Anthropic is working on developing an open-source technical system that enables verification of AI model identities and ensures that the answers issued were indeed produced by the declared model, relying on cryptographic mechanisms that provide mathematical evidence of the output source, thereby enhancing transparency and trust in these technologies.
Verifiable Inference System
The tech website TechCrunch reported on Tuesday that the new system is named "FlashProve" and aims to enable what is known as "verifiable inference," which allows users to confirm that the answer they received came from the specified model, not from another model used as a substitute without their knowledge.
Cryptographic Mechanisms and Transparency
The system relies on publishing a unique digital fingerprint for each model, allowing it to be cryptographically matched with outputs to verify their source, without the need to disclose the model's internal weights, which are the data that determine its operation mechanism and behavior and are usually kept secret by developing companies. This technology can also help prevent AI companies from replacing their models with cheaper ones without notifying users.
Development and Research
The system was designed to be technically lightweight, allowing verification processes to be conducted on local devices without relying entirely on external servers. Anthropic stated that the project is still in its early research stages, and there are currently no plans to implement it within its labs or at other AI labs.
Source: Syrian Arab News Agency (SANA)
Comments (0)