Could you prove you're a real person—or that you're authorized—without ever revealing who you actually are?

This explores whether you can prove something about yourself, such as being a real person or being authorized to act, without giving away who you are, and whether that holds up when the information is sensitive. The short answer from the collection: the cryptography mostly works, and the risk sits around it. One caveat first: the corpus has no material on whistleblowing or other high-stakes disclosure as such. What it does have is a set of building blocks and a set of warnings about where they leak.

The most direct building block is the personhood credential Can people prove they are human without revealing who they are?. A trusted institution issues it, and it lets you prove you're a real human rather than a bot without revealing which human you are. A related trick comes from agent auditing Can commitments protect sensitive agent data while enabling verification?. Instead of publishing a sensitive record, you publish a cryptographic "commitment" to it, a fingerprint that later proves the record existed and hasn't been altered. That separates proof from disclosure, which is exactly what a high-stakes discloser wants. The catch is that someone still has to hold the actual content. Questions about who can access it and whether it can ever be deleted don't go away.

The surprise is that the weak point is usually the system that checks your identity, not the math. One proposal anchors AI agent actions to the humans who authorized them using extended OAuth tokens, the same system behind "Sign in with Google" How can AI agents prove who authorized their actions?. The proposal's own authors admit that routing every check through providers like Google or Facebook gives those providers a new view of everything the agents do across services Do OpenID Providers gain surveillance power over AI agents?. A scheme can be anonymous to the service you're talking to and still fully visible to whoever sits in the middle. For a whistleblower, that middle party is often the exact actor they need to avoid.

AI adds new leaks around even a well-designed scheme. Language models that reason step by step tend to recall private user data inside their reasoning, and longer reasoning leaks more Do reasoning traces actually expose private user data?. Encrypted reasoning blocks that providers send back to users turned out to be swappable between models, so a weaker model could decrypt and repeat a stronger model's hidden reasoning Can cheaper models decrypt traces from stronger models?. Separately, most users never noticed that their chatbot had inferred and stored private details about them Do users notice when LLMs infer and retain private information?. If you draft a sensitive disclosure with an AI assistant's help, your anonymity can be lost long before any credential is checked.

One more twist: anonymity changes behavior in both directions. People inclined to cheat prefer reporting to machines over humans, because a faceless interface makes lying feel cheaper Do dishonest people prefer talking to machines?. The lower social cost that protects an honest whistleblower also lowers the bar for bad-faith reports. That's why schemes like personhood credentials try to prove one narrow fact, real person or authorized party, while hiding everything else. Whether that balance holds for high-stakes disclosure is an open question the collection doesn't yet answer.