Anthropic is developing technology that makes it possible to verify whether a response was actually generated by the artificial intelligence (AI) model that a company claims to use. Named FlashProve, the system uses cryptography to create a kind of mathematical certificate of the identity of the model responsible for a given output.
The aim is to give users a way to confirm that they are interacting with the AI advertised by the provider, rather than a different or less advanced version. According to Anthropic, the mechanism could also make it harder for companies to secretly replace their models with cheaper alternatives or present performance results that do not match the technology being used.
The project was presented in a company publication on Tuesday (3). The company also made a technical specification and parts of the code available in open format so that other researchers can examine and improve the initiative.
How does FlashProve work?
- The system uses the concept of “provable inference,” which consists of generating a cryptographic proof that a given response came from a specific model;
- The idea is similar to the digital signatures used in emails to confirm the identity of the person who sent a message. In this case, however, the verification would serve to identify which AI produced a piece of content;
- For this, model developers would need to publish some kind of fingerprint for each system released. Subsequently, this identification could be compared against the responses produced by the AI and verified through cryptographic mechanisms;
- An important aspect is that these fingerprints would not reveal the model weights — the internal values that determine its operation and behavior. Companies such as Anthropic and OpenAI keep this information private, but could take part in a verification system without giving up that trade secret;
- The technology was also designed to be lightweight, allowing verification to be carried out directly on a local device.
Read more:
- Anthropic says AI can accelerate its own development
- Anthropic's AI model ignores warning and sustains cyberattack in security test
- Anthropic changes Claude Code rules to prevent copies of existing games
System could prevent secret model switching
One of the problems FlashProve aims to address is the possibility of a provider presenting one model to customers while using a different one during the operation of the service.
The company also made available a technical specification and parts of the code in open format so that other researchers can examine and enhance the initiative – Image: gguy/Shutterstock
A company could, for example, advertise access to an advanced AI and replace it with a cheaper alternative to reduce costs.The mechanism could also complicate situations in which a company uses a model different from the one it claims to have used to demonstrate a certain level of performance.
Verification would have applications beyond the services offered directly by AI companies.
The researchers cite the example of a developer who claims to have used an advanced cybersecurity model to examine a software library for vulnerabilities. Another possibility would be a company stating that its business plan received approval from an AI suited to the task.
With FlashProve, it would be possible to cryptographically verify whether the announced model was actually used in these processes.
Technology has no release date yet
Despite the public presentation, FlashProve is still a research project. Anthropic has made a detailed technical specification and working code for different components available under an open-source license, allowing the community to examine and develop the system.
However, the company has not announced plans to implement the technology in its own services or on other AI platforms.
The researchers said the publication represents an intermediate step in the work, rather than the conclusion of a research agenda. The intention is to enable other teams to analyze the proposal, build on its components, and help improve it.
Rodrigo Mozelli is a journalist graduated from São Paulo Methodist University (UMESP) and currently works as a writer at Olhar Digital.
