Anthropic's FlashProve wants cryptographic proof of which AI answered you

Anthropic pulled back the curtain Tuesday on FlashProve, an open-source system designed to answer a question that gets stickier by the day: did the model you think you were talking to actually answer you? The research team published a technical spec and working code for chunks of the system on GitHub, per a blog post first reported by TechCrunch. Don't expect it in a product anytime soon, though. Anthropic says the work is ongoing and there are no plans to deploy it at its own lab or anywhere else.

The core idea is what the researchers call provable inference, a cryptographic proof that a specific model generated a specific output. Think of it as a signature for AI answers, except it verifies the model rather than the sender. To make it work, labs would publish a fingerprint of each model they ship, and that fingerprint can later be matched against real outputs and checked mathematically. Users could then confirm the answer they got really came from the model they believe they're using.

Here's the clever part. Those fingerprints don't spill a model's weights, the internal numbers that dictate how it behaves and that closed-model shops like Anthropic and OpenAI guard fiercely. That means labs could publish fingerprints without handing over their crown jewels, which is exactly what makes the scheme viable for commercial models and not just open ones. The system is also built to be lightweight, so verification can happen right on a local device.

FlashProve is aimed at impostors, and not just the usual ones. According to the researchers, it would hold up even when the faker is the lab that built and runs the model itself. That heads off a sneaky scenario: an AI company secretly slotting in a cheaper model to trim costs, or a pricier one to pump up its performance numbers.

There's a second use case for third parties making big claims. A developer might say its software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might say its business plan got a green light from the right model. The researchers pose the obvious follow-up: how can anyone verify that? FlashProve is meant to be the receipt.

The team is upfront that this is a work in progress, not a finished product. In the announcement, the researchers framed the release as a deliberate handoff, saying they're sharing intermediate progress so others can take it and improve on it. The code is out under an open-source license, meaning the real test of FlashProve may be what everyone else builds with it.