We live in a post-truth society.

With generative AI getting indistinguishable from reality, society is in a bit of a pickle.

Amazing technological advancements have led us into this ditch of "what's real?", maybe they can also get us out.

Below is a collection of ideas on how we as humans can address and fix some of these problems.

The idea

Imagine if all the videos/images/audio you saw from news outlets contained:

1. A link to the raw unedited source material
2. A full transaction history of how the media has been edited, cut, filtered, and spliced together

Now imagine this was baked into the editing & AI software we all use.

Opt-in full transparency mode for the world and web.

Background

How do we know the video we are watching or the image we are shaking our heads at is real? How did we ever?

In truth, this problem isn't new. Doctored images, audio, and video have happened since we were able to capture these mediums.

What has changed? Well, for one, producing doctored materials has been streamlined and democratized to the masses via generative AI. The barrier to entry and cost associated with fake media is at an all-time low and will only become more commoditized.

This presents a growing threat to societies around the world. Without a shared basis of reality, very serious problems are bound to emerge. A he-said-she-said of epic proportions with no real way to return to baseline reality.

Attempts have been made at addressing "mis" and "dis" information but all are really at the mercy of the foundational problem of "what is real". Fact checkers, censorship, and a ministry of truth are not a solution to the problem. These "solutions" may be well-meaning (or not) but suffer from their own biases and "mis/dis" information within their own worlds.

So how can we address these problems from a technological point of view?

First, and most disappointingly, we can't ever know for a fact what is true, BUT we can with far greater certainty know what is false.

The media in all its forms today lacks provenance, i.e. where it came from and, more importantly, how it was edited or manipulated. This is the heart of the issue that drives most narratives we see in the world today.

For example, a video of politician XYZ saying something out of context. Or an image passed through filters to tell a different story around its contents. Or outright doctored, highly editorialized content. I will let you fill in the blanks on where and when these things might have occurred in recent history.

All of these instances of "fake news" worked on a subset of the population as designed and elicited the desired response. They spread like wildfire and even if they are later "corrected" by the spreader of this information, it's already too late. The damage is done, again by design, and the story people tell themselves is mostly set. This can no longer stand.

Proposed solution

We need a new baseline protocol that all programs use to create an immutable log of all actions taken upon a piece of media.

For example, an image is edited in Photoshop and a person is cropped out of a photo. The steps taken to perform this action are in the .psd file but not in the final image that is then shared. What we need is a new mechanism in which any action taken upon a piece of media is logged and published alongside the final product.

This could be as simple as a pointer to the "media edit history" transaction log as a piece of metadata on the shared image.

Additionally, a pointer to the original unedited media should also be included as a piece of metadata on the newly created artifact.

Spec

Every published piece of media carries a manifest along these lines:

{
  "device": {
    "make": "Sony",
    "model": "A7 IV",
    "softwareVersion": "3.1.0"
  },
  "sourceURL": "https://cdn.example.com/raw/full-uncut-recording.mp4",
  "editHistory": "https://ledger.example.com/edits/8f2c91d4",
  "creators": ["Example News Org", "Jane Reporter"]
}

• device - type of device that captured the media. Make/model/software version
• sourceURL - immutable link to original source material.
• editHistory - pointer to immutable link of the entire edit history of the source file, which programs & LLMs were used, with timestamps.
• creators - (optional) list of all organizations and/or people associated with the creation of the underlying media. This is optional due to anonymized reporting etc.

Legislation

Government legislation around media provenance could happen in two stages.

First, require all media companies to include links to the full source material they are sharing. Such a simple change to media would have a dramatic effect on what people will believe. Call it the "Media Transparency Act" or whatever. This does create additional burden on media companies, so maybe it only applies to a subset of categories, specifically around politics & political campaigns.

Second, we might not even need the government here. Just brave companies that will commit to full transparency on the narratives they are producing. Those who do not will start to have their credibility wane over time (as we are already witnessing in the "news" business).

Caveats

What I'm proposing is a massive change to how all editing software works today. Implementing a standard around a "media edit transaction history" won't happen overnight and not everyone will implement it.

Additionally, not everyone will buy the edit transaction history of a given piece of media. That's fine. Tin foil hats exist. I get it. But media that is stamped with an "edit history" and a link to original unedited source material will stand far and away as more trustworthy than your standard "who knows if this is real" piece of media.

Will this stop lies? No, you can't stop lies. But you can make it increasingly hard to lie. Media companies could still spin their tales, but with pointers to original video and an edit history, the viewer is now in control of how they wish to incorporate the information they have just received.

What if media companies and others don't adopt this new transparency standard? That's fine. They can share media that has no trail, but the more people know about the alternative "media edit transaction log and original source" movement, the less seriously they will take untagged images.

This is one proposed solution. Open to others (as long as they don't rely on manual human intervention, which is insecure by nature).

Dual use problem

This solution doesn't come without its own risks.

The edit transaction log and provenance could be forged.^[1] So we'd likely need cryptographic security and signing up and down the entire chain, and ways to monitor the integrity of these processes.

flowchart LR
  A["Camera<br/>signs at capture"] --> B["Raw file<br/>+ signed manifest"]
  B --> C["Editing software<br/>signs every edit"]
  C --> D["Published media<br/>manifest + sourceURL"]
  D --> E["Viewer<br/>verifies signature chain"]

If exploited, this mechanism could be used for "verified" forgeries to be even stronger propaganda. No bueno.

Update: this is now happening

I sketched this idea in May 2024 and sat on it. Publishing it now, the wild part is how much of it has actually shipped.

The standard I was asking for exists: C2PA Content Credentials^[2] - cryptographically signed manifests that record where a piece of media came from and every edit made to it. Cameras from Leica, Sony, Nikon, and Canon sign photos at capture, and Google's Pixel phones now sign every photo by default with hardware-backed keys.^[3] The major AI generators embed credentials in everything they output - OpenAI stamps images with both a C2PA manifest and an invisible watermark.^[4]

That invisible watermark is the complementary half of the stack: Google's SynthID bakes the signal into the pixels/audio/tokens themselves at generation time and has marked over 100 billion pieces of content.^[5] The two approaches cover for each other's weaknesses. Manifests carry the full edit history but get stripped the moment someone screenshots or re-encodes; watermarks survive re-encoding but carry no history.

Regulators showed up too. The EU AI Act requires machine-readable marking of AI-generated content starting August 2026, and California's SB 942 is already in effect.^[6]

What still doesn't exist is the piece I care about most: the link to the raw unedited source material. Provenance today can tell you an image came from a real camera and was cropped in Photoshop. It still doesn't hand you the full uncut video so you can judge the context for yourself. That's the part worth building next.

Notes