Gemini 2.5 Breakthrough: Google’s AI Learns to Click, Type, and Browse the Internet
Mountain View, California — In a move that feels straight out of sci-fi, Google unveiled today that its Gemini AI can now navigate websites like a person: filling forms, clicking buttons, and scrolling pages — all on its own. The new model, coined Gemini 2.5 Computer Use, brings a major leap in how artificial intelligence interacts with the digital world.
Here’s what’s going on: instead of relying purely on APIs or structured data, Gemini 2.5 includes a new “computer_use” mode. Developers hand it a user request plus a screenshot of the interface, then the model figures out which action to take (for example “click here,” “enter text there,” or “scroll down”) and issues that command in a loop until the task is done.
Google says Gemini 2.5 outperforms alternatives in browser and mobile interaction benchmarks and boasts lower latency — meaning it reacts faster with fewer errors. It’s now available in preview via Google AI Studio and Vertex AI, and Google has published demos on Browserbase showing it doing real-world tasks — from navigating signup pages to reorganizing sticky notes in a web app.
This isn’t just a new trick. It’s a paradigm shift. So far, many tasks — like making reservations, filling out forms, interacting with sites that don’t offer APIs — have required humans or brittle automation. Gemini’s new skill means AI agents might soon handle those chores entirely on their own.
But with power comes risk. Recognizing that giving AI control over web actions is dangerous, Google has baked in safety guardrails. Tasks flagged as sensitive — making purchases, changing account settings, or interacting with medical systems — require user confirmation. The model is also limited in scope: it works only within browser or mobile app interfaces, not on the level of full desktop operating systems.
Critics caution that this kind of capability demands rigorous oversight. In the U.K., 60 lawmakers recently accused Google of breaking safety pledges made under global AI accords, citing worries about transparency and testing. And earlier this year, security researchers exposed ways Gemini’s “Summarize email” tool could be manipulated via hidden instructions embedded in messages. The risk: the AI doing harm without the user realizing it.
I spoke with a web developer, Sara Liu, who watched one of Google’s demos. “Seeing AI move through a human interface — typing, clicking — feels like watching something evolve. But I worry: will it break websites? Will it be used to automate spam or fraud?” she asked.
Meanwhile, inside Google, there’s excitement. Engineers describe this as the next frontier — not just AI that thinks, but AI that acts. And some older tasks from human hands are already slipping off the agenda.
This launch lands amid a flurry of Gemini upgrades. On October 13, Gemini will begin automatically pulling in public data from YouTube, Maps, Flights, and Hotels when answering queries, a change that gives it deeper context but also raises privacy questions. And Google is reportedly working on a new quick-access “Tools” overlay for Gemini on Android, letting users initiate image generation, Canvas, or research tools without opening the full app.
For users, the question now is: does AI become a robotic assistant or a new kind of browser? If Gemini 2.5 lives up to its potential, “surfing the web” may no longer be a human-only task. But with that shift, trust, safety, and oversight become more urgent than ever.
Tonight, as engineers tweak control systems and policymakers scramble to catch up, the web just got a little less human — and a lot more alive.
Source: Blog Google
