Home · Posts
We Owe it to the Future to Certify the Past: Video Authenticity in the Age of AI
April 15, 2024

AI-generated video is rapidly improving. In the future, it seems likely that the quality of video that AI can generate will surpass the quality of most genuine video existing today.

In the realm of physical artifacts, we have carbon dating and other material-based methods to tell us how old something is. These techniques are critical to our collective ability to make sense of the past. Without them, not only would we have trouble piecing together the story of what happened, given real evidence, but it would also become much easier for people to create fake evidence. History would be easier to rewrite.

As the digitization of our culture meets rapid advancement in AI-generated video, the lack of carbon dating for data threatens to create a crisis for future generations. At the risk of using a trite example, consider an Orwellian leader in 2100 who claims "We were never at war with Eurasia". Video evidence of war with Eurasia would be meaningless—it could easily have been artificially generated. Or, imagine a debate over whether coral reefs ever existed. Someone could show you videos of coral reefs that, if you looked closely, were obviously faked. "See, they're an old-timey hoax, right?" Reefs do seem fantastical… and more realistic video of them would be no more convincing: we only had the technology to build fake reefs in 2000, but in 2100 we can generate fully realistic video of these imaginary organisms in even greater detail.

Luckily, we have a technology that allows us to more-or-less prove to future humans that a given digital artifact existed at a certain point in the past. We can give them extremely strong evidence that a video—of coral reefs, say, or a war—came into existence no later than a certain point in time. In what might be a stroke of extreme luck, this technology happened to be invented about ten years before AI-generated video became possible.

Blockchains are a kind of database that can only be added to, not retroactively altered. Because these database additions happen at a predetermined rate, we’re able to know within a narrow range of time when any given entry was created. Combined, these two properties let us create a tamper-proof and decentralized certification that a digital artifact, such as a video file, existed at a specific point in time.

By accumulating more and more proof-of-work over time, and distributing their data openly over nodes around the world, they make faking these certifications so expensive as to be impossible. To modify a piece of data previously recorded on the blockchain, an attacker would need to redo all the mining work from that block onwards—which would be both prohibitively expensive and immediately apparent to participants in the decentralized network. These barriers would effectively stop attempts to rewrite history.

In particular, we should store the cryptographic hash (a small, unique fingerprint) of a video file in the blockchain, because storing the entire video file is beyond the ability of current platforms.

There are different options for which blockchain to use. The ones that currently seem most secure and most likely to survive are Bitcoin and Ethereum. But there are increasingly many options, including lower-cost “layer-two” chains built on top of these leading candidates.

If storing data on a popular level-one chain like Bitcoin or Ethereum is too expensive, relative to the benefit of certifying a particular video, multiple hashes can be combined into one. The list of hashes thus combined can be stored on a lower-cost blockchain, or offline, as long as any copy survives. This brings the per-video cost of certification close to zero. 

Under this scheme, in the future, a video file paired with a hash certified on a blockchain will be much more likely to have existed in the past than one without the certification.

Importantly, a committed hash isn't proof that a video is real, merely that it existed at a specific point in time—the time that it was committed to the blockchain. Faked videos can be certified just as easily. But with certification, we’re better able to evaluate authenticity through several methods. The main one is technological: comparing the quality of the video to the quality of generative AI tools available at the time its hash was committed. In addition, we’ll have cryptographic evidence of who committed that certification, and in some cases a sense of whether we trust that entity. Finally, we’ll still be doing what we do now: looking for consistency between this evidence and the rest of our evidence about the past. It will come down to a preponderance of evidence. If we have multiple certified videos of the same event, at a time when creating such sets of videos wasn't possible with AI, that's good evidence. If we have certifications committed by entities we trust, that's more credible too.

Note that the quality of real video we capture—in terms of resolution, frame rate, and color—is increasing as camera technology improves, both at the high end and on the mass consumer level. It's tempting to hope that genuine video quality will always outpace the capabilities of cost-effective fakery. After all, simulating realistic physics, lighting, and complex scene details at high resolutions is computationally intensive. But the quality of generated video is increasing too, and the latter will likely increase faster if it follows the same trends as software in general.

As with photographic evidence in the past—and Hollywood movies in the present—a sufficiently well-funded entity can create convincing fakes at any time. But with blockchain certifications, we’ll have a much stronger estimate of the point-in-time historical cost of doing so, and therefore much better evidence regarding what is true.

Implementing such a system will require careful choices of blockchain platforms and external storage options that balance security, scalability, cost efficiency, and ecosystem support. With the right choices, on-chain hash data will almost certainly survive. But the fact of their existence, what they mean, and the corresponding video files will have to be preserved as well.

We can give the future a defense against the erosion of truth. It’s not a perfect solution, and it won't eliminate the need for human judgment and social consensus, but it can provide a crucial layer of strong evidence for appraising digital artifacts. We owe it to future generations to act on this opportunity and defend the truth.