A Watermark Won't Save Your Credibility

Here is the story everyone is passing around right now. Google opened its SynthID detector to the public, so now anyone can check an image, a video, or an audio clip and see whether it carries a watermark from Google or a partner AI tool. The reaction has been predictable. Finally, people say, a way to separate the real from the fake. A truth machine for the age of synthetic everything.

I want to slow that down.

Because the belief underneath the celebration is the part worth testing. The belief is that detection equals trust. That if we can just label the AI stuff, audiences will know what to believe, and brands that make real things will be rewarded for it. That is a comforting story. It is also mostly wrong.

What SynthID actually does, and what it does not

Let us be fair to the tool first, because it is genuinely clever. SynthID embeds a signal into content created by certain AI systems. Not a visible stamp slapped in the corner, but something woven into the pixels or the waveform in a way that survives a fair amount of editing. Opening the detector to the public means a journalist, a buyer, or a curious person can now paste in a file and get a read on whether it came from a participating model.

That is useful. I mean it. For a reporter verifying a viral clip, or a platform trying to flag synthetic political ads, this is real infrastructure. I am not here to sneer at it.

But look closely at the sentence I just wrote. A participating model.

SynthID only catches what SynthID marks. Google's own tools, and the partners who have agreed to play along. The open source image generator somebody ran on a gaming laptop? Not marked. The screenshotted output stripped of its data? Weaker signal, or none. The tool built specifically to launder AI content so it reads as human? That entire category exists precisely to beat detectors like this one.

A detector that only catches the honest actors is not a lie detector. It is a politeness detector.

So the headline should really read something like, Google can now confirm that content made with Google tools was made with Google tools. That is a smaller claim. Still worth having. Just a lot less like a truth machine than the excitement suggests.

The belief worth challenging

Here is the conventional wisdom I keep hearing from marketers, and I think it deserves a hard look. The story goes like this. AI flooded the internet with cheap content. Trust collapsed. Now detection tools will restore order, and the brands that create authentic, human work will be verified and rewarded while the slop gets filtered out.

Nice narrative. The problem is that trust has never worked that way.

Trust is not a property you can scan for. It is not baked into a file. When someone believes you, they are not responding to a watermark. They are responding to a track record, a tone, a pattern of being right, a feeling that you have skin in the game. A detector can tell you how a thing was made. It cannot tell you whether the thing is true, useful, or worth your time.

Think about it from your own buyer's seat. If I read a case study from a company and it changes how I think, do I care whether a model helped draft a sentence in it? Not really. What I care about is whether the claims hold up, whether the numbers are real, whether the people behind it seem to actually know the thing they are talking about. A clean watermark report does nothing for any of that.

Provenance is not the same as credibility. Knowing where something came from tells you nothing about whether it is any good.

We have been here before

Every time a new way to fake things arrives, we reach for the same reflex. Build a detector. Catch the fakes. Restore the old equilibrium.

Photoshop was going to end our ability to believe any image. Then plagiarism checkers were going to clean up student work and content farms. Then the last couple of years brought a wave of AI detectors that schools and editors bolted onto everything, and we watched them flag real human writing as machine made and wave actual machine writing straight through. The false positives hurt the honest people. The determined cheaters adapted in a weekend.

I am not saying detection is pointless. I am saying it is an arms race, and in an arms race the tool is never the finish line. It is a move. The other side gets a move too.

So building your entire trust strategy on the hope that a detector will sort friend from foe is like building your house on the assumption that locks end burglary. Locks matter. Locks are not a security plan.

What this actually means for your marketing

Here is where it gets practical, because I do not want to leave you with a philosophy lecture.

The flood of synthetic content is real. Your buyers are swimming in it. And their defense mechanism is not going to be a watermark checker they run on every blog post. It is going to be something older and blunter. They are going to get skeptical of everything, and they are going to lean hard on the sources they already believe.

That is the shift that actually matters. Not detection. Distrust by default.

When everything could be fake, the brands that get believed are the ones who were already known.

So the question stops being, how do I prove my content is human. The question becomes, why would anyone believe me in the first place. And the answers to that are not technical. They are the unglamorous things we have always known and keep pretending we have outgrown.

A real name attached to the work. A point of view that costs something to hold. Specifics only an insider would know. Claims you can check. A consistent voice over years, not a brand that sounds like a different company every quarter. The willingness to say when you are wrong. These are the signals humans actually use to decide who to trust, and no model can fake the accumulated weight of them, because the weight is the whole point.

The AEO angle nobody is connecting

There is a second layer here that ties directly into how people find you now. More and more buyers are not scrolling ten blue links. They are asking an AI assistant a question and taking the answer. Getting recommended inside those answers, what some of us call answer engine optimization, is becoming its own discipline.

And watermarking collides with that in an interesting way.

As the web fills with synthetic content, the systems that generate answers have to decide what to treat as a reliable source. They are not going to lean on a watermark to make that call, because most of the junk will not carry one. They are going to lean on signals of authority. Who is cited elsewhere. Who has a real author with a real history. Whose claims line up with other trusted sources. Which brand keeps showing up as the one that got it right.

In other words, the same thing that earns trust from a human skeptic is the thing that earns a recommendation from a machine. That is not a coincidence. These systems are, clumsily, trying to approximate human judgment about credibility. So if you build for genuine authority, you are building for both at once.

Chasing detection is defense. Building authority is offense. One keeps you from looking fake. The other makes people want to quote you.

So should you ignore SynthID entirely?

No. Use it the way you would use any verification tool. If you are a publisher, checking suspicious submissions is smart hygiene. If you run paid media, knowing whether an asset carries a synthetic marker can matter for disclosure rules that are only going to get stricter. If you work in any field where a fabricated image or clip could do real damage, having one more way to sanity check a file is a good thing to have in the drawer.

Just hold it in proportion. It is a smoke detector, not a fireproof house. It does one narrow job, it does it only for cooperating sources, and it tells you nothing about the quality or honesty of what you are looking at. Treat it as a small piece of hygiene, not as the thing that saves your reputation.

And please do not fall for the inverse trap, the one I can already see coming. The idea that you should scrub every trace of AI from your own process so you come out clean on somebody's detector. That is optimizing for the wrong judge. Your buyer is the judge. Nobody ever bought from a company because its blog post passed a watermark scan. They bought because the work was good and the people seemed trustworthy.

Where the real line gets drawn

Let me land this plainly. The arrival of public detection tools is being sold as the moment trust gets restored to the internet. It is not. It is a modest, honest attempt to label some content made by some tools, and it will be routed around by exactly the people it is meant to stop.

The trust problem is not getting solved by a scanner. It is getting handed back to us, the people who make things and put our names on them.

That is actually good news, if you think about it. It means the moat is the same moat it has always been. Be a real source. Say true things. Show your work. Have a spine. Be the brand someone would vouch for without having to run a file through a detector first, because they have read you for two years and you have never wasted their time.

A watermark tells the world where your content was born. It says nothing about whether your content deserves to be believed. Those are two different problems, and only one of them has ever mattered.

Build for the one that matters.