As artificial intelligence becomes increasingly integrated into businesses, software development and everyday decision-making, one important question is becoming harder to ignore: How can users know which AI model actually generated the answers they receive?
Anthropic is exploring a potential solution with a new open-source research system called FlashProve, designed to provide cryptographic proof of which AI model produced a particular output.
The technology aims to address growing concerns about transparency, accountability and trust in the AI industry by allowing users to verify that an answer genuinely came from the model a company or service claims to be using.
According to a research announcement published on Tuesday, FlashProve introduces what researchers describe as “provable inference” — a process that uses cryptography to establish a verifiable connection between an AI model and its generated output.
If successfully developed, the technology could make it significantly harder for AI providers, software developers and other organisations to misrepresent which models are powering their services.
It could also address a particularly sensitive issue in the AI industry: the possibility that companies secretly substitute one model for another without informing users.
What Is Anthropic’s FlashProve AI Verification System?
FlashProve is an experimental system designed to establish the authenticity of AI-generated responses through cryptographic verification.
In simple terms, it would allow someone receiving an AI-generated answer to check whether the response actually originated from the specific model identified by the service provider.
This is important because users often have little visibility into what happens behind the interfaces of AI applications.
A company might advertise that its platform uses an advanced, expensive AI model, but customers generally have to trust that the provider is telling the truth.
FlashProve aims to reduce that dependence on trust by introducing mathematical verification.
Rather than simply accepting a company’s claim about the model behind an answer, users could potentially examine cryptographic evidence confirming its origin.
The concept is comparable to digital signatures used in secure email communications.
With cryptographic email signatures, recipients can verify that a message genuinely came from the person or organisation claiming to have sent it. FlashProve seeks to apply a similar principle to artificial intelligence.
The objective is to make AI model identity independently verifiable rather than something controlled entirely by the companies operating these systems.
How FlashProve Could Prevent AI Companies From Switching Models Secretly
One of the most interesting applications of FlashProve involves preventing AI companies from quietly changing the models used to generate responses.
AI services often provide access to different models with varying levels of capability, performance and operating costs.
More advanced models can require significantly greater computing resources, while smaller models may be cheaper and faster to operate.
This creates a potential transparency problem.
For example, a company could advertise access to a powerful frontier AI model but secretly route some customer requests through a cheaper, less capable alternative to reduce operating expenses.
Users might continue paying for a premium service without knowing that the underlying technology has changed.
The opposite situation could also create problems. An organisation might claim that a particular task was completed using a more advanced model than was actually involved, potentially exaggerating its technological capabilities.
FlashProve is designed to make these types of misrepresentation more difficult.
By attaching verifiable cryptographic evidence to AI-generated outputs, the system could allow users to confirm whether the model responsible for an answer matches the provider’s claims.
This could introduce a new level of accountability into the relationship between AI companies and their customers.
FlashProve Could Change How Businesses Verify AI-Powered Services
The potential applications of the technology extend beyond consumer-facing AI chatbots.
Anthropic’s researchers also highlighted situations where third-party organisations make claims about using particular AI models to perform specialised tasks.
For example, a software developer might claim that an application or software library has undergone security checks using an advanced cybersecurity AI model.
Similarly, a company could advertise that its business plan has been reviewed and approved by a particular high-performance AI system.
In both situations, customers and other interested parties may have no straightforward way to establish whether the claims are accurate.
FlashProve could provide a mechanism for independently checking the identity of the model involved.
This could be especially valuable in industries where the quality and reliability of AI-assisted work matter significantly.
Businesses could use model verification to provide additional evidence about the AI systems involved in their operations.
Software developers could potentially demonstrate which models were used during certain development or security processes.
Organisations relying on AI for specialised analysis could also benefit from having a verifiable record of the technology involved.
However, it is important to distinguish between proving which model generated an output and proving that the output itself is correct. FlashProve is designed to address model identity, not automatically guarantee the accuracy or quality of an AI response.
How the Cryptographic Verification Technology Works
For FlashProve to function, AI developers must publish a unique cryptographic fingerprint associated with each model they release.
These fingerprints serve as identifiers that can later be used to verify whether particular outputs originated from the corresponding model.
When a model generates an answer, the system is intended to provide cryptographic evidence that can be checked against its published fingerprint.
This verification process is designed to establish a mathematical connection between the output and the model responsible for producing it.
An important feature of the approach is that the fingerprints do not expose the underlying model weights.
Model weights are the internal numerical parameters that determine how an AI system processes information and generates responses.
For companies such as Anthropic and OpenAI, these parameters represent valuable intellectual property and are generally kept private in their commercial models.
Requiring companies to disclose their model weights would create a significant obstacle to widespread adoption of any verification system.
FlashProve addresses this concern by allowing developers to publish model fingerprints without revealing the underlying weights.
This means the technology could potentially support both open-source and proprietary AI models.
Commercial AI laboratories would therefore have an opportunity to participate without surrendering the confidential technical components that distinguish their products.
Why Local Verification Matters
Another important feature of FlashProve is its intended lightweight design.
Anthropic’s researchers say the system is being developed so that verification can be performed locally on a user’s device.
This could reduce the need to depend entirely on the AI provider to confirm the authenticity of its own outputs.
Independent verification is particularly important because a system controlled exclusively by the company making the original claim may offer limited protection against deliberate misrepresentation.
Allowing users to perform checks locally could make the verification process more accessible and reduce dependence on centralised infrastructure.
It could also make the technology easier to integrate into different applications and software environments.
Nevertheless, the practical performance, scalability and security of the complete system will depend on further research and development.
Anthropic Releases FlashProve Technical Specification and Open-Source Code
Alongside the announcement, Anthropic’s researchers released a detailed technical specification and working code covering different components of FlashProve.
The materials have been made publicly available under an open-source licence on GitHub.
This approach allows independent researchers, developers and other AI laboratories to examine the proposed system, experiment with its components and contribute improvements.
Making the technology openly available could also encourage collaboration across the AI industry, particularly among organisations interested in improving transparency and accountability.
Rather than presenting FlashProve as a finished commercial product, Anthropic described the announcement as an opportunity to share its ongoing research with the wider technical community.
The researchers emphasised that the project remains a work in progress.
They also made clear that there are currently no plans to deploy FlashProve at Anthropic or any other AI laboratory.
In their announcement, they described the publication as an opportunity for others to examine the work, build upon it and improve the underlying technology.
This suggests that the immediate objective is to encourage broader research rather than introduce a ready-to-use verification product.
Why FlashProve Could Be Important for the Future of Artificial Intelligence
As AI adoption expands, transparency about the technology powering digital services is becoming increasingly important.
Consumers, businesses and software developers want to know whether they are receiving the capabilities they have paid for and whether organisations are accurately representing the AI systems involved in their operations.
FlashProve could contribute to addressing this challenge by making model identity independently verifiable.
If the research succeeds, cryptographic verification could become an important part of how AI services establish credibility.
It could also encourage greater accountability among providers by making undisclosed model substitutions and misleading performance claims easier to challenge.
However, several questions remain.
The system will need continued testing to establish its reliability, security and practical usefulness across different types of AI models.
Widespread adoption would also depend on cooperation from AI laboratories willing to publish the required model fingerprints.
Without participation from model developers and service providers, the technology’s usefulness could remain limited.
Furthermore, verifying a model’s identity would not necessarily establish that its responses are unbiased, accurate or safe.
These issues will require additional safeguards beyond model authentication.
FlashProve Is Still Research, Not a Released Anthropic Product
Despite the attention surrounding the announcement, FlashProve should not be mistaken for a finished feature available to Anthropic customers.
The company has released research documentation and working code, but it has not announced plans to integrate the system into its commercial AI services.
Its development remains ongoing, with Anthropic encouraging external researchers and developers to participate in improving the technology.
The project nevertheless highlights an important direction for the AI industry: moving beyond simply asking users to trust claims about artificial intelligence and exploring ways to prove those claims mathematically.
As AI systems become more powerful and increasingly involved in important business and technical decisions, knowing which model generated an answer could become just as important as evaluating the answer itself.
FlashProve is an early research effort towards that goal, with the potential to introduce a new layer of transparency into an industry where the identity of the technology behind the screen is not always easy to verify.
