The transition from artificial intelligence that 'speaks' to artificial intelligence that 'acts' is the holy grail of modern technology. Anthropic, the company founded by former OpenAI executives with a mission of safety, recently took its most daring step: the release of a capability called 'computer use.' This tool allows the Claude 3.5 Sonnet model to interact with computer screens, move the cursor, click buttons, and type text, just like a human user.
The Technological Leap: Controlling the Screen
Until today, large language models (LLMs) were confined to a digital 'box.' They could analyze data, write code, or compose text, but implementing those outputs always required human intervention. Anthropic's new feature changes the paradigm. Claude doesn't just receive data via an API; it 'sees' the screen through continuous screenshots, interprets the position of icons and text fields, and decides which move to make next to complete a complex task.
For example, a user could ask the AI to take data from an Excel file, open a browser, log into a CRM platform, and fill out a form. Claude performs these steps autonomously, navigating between different windows and applications. This capability transforms the AI from a passive assistant into an active agent, capable of managing workflows that previously required hours of manual labor.
Why It Was Deemed 'Dangerous'
The delay in releasing this feature was not accidental. Anthropic itself had expressed reservations about the risks inherent in AI software gaining access to a computer's operating system. These risks fall into three main categories:
- Cybersecurity: Such a tool could be used to automate cyberattacks, discover vulnerabilities in real-time, or conduct sophisticated phishing campaigns.
- Prompt Injection: There is a risk that the AI could be guided by malicious instructions found on a website it visits, leading to unintended actions, such as deleting files or sending sensitive data.
- Unsupervised Autonomy: The possibility of the model misinterpreting a command and causing financial or operational damage to a business without being immediately noticed.
Anthropic claims to have implemented strict safety protocols, limiting the model's internet access during testing and developing systems that detect when the AI attempts to perform 'dangerous' actions, such as interacting with social media sites or registering for government services.
The Strategic Battle of the Giants
Anthropic's move is not just a technological achievement but a strategic response to competition from OpenAI and Google. While OpenAI focuses on developing the 'o1' model with improved reasoning capabilities, Anthropic chooses to emphasize practical utility within the corporate environment. Backed by billions from Amazon and Google, the company aims to become the first choice for businesses looking to automate their internal processes.
"We are not just building a chatbot; we are building a collaborator that can use the same tools as you," said a company executive during the presentation.
The challenge for Anthropic remains the balance between 'Constitutional AI'—a training method based on ethical principles—and the need for rapid innovation in a market that doesn't wait. The release of 'computer use' in beta suggests that the company now feels confident enough in its safeguards, or perhaps, that competitive pressure is forcing it to take calculated risks.
The Future of Work and the Ethics of Autonomy
As these tools mature, the debate over job displacement takes on a new dimension. If an AI can handle the software of an accountant, a programmer, or a customer support representative, what will be the human's role? The optimistic side sees the liberation of workers from repetitive tasks, allowing them to focus on strategy and creativity. However, the transition will be painful without adequate preparation from the educational system and governments.
At the European level, the implementation of the AI Act will play a decisive role in how such autonomous tools enter the market. Anthropic seems to be paving the way by offering transparency in its testing, but the real test will come when Claude begins operating on thousands of corporate computers worldwide, facing the unpredictable chaos of the real world.
