Anthropic is developing an open-source system that can verify which AI model produced a given output, according to a blog post the company published on Tuesday.
اضافة اعلان
The system is called FlashProve and aims to provide what researchers call "provable inference," a cryptographic proof that offers mathematical assurance that a particular output came from a specific model. It allows users to confirm that the answers they receive really came from the model they believe they are using.
FlashProve works in a similar way to encrypted email signatures, which let recipients verify that a message genuinely came from the sender it claims to be from. The system is designed to counter attempts to impersonate AI models, even when the party attempting the impersonation is the company that developed and operates the model.
The researchers point in particular to the system's potential to stop AI companies from secretly swapping a model for a cheaper one to cut costs, or from presenting results from a more expensive model as if they came from the original.
FlashProve can also be used to verify third-party claims that specific tasks were carried out using advanced AI models.
As the researchers wrote: "A developer might claim that a software library they built was checked for security vulnerabilities using an advanced cybersecurity model, or a company might claim that its business plan was approved by an appropriate model. How can a user verify these claims?"
For the system to work, AI labs would need to publish a digital fingerprint for each model they release, so that outputs can later be matched against it and verified using cryptographic techniques.
Importantly, these fingerprints do not reveal the model's weights, the internal numerical values that determine how it works and behaves, which companies developing closed models such as Anthropic and OpenAI keep secret.
This means AI labs could publish digital fingerprints without disclosing one of their most important technical and commercial assets, which would allow the system to be used with commercial models as well.
The system is also designed to be lightweight in terms of computing requirements, so verification can be carried out directly on local devices.
Alongside the announcement, the researchers published detailed technical specifications and runnable code for different components of the system, made available to the public under an open-source licence on GitHub.
Research on the system is ongoing, and there are currently no plans to deploy it at Anthropic or any other AI lab.
The researchers stressed in the announcement: "This research post aims to share the progress we have made so far, rather than presenting a completed research project."
They added: "We are publishing these results with a clear goal: to give others the opportunity to look at what we have done, benefit from it, and work to develop and improve it."
