Anthropic develops FlashProve to verify AI responses — TechCrunch - UA.NEWS
By Google News AI ·
Anthropic is developing an open system called FlashProve for cryptographically verifying which specific artificial intelligence model produced a particular response. As TechCrunch reports, the development is intended to provide users with mathematical confirmation that a result originated from a specific model.
Anthropic calls this approach “proven inference.” FlashProve is intended to work similarly to a cryptographic signature on an email: the system would make it possible to verify that a response truly came from the claimed model rather than from another model impersonating it.
According to the researchers’ concept, the system could help prevent the developer laboratory itself from secretly substituting a model. In particular, a company could covertly use a cheaper model to reduce costs or a more expensive one to inflate its performance metrics.
More current news is available on the UA.News Telegram channel Telegram.
For FlashProve to work, laboratories would need to publish a digital fingerprint of every released model. It could be matched against specific responses and verified cryptographically. At the same time, such fingerprints would not disclose a model’s weights — the internal numerical parameters that determine its behavior and usually remain closed in commercial developments.
The system is also proposed for verifying claims by third parties. For example, a developer could confirm that its software library was checked for vulnerabilities by an advanced cybersecurity model, while a company could confirm that its business plan was approved by the relevant AI model.
The researchers published a technical specification and working code for individual FlashProve components on GitHub under an open license. Work on the system is ongoing, and there are…
Originally published by Google News AI