OpenAI has, so far, stuck to making various LLMs and AI models and one consumer product: ChatGPT. It now wants to try browsers, though I’m not sure if you should try this one out.

OpenAI announced ChatGPT Atlas, a new web browser for macOS (it will eventually make landfall on Windows and mobile phones too) that, as you might expect, integrates ChatGPT right into its core. More AI. Yay. OpenAI promises that its AI can interact with web content in real-time. Unlike the existing method of copying and pasting information into a separate ChatGPT window, Atlas enables the AI to “see” the content on a user’s current page. This allows for in-context queries, summaries, and analysis without leaving your active tab. The browser also incorporates the user’s existing ChatGPT chat history, allowing the AI to draw upon previous conversations to inform its responses.

Credit: OpenAI

One feature that Atlas introduced is “browser memories.” When enabled, this function allows ChatGPT to remember information and context from websites a user visits. This retained context can then be used to answer complex, cross-sessional questions. For example, you could ask the browser to “summarize industry trends based on the job postings I viewed last week.” According to OpenAI, this feature is optional and user-controlled. You can view, archive, or delete all browser memories, and clearing the browser history will also erase any associated memories. Furthermore, users can prevent ChatGPT from seeing specific sites using a toggle in the address bar.

Of course, it couldn’t be an AI browser without an “agent mode.” Available in preview for Plus, Pro, and Business subscribers, this feature allows ChatGPT to take direct action on a user’s behalf. It can autonomously navigate websites, click buttons, and fill out forms to complete multi-step tasks. The company provided examples such as providing the agent with a recipe and instructing it to order all the necessary ingredients from an online grocery store, or tasking it with compiling a competitive research brief from various team documents and public websites. This is something that’s coming to some degree to browsers like Chrome and is already live in other browsers such as Opera Neon.

This is a browser built with AI in mind. If you’ve read some of my articles before, you’d know I’m not actually against the idea of AI in browsers—if done right, it could be genuinely helpful. Sadly, though, despite multiple attempts, agentic AI is not great and is prone to issues. Back when ChatGPT first showed its agentic AI capabilities, the first demos and user tests hilariously fell apart whenever a reviewer would ask it to do something outside of the very few test scenarios OpenAI envisioned internally, struggling to do simple things such as buying a white T-shirt for the user. I’m sure it’s gotten better, but maybe not to the point where you want your whole browser to depend on those capabilities. Just saying.

According to OpenAI, it doesn’t train its models on your browsing data, but you can opt into this. You better keep that toggle disabled, though, because that could be disastrous. And to add insult to injury, some malicious websites have caught onto AI agents and will try to manipulate them with hidden instructions. OpenAI has apparently not worked around this at all, because it’s warning users about this and advises them to remain vigilant, monitor the agent’s activities, and consider using it in a logged-out mode to limit its access to sensitive accounts. So, you know. That’s fun.

Source: OpenAI