Anthropic introduces FlashProve, a new open-source tool to verify AI models

Have you ever felt that a premium AI chatbot is giving you low-quality, cheap answers? Many in the tech industry worry that some AI providers might be secretly swapping expensive, high-end models with cheaper ones behind the scenes to save costs. To solve this trust issue, Anthropic, the maker of the popular Claude AI, has introduced an open-source prototype called FlashProve.

This new system works like a digital lie detector, giving you a mathematical guarantee that the answer you received actually came from the specific AI model you requested.How the FlashProve tech works

FlashProve acts similarly to a digital signature on a secure email. Before an AI lab releases a model to the public, it publishes a unique digital fingerprint of that model. When you ask the AI a question, the server sends back your text answer along with a special cryptographic proof. The brilliant part is that FlashProve uses a complex security technique called a "zero-knowledge proof" (zkSNARK).

This allows anyone to verify the AI's answer without the company having to reveal its highly guarded, multi-million-dollar underlying code, known as model weights. Because it completely protects a lab's most valuable secret assets, commercial AI developers can use it safely.Massive scale and open-source release

Anthropic did not build this massive project alone. They collaborated with researchers from MIT, EPFL, StarkWare, and Succinct to bring FlashProve to life. To prove the technology works on a massive scale, the team tested the system on a giant 400-billion-parameter AI model called Llama 4 Maverick.

According to the performance data, generating the proof for a 1,024-token request creates a file of about 19 MB at a speed of 110 tokens per second using a single powerful H100 graphics chip.

While making the proof requires heavy computing power from the AI company, checking it is incredibly easy for the end user. A person can run a quick check on a standard home computer CPU in just about 17 seconds to verify the AI's true identity.

To push the industry forward, Anthropic has released all the working code and technical details for free on GitHub under an open-source Apache 2.0 license, allowing anyone to study.The future of AI trust

Beyond catching sneaky cost-cutting, FlashProve has other major uses for businesses. For instance, a developer could use it to prove that their software was genuinely checked for viruses by a top-tier AI, or a startup could mathematically prove that an advanced AI approved its business plan.

While the technology is a major breakthrough, you will not see it in your daily AI apps right away. Anthropic states that FlashProve is currently an early research prototype limited to 1,024-token conversations. Generating these mathematical proofs is still quite expensive, taking about 72 to 95 times more computing time than standard AI operations.

For now, Anthropic has no plans to roll it out in a live commercial product or API.- EndsAlso Read | Anthropic rolls out Message Threads in public beta to make Claude AI much fasterAlso Read | Anthropic warns it is nearing Self-Building AI, what it means for the futureAlso Read | How Anthropic's $100 billion IPO underscores India's need to hike R&D spend on tech

I am a journalist covering the ever-evolving consumer tech landscape. From smartphones to laptops, I track the latest drops and decode the data to see if the newest gadgets are worth the hype.