Anthropic is developing an open source system designed to verify which artificial intelligence model produced a specific piece of content. The initiative, detailed in a blog post published on Tuesday, aims to provide users with a mathematical guarantee that an output originated from a claimed source. The system is called FlashProve.
Researchers describe the technology as a method for “provable inference.” It functions similarly to cryptographic email signatures, which allow recipients to confirm that a message truly came from the sender it claims to be from. This approach is intended to prevent impostors from disguising one model as another. The system is designed to be robust even against the labs that build and run the models themselves.
Preventing Model Swapping and Fraud
The primary motivation behind FlashProve is to stop companies from secretly swapping models to cut costs or inflate performance metrics. For example, a provider might be tempted to use a cheaper, less capable model while charging customers for access to a more advanced one. Conversely, a company might use a superior model to generate better results for a report, then claim the work was done by a less capable system to save money on future contracts.
The researchers note that the system could also verify claims made by third parties. A developer might assert that their software was scanned for security flaws by a top-tier cybersecurity model, or a business might claim its strategy was approved by a specific advanced AI. FlashProve would allow users to check these assertions cryptographically.
Protecting Proprietary Weights
To function, the system requires AI laboratories to publish a “fingerprint” for each model they release. These fingerprints can later be matched against specific outputs to verify their origin. A critical feature of this design is that the fingerprints do not reveal the model’s weights. Weights are the internal numbers that define how a model behaves, and they are closely guarded secrets for closed-model makers like Anthropic and OpenAI.
By keeping the weights private, labs can publish the necessary fingerprints without giving up their most valuable intellectual property. This allows the verification system to work with commercial models. The system is also built to be lightweight, enabling verification to take place on a local device rather than requiring heavy cloud infrastructure.
Alongside the announcement, researchers released a detailed technical specification and working code for various parts of the system. The code is available on GitHub under an open source license. Anthropic stated that the research is ongoing and there are no current plans to deploy FlashProve at their own facilities or at other laboratories.
“This research post is meant to share intermediate progress, and not a completed research agenda,” the researchers wrote in the announcement. “We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it.”
Source: TechCrunch
