The Internet Needs Proof of Origin

When generative media makes it easy to replicate anyone's voice, face, or brand, the question becomes whether you can verify who's actually behind a piece of media. We believe building that proof of origin into the internet is among the more urgent infrastructure problems we're facing.

Spoofing someone used to be hard. Faking a CEO's voice required a skilled impersonator. Forging a journalist's reporting required painstaking effort. Impersonating a public figure or counterfeiting a brand's communications convincingly enough to fool anyone required resources that most people simply didn't have.

That barrier is gone. In 2024, scammers used AI-generated deepfakes on a video call to steal $25 million from British engineering firm Arup by impersonating the CFO and other colleagues. In February 2026, a cloned voice helped fraudsters push the president of Italian private bank Fideuram into authorizing about €95 million in transfers. The same year, scam ads used a deepfake BBC anchor and cloned pages from the BBC, the Guardian and the Financial Times to sell fake investment platforms. The tools to create convincing fake images, audio, and video are now accessible to anyone with an internet connection and a few dollars.

Deepfakes didn't invent impersonation or manipulation, but they made it effortless. What once required real skill and resources now takes minutes and costs nearly nothing. A creator's reputation, a brand's visual identity, even someone's own face and voice are no longer sufficient to establish that a piece of content is genuinely theirs, because anyone can replicate them convincingly. The floor of trust has dropped out, and it changes the fundamental question we should be asking. It's no longer "Is this real?" It's "Where did this come from?"

This is the problem we've been working on at Verda.

Why Detection Isn't Enough

The instinctive response to the rise of synthetic media has been to build AI models that spot AI-generated content. If machines can create fakes, surely machines can catch them, right? We believe this is a losing game. Detection gains have historically been temporary. The dynamic is inherently asymmetric: detectors rely on artifacts specific to current generation models, but those artifacts tend to disappear as the models improve.

Intel built a deepfake detector that analyzed blood flow patterns under the skin, subtle color changes caused by your heartbeat that generators couldn't replicate. It claimed 96% accuracy. When the BBC tested it in 2023, it flagged several real videos as fakes.

There's a deeper problem, though. Detection can only tell you something might be fake. It cannot prove provenance. When a detector says "90% likely AI-generated," what about the 10%? And even setting accuracy aside, a real video can be edited to misrepresent someone, a genuine audio clip can be taken out of context, an authentic image can be mislabeled. Even perfect detection wouldn't answer the question that actually matters: who made this, and is it what they intended to say?

Detection asks "is this fake?" Provenance asks "who made this?" When spoofing someone is nearly free, the second question becomes more important. Judging content one piece at a time doesn't scale. Checking where it came from does.

Proving Provenance at the Source

What if content carried proof of its origin from the moment it was published?

This is what watermarking tied to a cryptographic signature does. The closest analogy is probably SSL certificates. When your browser shows a padlock icon, it's not detecting whether a site is legitimate by analyzing how it looks. It's verifying a cryptographic signature that proves the site's identity. You're not guessing, you're checking proof.

Watermarking works along similar lines for media. Before any content is watermarked, the person or organization behind a Verda profile completes a secure identity verification flow, so every watermark traces back to a confirmed identity. When a creator or brand publishes content through Verda, an imperceptible signal gets embedded throughout the media (images, audio, or video), linking it to a verified identity. This isn't metadata that can be stripped away or a visible logo that can be cropped out. The signal is woven into the content itself, designed to survive compression, resizing, reposting, and the messy reality of how media actually travels across the internet. Standards like C2PA already attach signed provenance to files, and Verda reads and writes them. The problem is that most platforms strip that metadata on upload. Verda's watermark carries provenance through the places metadata doesn't survive.

Verda becomes the canonical record: creators publish through it, distribute across platforms, and every copy that keeps the watermark traces back to a signed record of the origin.

How the Security Model Works

Content published through Verda carries a verified watermark. That watermark resolves to who published it. That's a positive claim, and it's the one that matters: a brand, a newsroom, or a creator can prove a piece of content is theirs, wherever it ends up. As adoption grows, the absence of that proof will start to stand out too, the same way a missing padlock does today.

But the watermark reveals more than presence or absence. It also reveals integrity. Say a politician gives a speech that gets watermarked throughout the entire audio file. Someone downloads it, carefully alters a sentence in the middle, and reposts the edited version. The watermark pattern, which should be continuous across the full recording, now has a discontinuity. That discontinuity is a sign the content was altered, and a clue to where. Video works the same way: if someone edits specific frames to change what a protest sign says, or cuts a speech to reverse its meaning, the edit disrupts the watermark pattern around those frames.

So when manipulation is attempted, there are a few outcomes. The watermark is disrupted, which signals tampering. Or the watermark persists through a subtler edit, in which case anyone can compare the circulating version against the original in Verda's provenance registry and see what changed. Or someone strips the watermark entirely, which leaves the content with no way to prove where it came from. A forger can remove a verified origin, but can't borrow one.

What This Infrastructure Enables

We think about Verda as an identity layer for media on the internet.

Creators and brands get verifiable provenance that doesn't depend on any single platform. When they publish through Verda, a watermark linked to a signed record of their verified identity travels with the content, wherever it ends up.

A video forwarded across three platforms, a screenshot circulating without context, an audio clip in a group chat. Audiences can check any of these against Verda's registry to find who originally published it, and see the original they published.

Platforms get a path through the moderation challenge. Rather than making subjective editorial judgments, they can surface provenance signals. California's SB 942, in force since August 2026, already requires large generative AI providers to embed provenance disclosures in generated content, and the EU and India are moving the same way.

To be clear, Verda is not a truth arbiter. We provide verifiable provenance so that people and platforms can make more informed decisions. The trust layer, not the truth layer.

The Window is Closing

Provenance infrastructure is most effective when it's established before the noise becomes overwhelming. Building a registry of verified content is easier when people can still mostly distinguish authentic media from synthetic. That window is narrowing quickly, and you can't go back and watermark content that was published without it.

Think about SSL. Most websites once operated without encryption and that was considered normal. Today, the absence of a padlock isn't just unusual; it's a warning. The default expectation flipped.

We believe media is heading toward a similar inflection. Right now, most content carries no proof of origin, and that's considered normal. As synthetic media saturates the internet, the default will shift. Content with verified provenance will be trusted. Content without it will be questioned. The infrastructure that makes that shift possible needs to be in place before it's needed, not after.

That's what we're building.