Anthropic is working on a system to prove AI models’ identity
techcrunch ·
Anthropic is developing an open-source system called FlashProve to cryptographically verify which AI model generated a specific output. This technology aims to provide 'provable inference,' ensuring that users can confirm the authenticity of responses from claimed models. The system protects against impostors, including scenarios where a provider might secretly swap models to cut costs or inflate performance metrics. To function, labs must publish unique fingerprints for their models without revealing proprietary weights, allowing verification on local devices. Researchers have released technical specifications and code on GitHub as part of ongoing work rather than a final product. The goal is to enable third parties to verify claims about which advanced models performed specific tasks. Anthropic explicitly shares this intermediate progress to encourage community improvement and collaboration.
techcrunch