When AI Makes the Ad, Does It Need a Label?

A few days ago, a friend who runs an e-commerce shop complained to me.

He'd used an AI tool to retouch a product photo — just brightened the sky in the background a little, and tweaked the colors of the product itself.

His compliance team asked him: does this image need an "AI generated" label?

He froze.

I froze for a second too. Then it clicked.

What's really giving everyone a headache about AI in advertising isn't the technology — it's figuring out where the line is.

Today, let's really unpack this.

First, an overlooked fact

When most people hear "AI making ads," their first reaction is: this is something new, the law hasn't caught up yet, right?

Wrong.

The law caught up a long time ago. And it caught up hard.

Over in the EU, there's been the Unfair Commercial Practices Directive since 2005, and it states plainly: any commercial practice that "materially distorts the economic decision of an average consumer" is prohibited. What does that mean? It means if your ad misleads a consumer into making a purchase decision they otherwise wouldn't have made — whether you painted it by hand or generated it with AI, it's a violation.

In 2006, there was also the Misleading and Comparative Advertising Directive, defined even more directly: any advertisement that influences consumers' economic behavior in a deceptive way is illegal.

The UK, India, Brazil, South Africa — they all have similar statutes.

In other words, this underlying rule of "don't deceive people" was put in place decades ago, and it has nothing to do with whether you use AI.

So what did AI actually add?

It added another layer, called "transparency."

What does "transparency" actually mean?

In 2024, the EU passed a pretty important law called the AI Act. Article 50 of that law is devoted to exactly one thing.

It splits the requirement into two parts.

The first part (Article 50(2)) is for system vendors. If you build an AI system that can generate synthetic content — whether text, image, audio, or video — you have to ensure that what the system spits out is, at the machine level, "recognizable as AI-generated." Simply put, there has to be a machine-readable watermark embedded at the file level.

The second part (Article 50(4)) is for people who use AI, and it targets "deepfakes." If you use AI to generate or alter content that looks like a real person, a real event, or a real place — and an ordinary person looking at it would assume it's real — then you have to come forward and disclose that it's AI-made.

Sounds reasonable, right?

But the devil is in the details.

The phrase "standard editorial editing" is tripping up everyone

There's an exception inside Article 50(2).

It says: if the AI is merely "assisting with standard editorial editing," or "does not substantially alter the original data or its meaning," then it doesn't count.

Sounds simple enough.

But what counts as "standard editorial editing"?

Think about it.

Does brightening the sky in a product photo count? Does removing background noise from a video count? Does auto-generating subtitles with AI count? Does adjusting brightness, contrast, or saturation count?

By all rights, all of these should count. Because they're exactly what people have been doing for years with Photoshop, with darkroom chemicals, with hand retouching.

Even earlier, in the era before AI was even born, the people who shot food ads had an open secret.

They used motor oil instead of maple syrup on pancakes, because real syrup would melt under the hot lights. They used glue instead of milk on cereal, because real milk was too thin to cling. They sprayed glycerin, not real condensation, on cold-drink bottles to get those water droplets.

These "tricks" have been used in the advertising industry for decades, and nobody ever demanded they slap an "artificially synthesized" label on them.

So why, when someone uses AI to nudge the brightness, should they suddenly have to label it?

The law has left a giant gap for interpretation here. "Standard editorial editing" is a fuzzy phrase, and it's not clear who gets to draw the boundary. That's landmine number one.

Landmine number two: defining "deepfake"

Article 50(4) requires public disclosure for "deepfakes."

So what counts as a deepfake?

The legal definition is: images, audio, or video that AI has generated or altered, which appear to depict a real person, object, place, entity, or event, and which would lead someone to believe they are real.

Notice that there are two conditions stacked here.

First, it looks real. If you generate an image of a completely fictional alien walking on the rings of Saturn, nobody's going to take it for real — so it's not a deepfake.

Second, it would mislead someone into thinking it's real. If you add some dramatic storm clouds to an ad for atmosphere, no consumer is going to believe there's literally a sky that exaggerated out there, and they're not going to change their purchase decision because of it. So that shouldn't count either.

What really needs regulating is the case where you produce a video that looks like a real person and a real event, and that video leads consumers to form a false belief that changes their purchase decision.

That's the actual intent of the legislation.

But the literal wording of the law is so broad that, if applied mechanically, almost any AI-generated commercial image could be swept in. That's landmine number two.

The good news: the law itself left a window open

Hidden inside Article 50(4) is one crucial sentence.

It says: if the content is evidently of an artistic, creative, satirical, fictional, or similar nature, then the transparency obligation is downgraded — you only need to disclose it "in a manner that does not harm the presentation and enjoyment of the work."

That sentence carries a lot of weight.

Why?

Because advertising is, by its nature, a deeply creative industry.

You shoot a funny commercial and use AI to push an actor's facial expression further into caricature. You design a dramatically charged poster and use AI to generate a fictional cityscape backdrop. You run a satirical current-events marketing campaign and use AI to synthesize a conversation that obviously couldn't have happened.

All of these should land in the "creative" category.

The European Court of Human Rights long ago recognized that when commercial advertising carries creative, satirical, or artistic elements, it's equally protected by freedom of expression. The US Supreme Court has also ruled that commercial advertising enjoys First Amendment protection — though under a somewhat more lenient standard of review than political speech.

In other words, advertising is not an orphan under the law. It has its own freedom of expression.

So what does "disclosure in a manner that does not harm the presentation" actually mean?

It means you can't require an advertiser to stamp a giant red "AI GENERATED" seal in the dead center of a poster they've painstakingly designed. That would destroy the entire piece.

So how should disclosure happen?

It could be a secondary layer of information you only see when you click through. It could be machine-readable metadata. It could be a small icon tucked into a corner that doesn't compete visually. It could be a "content credentials" standard like C2PA — pushed by an industry alliance whose participants include Google, OpenAI, Microsoft, the BBC, and Sony, and which uses digital watermarks and cryptographic signatures to record a piece of content's origin and modification history.

A flexible approach is the reasonable approach.

Label fatigue: the overlooked big risk

Now let me tell you something counterintuitive.

The more labels you slap on, the less people believe them.

Sounds wrong, doesn't it? More labels should mean more transparency, which should mean more trust.

But that's not how humans work.

Think about it. You scroll through your phone for a day, you see a hundred ads, and every single one is tagged "AI enhanced," "AI-assisted creation," "may contain AI-generated content." Are you really going to scrutinize every single one more carefully?

No.

You're going to tune out. Just like when you walk past a hundred "wet floor" signs — by the tenth one, you've stopped looking.

That's what's called "label fatigue."

What's worse is that when a genuinely dangerous deepfake shows up — say, a video faking a public figure endorsing a fraudulent wealth-management product — its label looks identical to the label on "I brightened the sky a little." Consumers can't tell which one to take seriously and which one they can ignore.

The value of a label lies in its scarcity.

Labeling all AI is the same as labeling none of it.

The original intent of the legislation was to build trust. But the outcome of mechanical enforcement could be the opposite — it hollows out trust.

SMEs are the most likely to get crushed

There's another overlooked angle.

The EU itself says it has more than 6,800 AI startups on the continent, and its ambition is to be the global leader in AI. A report on EU competitiveness specifically warned in 2024: stop piling compliance costs onto small and medium-sized enterprises — especially in the digital domain. Doing so is self-defeating.

Why?

Because embedding a machine-readable watermark in an AI-generated ad image, doing content provenance, and running a compliance review — those are costs for a big company, but for a small shop owner they may simply be undoable.

A coffee-shop owner who uses AI to retouch a photo of a new cake for a social media post — if you demand they label it, trace its origin, and retain metadata, they'll just stop using AI. Or worse, stop running ads altogether.

Rules are a good thing. But when rules suffocate innovation, the rules themselves lose their meaning.

So what should we actually do?

Let me boil it down to five principles.

Principle one: technology neutrality.

What the law should be policing is "did anyone get deceived," not "which tool was used." Deception is deception — whether you used a paintbrush, Photoshop, or AI.

Principle two: risk-based proportionality.

Disclosure obligations should be tied to the actual risk of consumers being deceived. For low-risk day-to-day editing, no mandatory disclosure; for high-risk photorealistic synthetic content, disclosure is required.

Principle three: protection of fundamental rights.

The creative, satirical, and artistic expression inside advertising is itself part of freedom of speech. No disclosure requirement should be allowed to choke off that expression.

Principle four: legal certainty.

"Standard editorial editing," "deepfake," "appropriate manner" — the current definitions of these terms are far too vague. Regulators need to issue tighter guidance so practitioners know where the lines are, instead of everyone being afraid of their own shadow.

Principle five: protecting innovation and competitiveness.

Compliance solutions have to be ones that SMEs can actually carry out. A one-size-fits-all hard mandate will only make the big companies more monopolistic.

Back to that photo from the beginning

So, returning to my friend's dilemma.

He used AI to brighten the sky in a product photo — does it need an "AI generated" label?

By my reading of the legislation's intent, no.

Because this image doesn't look like a real person or a real event, and it wouldn't cause a consumer to change their purchase decision over a tweak in brightness. It's a standard image edit — just done with a new tool.

But by the literal wording of the statute, the answer is unclear.

And that's the biggest problem of all.

Technology is moving fast. The law is trying to keep up, but its wording is too broad. The industry wants to comply, but it can't find a clear boundary. AI has brought unprecedented productivity to the advertising industry — and, in particular, it has given SMEs their first real chance to compete with the big brands on creativity.

Don't let a vague label slam that freshly opened door shut again.

Maybe one day we'll find a balance that blocks genuine deception without strangling ordinary creativity. Until then, I'd rather have fewer labels — but every one of them counts.