What Can Be Built with Auto ID
Last updated
Last updated
Auto ID allows honest users who publish content online to demonstrate if and how the content was AI generated, in a manner that anyone can verify, with a high degree of certainty. Lets explore an example with our trusty friends Alice and Bob. Note that most of the below workflow woud be automated in any actual implementation.
Alice routinely publishes content on X (formerly Twitter) and uses OpenAI’s ChatGPT to help write her tweets and generate associated images. Alice has nothing to hide and would be happy to show the world that she produces content with the help of AI, but how can she do this? First, Alice creates her own Auto ID and links to her OpenAI account. OpenAI also creates its own Auto ID and uses it to issue IDs to each AI service it exposes, such as GPT-4 and DALLE. Alice and OpenAI then jointly issue an Auto ID to her own instance of ChatGPT.
Now every time Alice generates content using her GPT she can produce a digital signature of the content that is linked to her Auto ID. She could sign the content directly (or more likely, its cryptographic hash) but then if she changes the content even slightly (as most users do) it will have a completely different hash, and no one will be able to verify its origin. Instead, she can use a different OpenAI service to generate a short embedding, or semantic fingerprint of her content. She can go one step further and transform the embedding into a tiny semantic hash, using a Locality Sensitive Hashing (LSH) technique, which she can then sign using her Auto ID. If she does this for both the original content from her GPT and the modified form that she edits, then these two semantic hashes will be almost exactly the same, as they both refer to text with the same semantic meaning.
Now when Alice publishes her content to X, she can include the content along with a short proof-of-provenance which includes the joint signature of the original semantic hash, and her signature of the final semantic hash. The X platform may then verify this proof, using the public keys for each Auto ID in the registry, and compare the similarity between the two hashes. If they fall within a pre-defined range then X may render a seal, perhaps a green smiling robot icon, that demonstrates this is verified AI content. Now Bob, who follows Alice religiously on X, can then inspect the content and see further details such as the AI model, provider, and timestamp for the signature and be assured that it was AI generated and know that Alice honestly volunteered this information. Note that this not only applies to text, but any form of AI content which a semantic embedding may be computed over, including images, audio, and videos.
At Autonomys, we call this provenance detection and demonstration product Auto Verify.
Auto ID also provides a framework that allows honest humans to prove to existing apps and platforms that they probably actually are a human, and not a bot or AI account. This would allow users to more easily authenticate online by removing cumbersome sign-up and sign-on forms, as well as the increasingly difficult captcha puzzles we have all grown to hate.
Let’s say Alice is an honest human who would simply like to create an account on an a new platform. This platform is inundated with bots and fake users so it employs a series of captchas to deter attackers, but at the cost of exhausting honest users. How can Alice use her Auto ID to signup with a single click?
Let’s assume Alice has several existing online identities through well-known identity providers (such as Google or Twitter) who can attest that she has a verified email and phone number, and she has already linked these accounts to her Auto ID. If the new platform implements sign-on with Auto ID, then it may easily see that Alice does have verifiable accounts with many well-known providers.
The more providers Alice has linked to her Auto ID, the more certainty the platform can have that she actually is a human, and even that she actually is Alice, should Alice choose to make her identity public. Alice can even link emerging proof-of-personhood systems to her Auto ID, such as World ID, which leverages biometrics to register unique human identities in a privacy preserving manner. We can even quantify the notion of how likely an Auto ID is to belong to a human as its human score, as a probability between 0 and 1. We could say that any Auto ID with a sufficiently high human score is an implicitly verified Auto ID. Any platform that integrates with Auto ID could then require that valid accounts have a certain human score, potentially present if a user holds an Auto verified account on their profile, and even allow the user to expose their actual human identity to promote further trust in their profile.
We now turn to Eve, who runs a bot farm and tries to generate as many fake accounts as possible on platforms for a variety of reasons. Before, platforms could only stop Even with captchas. Now Eve must generate or purchase multiple verified accounts across several well-known providers to obtain a reasonable human score, before she has any chance of even being able to create a new account, significantly raising the cost of sybil attacks.
At Autonomys, we call this proof-of-personhood product Auto Score.
