Jonathan Mall

Your Next Customer Does Not Read Labels, and Your Trust Strategy Assumes They Do

Hand-drawn illustration: a wrapped product with a large blank amber tag, a human hand and a small robot hand both holding the tag string
Key takeaway: The label debate has a short shelf life, because your customer is starting to delegate the shopping to an agent that never reads labels. Sina Frizzi-Harms, who leads marketing at a Swiss online grocery retailer, calls the real task creating a login effect so the agent buys through your platform. That reframes everything else: for as long as humans still look, disclosure has measurable costs (the "AI-generated" tag is penalized, "human-made" earns nothing, and the hybrid label performs worst of all), and since August 2, 2026 the EU AI Act requires far less visible labeling than your feed claims. Machine-readable provenance is the durable move; visible badges are the transitional one.

Every trust strategy in marketing rests on one unexamined assumption: that a human being is on the other end, looking. Labels, badges, disclosures, brand tone, the careful wording of an origin claim, all of it is built for a pair of eyes and the judgment behind them. That assumption is beginning to expire, and the people who will feel it first are already rearranging their teams around it.

The timing makes the contrast almost comic. Three days before this article, on August 2, 2026, Article 50 of the EU AI Act became applicable, and within hours my feed filled with confident posts announcing that all AI-generated marketing content must now carry a label. That claim is false, and I take it apart below with the actual article numbers. But the deeper reason it matters less than it looks came up in a conversation with a marketing leader who is already planning for the world after the label.

The question behind the label question: what if nobody is looking?

The most consequential thing happening to AI in marketing is that the reader may stop being a person. Sina Frizzi-Harms, who leads marketing at a Swiss online grocery retailer, does not frame her problem as "will customers notice AI in our ads." She frames it as what happens when the customer stops doing the shopping at all. "You still have to create that login effect," she told me, "so that the agent then shops via your platform." Agentic purchasing already works in her category, in her words "it's not great yet," and her product teams have tested how it behaves. Her team is not theorizing about this; they are asking what their answer is when the scenario arrives.

Sit with what that does to disclosure. Every experiment in the acceptance literature measures a human looking at a label and flinching. If an agent assembles the basket, the flinch never happens, and the only question left is what the agent can verify: structured product data, provenance metadata, signed content credentials, return rates, delivery reliability. The visible badge, the thing most of the industry is currently arguing about, becomes invisible to the buyer that matters. The machine-readable marking, the thing almost nobody is implementing with any care, becomes the entire game.

Hand-drawn illustration: an amber robot arm pushing a shopping trolley past shelves of blank products while a person sits apart looking at a phone

She also names the limit of that substitution, and it is the most useful brand argument I have heard on the topic. Groceries are not televisions. For a TV, in her words, providers are "more or less interchangeable if they're reputable," so an agent optimizing price and delivery can pick almost at random without loss. For food, freshness perception and quality are brand-specific and still decide the choice, which means perception carries real information that an agent has to source from somewhere. Categories where the brand encodes something an agent cannot compute keep a moat. Commodity categories lose it first, and they lose it to whoever feeds the agent the cleanest data.

Her organizational reality tells you how fast this is moving. Marketing at her company now sits inside the Product department, reporting to the Chief Product Officer, closer to product areas and engineering teams; her team has been fully on Claude for about three months and uses it at every level, from analysis to presentations to text optimization. And in the same breath she reports the honest friction: two colleagues told her unprompted that they see no added value in it for their own work. That is the real adoption curve, both halves of it, from someone who has no reason to sell either.

Where we are: the buyer is beginning to delegate, which makes verifiable data more durable than any visible badge, while brand perception still protects categories where it carries information machines cannot compute. That is the horizon. The rest of this article is about the years before we get there, when humans are still the ones looking, and where the evidence about what they do when they see a label is stranger than most marketers assume.

What customers actually think of AI content, while they are still the ones looking

Customers accept AI where they expect competence and resist it where they expect feeling, and labels have an asymmetric price: an "AI-generated" tag costs perceived effort while a "human-made" tag earns nothing extra. The science behind that summary rests on a two-decade seesaw. People abandon algorithms faster than humans after seeing identical mistakes (Dietvorst, Simmons and Massey, 2015), yet in other settings weight algorithmic advice more heavily than human advice (Logg, Minson and Moore, 2019). The resolution is context. The cleanest map is the word-of-machine effect (Longoni and Cian, nine studies): people prefer AI recommenders for functional, utilitarian attributes and human recommenders for emotional, hedonic ones. Machines are believed competent; humans are believed to understand feeling.

Labels add their own physics. In a 2026 experiment with 618 short-video users, marking identical content "AI-generated" cut its perceived effort substantially, while the "human-made" label earned no premium over no label at all (Frontiers in Psychology). Read that asymmetry twice: disclosure punishes the machine without rewarding the human. And before you conclude disclosure is always a tax, one study found the opposite in hedonic e-commerce, where disclosing AI curation lifted purchase intention (JTAER), and another found the disclosure penalty concentrates in consumers who already distrusted AI (it is a moderator, not a constant). The honest summary: effects are real, heterogeneous, and smaller than the discourse implies.

And one caveat from practice belongs next to all of these experiments, because it cuts against overreading them. Label studies force the origin question: they hold "AI-generated" in front of a participant and then measure the wince. In the wild, almost nobody asks that question. When an ad is good, when the content genuinely entertains, people watch it, share it and buy from it without ever running an origin check; the evaluation heuristic the lab installs is rarely the one consumers apply at the feed. I say this as field observation, not as a study, but it fits the data: the penalty shows up when a label is salient, platform labels get scrolled past, and origin gets interrogated mainly at moments of doubt, a questioned claim, a scandal, a testimonial. Quality and entertainment carry the everyday case. Which is exactly why the next two sections matter: the law decides when the origin question gets forced into view, and your craft decides whether the work survives it.

Hand-drawn illustration: three identical paintings in a gallery, viewers crowding the one with a human plaque, the hybrid-labeled one alone

Where we are: customers accept AI where they expect competence and resist it where they expect feeling, and the AI label costs perceived effort while the human label buys nothing. Every marketer's instinct at this point is the compromise: admit the AI, emphasize the human, label it a collaboration. The data on that instinct is the strangest result I know in this field.

The hybrid penalty

Labeling identical work as a human-AI collaboration gets it rated worse than labeling it purely human or purely AI. In a controlled study I wrote about in detail in German, 300 participants rated identical work labeled three ways. Human-labeled work scored 5.8 out of 7. AI-labeled work scored 3.9. The hybrid label, "created by AI together with a human," scored 3.4, worst of all three, with the penalty landing on moral legitimacy rather than perceived quality, and a large effect size. The mechanism is old: we price work by the effort we imagine behind it (the effort heuristic, Kruger et al., 2004), and a hybrid label tells the reader neither story cleanly, so they imagine the worst of both.

Full transparency, in the spirit of this series: I have not found a peer-reviewed study that runs exactly this three-label test on trust or purchase outcomes, so treat the hybrid penalty as a replicated-in-my-lab finding with directional support rather than settled literature. The support is real though. Longoni and Cian found hybrid framing neutralizes the word-of-machine effect in both directions, and a meta-analysis of 106 experiments found human-AI combinations typically underperform the better of human alone or AI alone on task outcomes (Nature Human Behaviour, 2024). Combination rarely captures the best of both. In perception, it can capture the worst.

The wrong lesson is the one most commentators draw: force a choice, go pure human or pure AI. The right lesson is that origin labels measure a bias, and the fix for a bias is rarely to reorganize your production around it. Which raises the question the label-averse have been avoiding since Saturday: when do you have no choice?

Where we are: labels move perception, the hybrid label moves it most cruelly, and honesty about origin is currently taxed by your audience's own heuristics. So the temptation is to label nothing. As of this week, in Europe, that is sometimes illegal, and the boundary between "sometimes" and "mostly not" is exactly what almost everyone is getting wrong.

AI labeling since August 2026: what the law actually requires, and what it does not

Most AI-assisted, human-reviewed marketing needs no visible label; the law requires machine-readable marking from tool providers by default, and visible disclosure only for chatbots, deepfakes, and unreviewed AI text on matters of public interest. Specifically, Article 50 creates four separate duties, each with its own trigger, applicable since August 2, 2026, with fines up to €15 million or 3 percent of worldwide turnover (Article 99):

One: chatbots must be recognizable. If people interact with your AI directly, they must be able to tell, unless it is obvious to a reasonably informed person. Name the bot a bot. Never let it play a human agent.

Two: synthetic media must be machine-readable. Providers of generative systems must mark AI-generated audio, images, video and text in a machine-readable, detectable way, think embedded metadata such as C2PA Content Credentials, not a visible banner. This duty sits mostly with tool providers, and systems that only assist standard editing, or do not substantially change the input, are exempt. Systems already on the market get a grace period to December 2, 2026.

Three: emotion recognition and biometric categorization require informing the people exposed. Niche for most marketers, decisive for some retail-analytics ambitions.

Four: deepfakes and AI text on matters of public interest need visible disclosure. A photoreal synthetic person, or fabricated content that would pass as authentic footage, must be labeled. AI-written text "to inform the public on matters of public interest" must be disclosed, unless a human has editorial review and responsibility. That human-review exemption is the sentence most LinkedIn coverage skips, and it changes everything for marketing: a typical AI-drafted, human-edited blog post, product page or campaign email requires no visible AI label under the Act. The Commission's July 2026 guidelines and the voluntary Code of Practice (roughly 190 signatories) fill in the details; signing the Code creates a presumption of conformity for the marking duty.

Around the EU core, the rest of the map: China has required visible labels on synthetic content since September 2025 (stricter than the EU). The US has no federal labeling law; the FTC polices deception, and its fake-review rule bans AI-fabricated reviews and testimonials outright. California's transparency act took effect the same day as Article 50, deliberately. In Germany, the UWG runs in parallel: undisclosed AI use that deceives can draw an Abmahnung from a competitor regardless of AI Act compliance (German guidance). And the platforms are de facto law with faster courts: YouTube suspends channels over undisclosed realistic synthetic content, and Meta auto-detects AI in ad creative since June 2026 and requires explicit written disclosure for AI in political ads (Meta policy). Meta's detection has also mislabeled real photographs as AI (its own labeling saga forced a rename from "Made with AI" to "AI info"), which is worth remembering: the labeling infrastructure judges you probabilistically, whether you disclosed or not.

Hand-drawn illustration: a large paragraph section sign woven from amber circuit-like lines, ringed by small stars

Zombie claims, briefly. "All AI content must now be visibly labeled in the EU": false, the human-editorial-review and assistive-editing exemptions cover most professional marketing workflows. "Signing the Code of Practice is mandatory": it is voluntary. "The AI Act requires labeling synthetic research data": it does not; but republishing synthetic voices as real customer testimonials is plain old deception law, FTC and UWG, and was illegal before AI existed. "Virtual influencers get three times the engagement of humans": vendor-aggregator numbers with no disclosed methodology, and the best-known virtual influencer monetizes at a fraction of comparable humans. As always: if the argument needs inflated claims, it is not an argument.
Where we are: the law asks for less visible labeling than feared, the platforms ask for more than the law, and your audience penalizes the labels either way. That triangle looks unresolvable until you stop treating disclosure as a compliance cost and start treating it as a design material. The playbook that falls out of the evidence is short, and the first rule is the one most companies have backwards.

The playbook the evidence supports

The whole playbook fits in one sentence: mark everything machine-readably, label visibly only where the law triggers it, never counterfeit humanness, put disclosure where it helps, and spend the savings on understanding your audience. Now each rule with its evidence.

Mark everything machine-readably, label visibly only where triggered. Embedded provenance (C2PA) satisfies the provider duty, feeds the platform detectors on your terms, and costs no perception penalty because customers never see it. Reserve visible labels for the actual triggers: photoreal synthetic people, content that could pass as documentary footage, public-interest text without editorial review, and any chatbot.

Never counterfeit humanness. The chatbot rule and the fake-review rule share one principle, and it matches the acceptance data: deception about origin is where trust dies and where every regulator converges. This includes research: I build digital-twin simulations of audiences, and their outputs belong in strategy decks, not dressed up as real customer quotes. Synthetic insight is legitimate; synthetic testimony is fraud.

Put the disclosure where it helps. The evidence says disclosure of AI curation can lift hedonic purchases, that giving users control over an algorithm restores trust after errors (Dietvorst et al., 2018), and that the penalty concentrates among the already-averse. Disclose in functional contexts confidently, in emotional contexts carefully, and in hybrid workflows describe the process ("drafted with AI, decided by us") rather than hanging an ambiguous co-authorship label on the artifact.

Spend the savings on understanding, not volume. Generation is commoditized; the AI-written share of the web plateaued at parity partly because undifferentiated content stopped working. The durable advantage is knowing what your audience will feel before you publish, and word-level understanding moves real numbers: in my own client work, rewording a single search prompt lifted inquiries by roughly 30 percent, and one word swap moved social engagement by 28 percent (documented here). AI's job in marketing is to make you understand people faster. The decision of what to say, and the responsibility for saying it, stays with you, which is the same conclusion the relationship and society evidence forced, and by now I think it is the conclusion.

Where we are: mark invisibly, label where triggered, never fake a human, disclose where it helps, and pour the dividend into understanding your audience. One question remains, and it decides whether any of this playbook survives contact with the next five years: which way is acceptance itself moving?

Where acceptance goes next

Acceptance will drift toward indifference about origin as provenance becomes ambient infrastructure, and the agentic shift Sina Frizzi-Harms is preparing for will arrive faster in commodity categories than in ones where brand perception carries information a machine cannot compute. Adoption is no longer the variable: AI now powers roughly a quarter of marketing activities in the most credible tracking survey (Duke CMO Survey), and the fights ahead are about provenance infrastructure and audience trust, not tooling. Camera makers now sign images at capture, platforms read the signatures, and a generation raised on AI companions will bring very different baselines for what "authentic" even means; the label experiments above measure today's heuristics, not eternal law. My bet, stated so it can be wrong: origin labels will fade in importance as provenance becomes ambient, and the brands that win will be the ones that used the transition years to become verifiably honest, because in a world where anything can be generated, being the account that never needed the benefit of the doubt is the scarcest asset in marketing.

The label question, in the end, is the small version of the question all three articles in this series keep landing on. Machines to understand humans, humans to answer for the decisions. Everything else is implementation detail.

Sources