Google has introduced a preview version of its latest artificial intelligence model, Gemini 2.5 Computer Use, designed to perform tasks on web browsers much like a human user. The system can move the mouse cursor, click buttons, fill out forms, scroll through pages, and type text marking a major step toward giving AI direct control over computers.
According to Google, developers can now use the new Gemini 2.5 Computer Use API to create agents capable of interpreting user interfaces through screenshots and translating commands into visual actions such as clicking, dragging, and typing. At this stage, the AI is primarily optimized for web browsers and cannot yet operate independently across full desktop environments.
Google emphasized that the model includes strong safety protocols. When faced with sensitive actions, such as sending an email or making a financial transaction, the AI will first request explicit confirmation from the user before proceeding.
The development signals a shift in artificial intelligence from conversational chatbots to more advanced “AI agents” capable of performing practical digital tasks. These agents can act beyond the chat window to execute real-world operations in software systems, similar to how a human assistant would navigate a computer.
Google claims that Gemini 2.5 Computer Use performs better than competing systems in terms of precision and speed, especially when interacting with web and mobile applications. However, the company admits the model is not yet advanced enough to control an entire operating system or handle file management and system configurations.
Despite its promise, the new model raises significant concerns about potential misuse and security vulnerabilities.
Google has acknowledged risks such as malicious use, unexpected behaviors, data fraud, and manipulation within online environments. To minimize these threats, developers can limit the AI’s permissions and must require user authorization for sensitive commands.
The Computer Use model remains in its experimental phase and may still produce errors. Google has advised developers and users to avoid applying it to high-risk or mission-critical operations without supervision.
Currently, the model is available to developers through Google AI Studio and Vertex AI. This rollout reflects Google’s broader strategy to advance autonomous AI agents capable of performing complex workflows, a move that could redefine how people interact with technology in the near future.
