Cryptographic proofs can verify which model generated an AI response

Anthropic, alongside academic and cryptography researchers from MIT, EPFL, and StarkWare, announced an open source project named FlashProve. According to a research update published by Anthropic, the system relies on zero knowledge proofs to generate a verifiable digital receipt whenever an artificial intelligence model generates text. This receipt confirms the precise AI model that produced the answer without exposing the model's underlying software weights.

When people use third party services or aggregators to query AI tools, they currently must trust that the middleman actually queried the promised system. Middlemen could secretly direct queries to cheaper, lower quality alternatives to save money. Cryptographic proofs like FlashProve eliminate this blind trust by letting users verify that their responses came directly from the expected model. The technology could also assist independent auditors and regulators who need to confirm that companies are serving approved AI versions.

However, achieving this level of authenticity requires substantial computing power. Anthropic reported that generating these receipts takes between 70 and 95 times more graphics processing unit time than simply generating the text itself. It remains to be seen whether consumer demand for guaranteed authenticity will ever justify such a massive hike in hosting expenses. Unless hardware costs drop dramatically or proof generation becomes significantly faster, verifiable inference will likely remain restricted to high security enterprise uses rather than everyday consumer chat applications.