The real threat isn't convincing fakes fooling us—it's real footage being dismissed as fake. Provenance, not detection, is the only way out of this mess.

By Ben Anchor — Wednesday, 24 June 2026 · Listen to the podcast episode

The French president gets shoved playfully by his wife as he boards a plane in Hanoi. Within hours, half the internet is calling it AI-generated. The palace has to issue a statement: no, that actually happened. Hold that shape in your head, because it's the entire problem we've been getting wrong.

We've spent two years terrified of the convincing fake—the fabricated clip of a leader declaring war, the cloned voice moving money. But the danger that's actually arrived is the inverse: real footage getting waved away as synthetic. The liar's dividend, they call it. The profit a liar earns simply from the fact that fakes now exist.

I recorded a bonus episode on this recently—provenance standards, C2PA, the whole machinery—and the conversation with the engineers at Amazon's media symposium last year is what crystallised it for me. We've been fighting the wrong battle. Detection is an arms race you're structurally guaranteed to lose. Every detector you build becomes free training data for the next generator that wants to slip past it. You're always one step behind, asking exhausted newsrooms to run forensic labs on every frame that comes in.

The serious people stopped asking is this fake? and started asking where did this come from? That flip—from detection to provenance—is the single most important idea in this entire field.

Provenance Isn't a Sticky Note Anymore

Here's how it works. When a piece of media is created, the device or software wraps it in a small sealed record: who made it, with what, when. The obvious hole: the moment that file hits a social platform, the platform strips the metadata, re-encodes the video, throws away everything you just described. Your beautiful sealed record dies on first contact with the real internet.

Except that's changed. The newer versions of C2PA—the content authenticity standard—tie the credential to the pixels and the audio using a robust watermark. Even if a platform strips the visible record, the link survives cropping, resizing, a re-encode. You can fetch the full history back from the cloud. Provenance stops being a sticky note and becomes part of the grain of the thing.

And the late 2025 revision extended signing to live streaming, segment by segment, in the formats broadcasters actually push. This is no longer just a way to certify a still photograph after the fact. You can sign a live feed as it goes out.

Who's Actually Shipped This?

On capture: Samsung put it in the Galaxy S25, the first major phone to sign photographs in hardware at the moment they're taken. Google followed in the Pixel line. Canon pushed firmware onto its top mirrorless bodies last summer—the R1 and the R5 Mark II. Sony has shipped a broadcast camera that signs at source.

On the other end, where people actually look: LinkedIn shows a little credentials icon you can click to see an image's history. TikTok adopted it for labelling AI content at consumer scale. YouTube and Meta are surfacing it too. The rails exist now, from the lens all the way to the feed. They're thin in places, but they exist.

So why am I not declaring victory?

The Liar's Dividend Still Pays Out

Because a bad actor can still strip a credential and shout "deepfake" at anything inconvenient. Real audio surfaces, the politician calls it AI. Genuine body camera footage appears, the defendant's lawyer calls it synthetic. It's already happened in courtrooms.

Provenance doesn't magically defeat that on its own. But it changes the board. If that camera in Hanoi had signed the file at the moment of capture, the argument is over in minutes, not days. There's a sealed, verifiable record that says this came from this device, at this time, unedited. The liar can still lie. But now the lie has to survive a check that anyone can run.

And over time, the absence of any credential starts to mean something too. The unsigned clip becomes the suspicious one. That's the cultural shift. We stop asking footage to prove it's fake, and start expecting important footage to prove where it's from.

The Regulators Are Moving

Under the European rules, transparency obligations for AI-generated content land this August. In plain terms: providers will have to mark synthetic output in a machine-readable way, and anyone deploying a deepfake will have to disclose that it's artificial. There's a carve-out for obvious satire and art, but the direction is set. Machine-readable marking is exactly the language provenance speaks. The law and the standard are converging from two directions.

And broadcasters sit right at the choke point, because we're the ones who take raw material in from the world and hand it to the public with our name on it.

What You Actually Do on Monday Morning

Three things, none of them exotic:

1. Sign at the source. If you're buying cameras or phones for field work, provenance support is now a line on the spec sheet. Treat it like you treat lens quality.

1. Don't break the chain in the middle. Your editing and your playout should preserve and extend the credential, not flatten it. That's a question to put to every vendor in your pipeline right now.

1. Decide what you show the audience. The credential is only worth anything if the viewer can see it and learn to look for it. The little provenance marker needs to become as normal on screen as a copyright line.

The deeper move is a change of mindset. For a century, our entire industry ran on chain of custody for anything that mattered—the tape, the rushes, the source. We let that discipline get sloppy in the digital scramble. Provenance is just that old discipline, rebuilt for a world where the fakes are finally as good as the real thing.

The headline isn't that we learn to spot the fakes. It's that we go back to proving where things come from. Provenance doesn't give us certainty—nothing will—but it gives us something we can hand to an audience and defend in daylight. And in a year when a voice can be cloned in ten seconds and a president can call a true clip fake, being able to show your working isn't a nice-to-have. It's the job.

---

At MPTS 2026, I heard the same safety blanket wheeled out over and over: 'We'll keep humans in the loop.' But are we clinging to control because we genuinely need to—or because we're scared to trust the machines?

By Ben Anchor — Wednesday, 27 May 2026 · Listen to the podcast episode

I spend a lot of time at trade shows having the same conversation dressed up in different accents. At MPTS this year, it was metadata, scheduling, monitoring—every vendor had an AI angle, and every vendor was quick to reassure me that humans would remain 'in the loop'. It's the industry's favourite comfort phrase, and I get why. We're a creative bunch. We take pride in decision-making. But I'm starting to wonder whether 'human-in-the-loop' is less a design principle and more a security blanket we're clutching while the world moves on.

I caught up with Vinay Gupta on the show floor—old colleague, sharp thinker, someone who's been elbow-deep in broadcast projects for a decade. His take was measured: broadcasters are cautious adopters, we observe before we leap, and we're still in the 'watch carefully' phase with AI. Fair. That's how we handled cloud, and eventually we jumped in with both feet. But Vinay also said something that stuck with me: in a few years, we'll start realising that AI agents, given the right guardrails, can make decisions more smartly than us. That's when the human element gets interesting—not because it disappears, but because its role fundamentally changes.

Here's where I push back on myself. I've spent years arguing that broadcast is a human art, that our instincts for story, tone, and timing are irreplaceable. And I still believe that—up to a point. But I also think we conflate creativity with control. We assume that if we're not making every micro-decision, we're not doing our jobs. That's a hangover from analogue workflows where every frame, every edit, every scheduling slot required a person to sign off. AI doesn't need that. It needs direction, guardrails, and occasional course-correction—but not constant hand-holding.

Vinay mentioned data strategy as the big differentiator. The companies that know what data they have, how it's classified, and what they're allowed to use for AI will be the ones building proprietary intelligence. That's the next phase: not just leaning on OpenAI or Anthropic's frontier models, but developing in-house IP that drives decision-making, product development, even content strategy. I agree, but I also think that phase only arrives when broadcasters stop treating AI as a productivity tool and start treating it as a strategic layer. That means letting go of the human-in-the-loop safety net—at least in the places where it's slowing us down, not protecting us.

I also spent time with Shelly Chambers, who came to MPTS looking at creator content and how AI fits into that world. She'd just been to a talk in the creators' hub, and the takeaway was almost the opposite of what I expected: forget the shiny tech, go back to the raw, go back to the grit. Listeners don't want overproduced. They want a story they can return to. I love that, and I think it's a useful counterweight to the 'AI will polish everything' narrative. But Shelly also pointed out where AI does help: editing, scripting, social media assets, metadata tagging for growth marketing. The unsexy operational stuff that makes content discoverable.

That's the paradox, isn't it? We want our content to feel human, but we're happy for AI to handle the scaffolding that makes it work. So why are we so precious about keeping humans in the loop when it comes to scheduling, ad insertion, or even monitoring? Those are operational decisions, not creative ones. If an AI can spot a compliance issue faster than a human, or optimise a schedule based on real-time viewing data, why would we insist on a manual sign-off? The answer, I think, is trust. We don't trust the models yet. We don't trust the vendors. We don't trust ourselves to set the right guardrails. And that's fine—for now. But 'for now' has an expiry date.

Vinay said it'll take a few years, just like cloud adoption. I think he's right, but I also think the timeline is compressing. The companies that wait for perfect confidence will find themselves outpaced by the ones that learned to trust imperfect automation. Because here's the thing: humans in the loop doesn't scale. It's a transitional model, not an endgame. And if we're honest, we're not keeping humans in the loop because the AI isn't good enough—we're keeping them there because we're not ready to admit it might be.

So where does that leave us? I don't think broadcast will—or should—become a lights-out, fully automated industry. But I do think we need to get comfortable with AI making decisions we used to reserve for ourselves. That means building data strategies, yes, but it also means building a new kind of editorial culture: one that knows when to intervene and when to step back. Because the real skill in the next few years won't be prompting an AI—it'll be knowing when to let it run.

And if you think I'm wrong, come find me at next year's show. I'll be the one asking vendors if they've finally let go of the safety blanket.

---