- According to the source, Anthropic is developing FlashProve, an open-source system that checks which AI model a specific output came from. The goal is to use cryptographic proof to confirm whether the answer a user received came from the model they believe they are using.
- The source explains that it could be used to stop AI companies from secretly switching to cheaper models to cut costs, or from using more expensive models to inflate performance figures. This is a possible use suggested by the researchers and is not a confirmed real case.
- The method is to publish a fingerprint for each model when it is released, without revealing the weights. Outputs are then checked cryptographically against this fingerprint. The source says the system was designed to be lightweight so it can be verified even on personal devices.
- Part of the technical specification and working code is published on GitHub under an open license. The source says the research is at an intermediate stage and that there are no plans to deploy it at Anthropic or other labs.
O: This works for a scene where a planner who uses AI output as a basis for work records the model name and call time of the answer they received, and sorts out in advance which items to check if a verification tool like this appears later.
X: The original says this is a mid-stage research share and that real service use or verified performance has not been confirmed, so you should first check the scope of use before putting this tool into actual work.