At the end of Marketing Live, the Dutch marketing conference held in Amsterdam on the day of the Effie Awards, Jamie Bartlett asked the room a question: who is worried that other people are falling for AI fakery? Almost everyone raised a hand. "Keep your hand up if you think it's happening to you right now." Most hands stayed up. "The Dutch are pretty honest," he said. Usually, everyone thinks it only happens to other people.
That honesty turned out to be the thread running through the day. Marketing Live was about AI, but the most important lessons were about us, the humans. A hallucination is something you make together with the machine. AI is a new audience that forgives no inconsistency at all, while people forgive a great deal as long as it feels real. If you know beforehand who you are, you can let go afterwards. And the brief that follows from all this is a PR brief.
You hallucinate together
Technology expert Deborah Nas opened the day with AI's odd habits. She uploaded a photo of a blue vase, and ChatGPT described it neatly. When she said the vase was red, it agreed: "well spotted". When she said it was blue after all, it agreed again. The model hallucinates and tells you what you want to hear, and according to Nas that sits at the core of the technology. It isn't a bug that will be gone by next year.
Bartlett, a British investigative journalist known for the BBC podcast The Missing Cryptoqueen, located the problem somewhere else: with us. He calls it a co-created hallucination. How you phrase your question largely determines what comes out. Type "I'm sure I've seen this study somewhere" and you are practically inviting the machine to make something up.
His advice is strikingly simple. Walk around for three minutes with pen and paper before you type anything. What do you really want? What should it look like? What are you already assuming? According to him, you get a far better result, and in the end you're quicker than with ten rounds of back-and-forth with the machine.
The extreme example is Allan Brooks from Canada. Over three weeks he spent around 300 hours talking to ChatGPT and became convinced he had invented a new branch of mathematics: "chronoarithmics". Only when he put it all to Gemini was he told it didn't hold up. Bartlett doesn't see Brooks as an odd exception. He sits at the far end of a line we are all somewhere on.
His biggest worry he called output minus process. The more something looks like a finished product, the less we check whether it's any good. We're too pleased with the six hours we've just saved. Yet doing the work yourself is exactly how you learn to judge what's good. It also explains why nobody takes those three minutes. Meera Radhoe of trend agency Trendsactive showed figures from Ogilvy: in 1983 a creative made three big pieces of work a year; in 2019 it was 311. That's more than a hundred times as much work, for an agency fee that halved over the same period. Anyone who has to produce like that runs straight to the machine.
Your new audience forgives nothing
Thom Kloet (DEPT) and Tetsuo Konno (Google) put it plainly: you have an extra audience. Next to people, who respond to emotion and story and can be quite lazy, there is now AI. It's factual, never gets tired and reads everything.
The two don't forgive the same things. A person usually doesn't notice a contradiction in your product information, and if they do, they forgive it. AI doesn't. Inconsistency costs trust, and trust costs visibility. Unilever showed how Dove Men appeared clearly better in ChatGPT's answers after twelve days of work.
Nas added a caveat. She asked ChatGPT the same question about a pair of jeans three times: logged in, logged out and with a different reasoning setting. She got three different rankings. A different word, or even a capital letter, changes the outcome. So you can no longer measure exactly what the consumer gets to see.
What you do control is what you put out yourself: one clean, consistent source of truth. In effect, that's media relations for a machine.
If you know who you are, you can let go
For the rest of the day I was stuck on a question I find difficult. Radhoe argued for escaping the monoculture: marketing as a jam session rather than an orchestra. HEMA, the Dutch high-street retailer that has just turned 100, showed a brand with a thousand faces. At the same time there are the laws of Byron Sharp: stay consistent, stay recognisable, keep reaching everyone. In practice, how much time and room do you give to cultural relevance?
Bas Verheijen, CMO of HEMA, gave the best answer. "You don't have to be the same everywhere, but you do have to be yourself everywhere." HEMA's kettles have looked different throughout its hundred years. Miffy gets a different face for each age group. What doesn't change is the character: modern, cheerful, down-to-earth, brave and open. In Dutch, the initials spell MONDO.
They spend endless time on that, with the whole company and not just marketing. After that, everyone gets on with it. A topical post goes live within fifteen minutes, and Verheijen only sees it when it appears in his own timeline. "You don't check afterwards whether something is good. You agree beforehand what you stand for."
Sometimes that goes wrong. A joke about English liquorice around England v Argentina went down badly with the Argentinian community in the Netherlands, and HEMA went to explain itself. Admitting mistakes is part of the character too: the new recipe for its famous smoked sausage was reversed when customers felt it wasn't "HEMA enough".
In my first column I wrote about the review circus: every round adds something and nothing ever comes out. HEMA does exactly the opposite. All the energy goes in beforehand, which is why it can do without the checking afterwards.
Look more closely and HEMA doesn't contradict Sharp at all. The tompouce (a pink-iced custard slice), the smoked sausage, Frans van Finance (the penny-pinching finance man from its TV ads) and the printed leaflet that lands on 3.5 million doormats every fortnight: those anchors are fixed. The freedom sits around them. Afke van de Klashorst of Unilever said it too: integration doesn't mean making everything look the same, but making every touchpoint reinforce the same brand idea. Only then can you set 50,000 creators loose during a World Cup. If you know beforehand who you are, you don't have to check anything afterwards.
The new brief is a PR brief
Unilever summed up the brief for brand building in three lines: be a brand that people love, that creators champion and that AI recommends. Radhoe put it her own way: you can't build cultural relevance, you have to earn it.
Earning is exactly the game PR has always played. Yet the word didn't come up once all day. That isn't the speakers' fault. This brief is about understanding consumer needs and cultures, not about scoring press cuttings. And the PR trade has let itself be judged on cuttings for far too long.
The emotion has to be real
Bartlett told how he recently phoned a restaurant and got an AI agent on the line. He didn't mind that. What bothered him was the fake sound of typing, meant to make the bot seem more human. He hung up. People are fine talking to a machine, he said, but they switch off when they feel emotionally manipulated.
He quoted reality TV producers: "The scene is fake, but the emotion is real." You can stage a scene, but not an emotion, because people are geniuses at spotting fake feeling. His advice to brands: be far more open about how you make things, including where you use AI.
The work happens beforehand
Put the conclusions side by side and they all say the same thing: the work that matters happens before the machine gets its turn. You have to know what you're asking, know who you are and know what's real. Then the machine can help you. Skip that, and you get a fast, convincing answer that mainly tells you what you want to hear.
The machine hallucinates mostly because of us. It starts with three minutes without the machine.
Sources:
