跳到正文
DEV Community· Ahtisham Lala·· 15 天前AI 評分57

Anthropic 的 FlashProve 可用密碼學證明驗證 AI 回答由哪個模型生成

Anthropic's FlashProve Can Cryptographically Prove Which AI Model Wrote Your Answer

AI 導讀

Anthropic 介紹 FlashProve,這套研究系統可為特定回覆生成密碼學證明,驗證它由指定 Claude 模型在指定指令下產生,而不披露輸入、輸出或提示詞。它使用 SNARK,證明涵蓋模型、權重、系統指令及上下文,第三方可在內容保密時進行驗證。Anthropic 表示 FlashProve 仍屬早期研究,尚未作為產品推出。

正文

Ahtisham Lala

Originally published at ConvertElementor.

Anthropic has introduced FlashProve, a system that can generate cryptographic proof that a specific response was actually produced by a specific Claude model with specific instructions — without sharing the input, output, or prompts.

The idea is simple to state and hard to build: we trust AI answers more when we can verify where they came from. Today, a screenshot or a copied answer proves nothing. FlashProve changes that.

How FlashProve works

  • It uses a SNARK (Succinct Non-Interactive Argument of Knowledge) to produce a proof that is tied to the model, its weights, and the system prompt used.
  • The proof does not disclose the prompt or the output — a third party can verify it while the content stays private.
  • The system proved Claude's answers could be authenticated without revealing the input, output, or prompts.
  • Crucially, it also includes system instructions and context in the proof, so you can check that the model answered with the right context, not just that some model answered.

Anthropic describes FlashProve as an early research step, not a shipping product yet. It is built on technology from Project Moonshot, a research effort that is explicitly tackling the hardest open problem in general intelligence: how to build AI systems people can actually trust.

Why this matters for developers

If this kind of proof becomes practical, it could transform regulated work — finance, healthcare, legal, and any workflow where an audit trail matters. A developer could prove which model version answered a question, with which instructions, long after the conversation ended.

It also raises the bar for AI-generated content: verifiable provenance, instead of "trust me, Claude said it."

Read the full article on ConvertElementor: https://convertelementor.com/blog/anthropic-s-flashprove-can-cryptographically-prove-which-ai-model-wrote-your-answer-and-that-changes-the-trust-game-for-developers-november-2026-20261106

來源:DEV Community · dev.to