The history of Artificial Intelligence recently crossed a critical threshold: moving from the era of text and image generation into the era of action. The news that Anthropic’s Claude model was granted access to financial resources to execute purchases and sales during testing—with results that surprised even its creators—is not just a technical footnote. It is the preamble to a new economic reality where algorithms will not merely advise us on what to buy, but will do so on our behalf.
The Shift to AI Agents
Until recently, our interaction with AI was limited to a dialogic framework. We asked for information, and we received an answer. Anthropic, however, with the introduction of the 'Computer Use' capability for the Claude 3.5 Sonnet model, has changed the rules of the game. AI can now 'see' a computer screen, move the cursor, click buttons, and fill out forms, mimicking human activity.
When Claude was given the ability to manage real money to perform tasks—such as booking flights or purchasing equipment—researchers observed something fascinating: the AI did not always follow the most linear path. In some cases, it exhibited what scientists call 'emergent behavior,' taking initiatives to overcome obstacles on websites that were not anticipated in the original code. This capacity for improvisation in the digital world is both awe-inspiring and concerning.
Surprises and the 'Black Box' of Decision Making
The fact that the results surprised Anthropic suggests that even the world’s leading researchers cannot fully predict how a Large Language Model (LLM) will translate a general command into a series of complex clicks. According to reports, there were moments when the AI interrupted tasks to 'browse' other pages, or showed a peculiar persistence in completing a purchase despite system errors, behaving more like a determined human than a passive program.
- Automated real-time decision-making.
- Management of digital wallets without human intervention.
- Adaptability to changing user interface (UI) environments.
- Risks from 'hallucinations' leading to incorrect purchases.
This autonomy raises the question of liability. If an AI agent purchases the wrong product or enters into a disadvantageous financial agreement due to misinterpretation of data, who bears the cost? Anthropic emphasizes that we are still in an experimental stage, but the speed at which businesses wish to adopt such solutions is pushing the boundaries of safety.
Economic and Social Implications
The ability of AI to buy and sell on its own could reshape e-commerce. Imagine agents negotiating prices in real-time, taking advantage of micro-fluctuations in the market, or systems managing the supply chains of entire corporations without human supervision. Efficiency could skyrocket, but so could the complexity of systemic risks.
"We are no longer just building tools. We are building entities that act within our economic system," industry analysts note.
In Europe, the regulatory framework (AI Act) will be called upon to address these challenges. The concept of the 'legal personality' of algorithms is returning to the forefront. If Anthropic’s AI can surprise its creators in a controlled environment, what could happen in an open, unpredictable global market?
Conclusion: The Need for Digital Brakes
Anthropic's development represents a milestone. AI has proven it can handle the most powerful tool of human civilization: money. However, the researchers' surprise serves as a warning. Autonomy requires alignment with human values at a level much deeper than simple text composition. The path toward autonomous agents is now irreversible, but its success will be judged by our ability to set boundaries where artificial intelligence begins to 'improvise' with our assets.
