Why Blockchains Are Useful Counterbalances to AI Sophistication

The Problem

The problem with the internet is that it is an amazing replicator. We can share millions of copies of the same content, through thousands of channels at an extreme pace. This mass transfer of information is an immense equalizer, as the entire world can learn about something simultaneously. 

However, there is a big flaw in the creation of these copies. Time. It is nearly impossible to find the origin of anything, and infinitely harder to deterministically prove that origin point when it is declared.  

This used to mean that we couldn’t say with confidence ‘when’ or ‘where’ a photo was taken, but the sophistication of AI means we have an even larger problem. We can now no longer tell when, where, or ‘if’ something we are seeing happened.

This landscape lends itself to an internet that will be fundamentally broken, an internet that will be populated with synthetic content where the humanity and identity of individuals and the things they create will be impossible to navigate.

Current Solution and Its Flaws: Building on Top of a Broken Infrastructure

So far, the industry's approach has been to play whack-a-mole using deepfake detection. This approach will not work for much longer and is already failing. The signals deepfake detectors use are becoming harder to read, so the images and videos that are created are besting not just our own eyes, but the technology stack that currently exists to solve this problem.

What you are left with is to trust content for reasons other than the pixels. There are two ways you can generate this trust:

• You trust it because someone says so.

• You trust it because there's a technological basis.

To do this, we need to reliably verify, at a pixel level, the actual components that create a photo—a pixel-perfect trust.

Limitations of Detection

The reason detectors are so popular is because they don’t require you to change any workflows. They work post-hoc. Someone has sent a video, you don’t know where it came from, and you have no source of truth. These detectors are your only option left. They are a perfect solution in the ideal format. But that does not exist. If this deepfake technology is fooled, that creates a huge problem, and the onus will be on the photo editors who get fired because they publish fake content. The flip side may be even more harmful. Since these detectors don’t want to give out any false positives, they get stricter and stricter, and they start labeling real content as fake, and that becomes censorship for no reason at all—just because you wanted to portray an image of a highly resilient deepfake detector.

Root of the Problem: Adversarial Learning

The way these AI systems get better is by using adversarial learning. You pit the detector against the generator until the generator fools the detector. The better the deepfake detector, the better the generation in the next cycle. 

One early example was that deepfake generators were very bad at blinking. Detectors were initially formed to look at the blinking cadences in these videos and determine if they were real. By using adversarial learning, that was corrected very quickly. This has gotten so advanced that detectors are looking at the blood underneath the skin and seeing if it pumps through the body, and even this is being corrected. Even more concerning is that no human is going in and correcting it; these generators simply try a bunch of variations until they start fooling the detector. 

We have entered an arms race where these models are learning how to become more human, and how to better fake content by using the very detectors themselves.

Rethinking the Internet’s Infrastructure

When we think of the physical world, we have fossil records, these help us determine when something in fact happened and how. We need the same technology for the internet. A way of building digital fossil records, definitive and immutable time-stamps for content to prove that it happened, and to prove when it happened. 

The solution is to rethink how we approach the internet and the infrastructure of media circulation. In its current format, the internet is extremely good at replicating content. When something’s posted on the internet, you can screenshot it and send it through a thousand different channels. It is easy to share across the internet, but it’s extremely hard to find the origin points and provenance of that content. In a world where the photos and videos on the internet lose their evidentiary value, you have to look at restructuring the infrastructure behind the internet itself—building origin point content into media circulation.

We are helping facilitate a new generation of the internet that looks to build origin points and build provenance into all media assets online so you can always tell where something came from and if it is real.