuted earlier, and The Browser Company’s Dia, which arrived mid-year.
Traditional browsers like Google Chrome and Microsoft Edge now offer chatbot features as add-ons—Gemini and Copilot, respectively. But fully fledged AI browsers place agentic AI at the forefront, inviting users to type a prompt before entering a URL.
The Verge evaluated all of these options. One common test was using AI browsers to organize and summarize emails—a frequently touted benefit. The results were discouraging: despite numerous prompt refinements, the AI systems often filtered out important messages or produced unhelpful summaries. It took a very specific prompt to achieve any meaningful result:
“Find unanswered emails in which I previously expressed interest or provided personalized requests/feedback,” the prompt began. “Then determine which ones I should respond to based on timeliness and keywords like ’embargo’ with dates in the next two weeks. Ignore emails with multiple follow-ups I haven’t addressed.”
Does that sound practical? Comet and Dia did flag some relevant messages, but several results wandered into spam territory. Atlas offered only a technical explanation for why the request couldn’t be completed and suggested further refinement, which led Song to declare the effort a lost cause.
Another arena where these tools struggled was shopping—an area many expect AI browsers to excel. They could conduct rapid research for a stylish running shoe, yet still goof on basics like color accuracy. The purchasing process proved equally error-prone: Atlas repeatedly pressed the user to confirm items in the cart and even spent an entire minute trying to close a window to resume shopping.
In short, AI browsers face the same hurdles as AI agents: they’re slow, demand constant supervision, and push critical decisions back onto the user, undermining the very idea of autonomous assistance.
Security is another major concern. Numerous studies show that AI browsers are highly vulnerable to prompt injection attacks, where malicious content embedded in web pages can steer the AI to harmful actions. In one notable case, Perplexity’s Comet could be manipulated to expose bank account access via a Reddit post. Other research demonstrated that feeding Atlas fake URLs could trick it into accessing and deleting files in a Google Drive account.
Safety must come first. Yet the bigger question remains: can these tools ever reach a level of seamless ease that persuades widespread adoption? For now, AI browsers feel more like expensive copilots that require constant babysitting rather than true personal assistants.
As Song put it, the experience often means doing tasks for AI so it can occasionally do something for the user. The focus shifts from how AI fits into life to how everyday routines must adapt to accommodate AI’s growing presence.
If these topics spark debate, share your thoughts in the comments: Do AI browsers represent a real leap forward, or are they a temporary detour on the road to truly autonomous browsing? Are there use cases where they genuinely outperform traditional workflows, or do the current flaws outweigh the benefits in most situations? Consider what improvements would need to happen for you to trust and rely on AI-powered browsing in daily tasks.
