Anthropic's recent study on the inner workings of Claude, an artificial intelligence model, has reignited discussions among experts about the use of anthropomorphic terms to explain AI systems. The research suggests that Claude can perform intermediate steps of mathematical logic and detect errors in codes without making this reasoning explicit to the user.
This raises questions about the extent to which the use of terms like 'think', 'anxiety', and 'consciousness' helps explain complex technology and when it starts to attribute human characteristics to software. Experts like Gustavo Torrente, professor of new technologies and AI at FIAP, warn that using these terms may lead the public to conclude something the article itself does not state.
1. What is 'J-Space' and how does it work?
The interpretability team at Anthropic identified an internal space shared by different components of the model, dubbed 'J-Space'. This space appears to play a role similar to the neuroscientific theory of 'Global Workspace', functioning as a shared environment where different parts of the model exchange information before the final response is generated.
Imagine a large control center of a plant, where different sectors communicate to ensure efficient operation. 'J-Space' works similarly, allowing different parts of the AI model to communicate and generate responses more effectively.
2. What is the impact of using anthropomorphic terms?
The choice of biological and psychological terms by major AI labs divides researchers on the real impact of these expressions on the public. Some argue that using terms like 'anxiety' and 'thoughts' may attribute human characteristics to systems that remain statistical models.
For example, if a restaurant adopted an AI system to manage orders and deliveries, using terms like 'the system is eager to serve' may create unrealistic expectations about the system's ability to really 'feel' or 'think' like a human.
3. What are the implications for digital security?
The study also represents an important advance in model interpretability. Reverse engineering J-Space allows monitoring internal activations that do not appear in the text generated by Claude. This makes the artificial intelligence less like a black box and turns it into a glass box.
Imagine an insurance company using an AI model to assess risks and issue policies. With the ability to monitor internal activations, the company can identify potential biases or problems in the evaluation process, ensuring a fairer and more secure decision-making process.
4. How does this affect market perception?
The disclosure of the research occurs at an important time for Anthropic, which is experiencing a period of strong valuation and frequently appears in international press reports as one of the main candidates for a future initial public offering (IPO).
The narrative about AI advances may strengthen investors' perception of the company, but it also increases expectations and pressure for transparency, independent auditing, and legal responsibility.
Frequently Asked Questions
What is Claude?
Claude is an artificial intelligence model developed by Anthropic, designed to perform complex tasks such as mathematical logic and error detection in codes.
What are the risks of using anthropomorphic terms?
The risks include attributing human characteristics to systems that do not possess them, creating unrealistic expectations about the system's ability to 'feel' or 'think' like a human.
How does this impact digital security?
The study represents an important advance in model interpretability, allowing monitoring of internal activations and identification of potential biases or problems in the evaluation process, ensuring a fairer and more secure decision-making process.
