Published June 12, 2023 · By CampaignsLive · Industry
Heinz’s “Draw Ketchup” campaign, launched in mid-2022, was probably the first piece of work that taught the wider advertising industry to take generative AI seriously as a brand creative tool. The mechanic was deceptively simple: Heinz’s agency at the time, Rethink, fed DALL·E 2 prompts like “ketchup, but Renaissance”; “ketchup, sketch”; “ketchup, in space.” The model returned variations on what the world had collectively decided ketchup looked like — and almost every one of them looked like a Heinz bottle.
The campaign reframed itself around that finding. The line ran roughly as: even an AI thinks of Heinz when it thinks of ketchup. It was a smart, low-risk, high-distinctiveness use of an early-generation tool that did one thing well — pull from the median visual concept attached to a word — and a great deal of other things badly. The campaign avoided the badly part by design.
For an industry that had spent eighteen months watching DALL·E and Midjourney produce uncanny portraits and surreal composites in the consumer space, “Draw Ketchup” was a useful proof of concept. It showed a real brand had figured out how to use the tool without producing the uncanny-portrait look that had defined the consumer side of the wave.
What Heinz proved that the rest of the industry was still figuring out
Three things, in retrospect, were the lessons.
The first was that the smart move with first-generation image AI was not to ask it to produce finished creative. It was to ask it to do something the audience knew was AI, frame the work around the AI being there, and let the discovery — that even the model defaults to Heinz — be the payoff. The creative thesis was about brand recognition; the AI was the instrument.
The second was that the work succeeded because it was constrained. Heinz did not ask the tool to generate a hero image for a thirty-second TV spot. It asked the tool to do a single, narrow, well-defined thing. The constraint protected the brand from the tool’s weaknesses.
The third — least discussed at the time, most important in retrospect — was about training data. The DALL·E result hinged on the model having ingested enough of the world’s images of “ketchup” to converge on the brand-typical version. That is a story about what the model had seen. It generalized to a broader question: if a model can hold the median visual concept of ketchup, what else can it hold? And what happens when you train one on a narrower, more disciplined visual corpus?
Cadbury’s parallel experiment
Around the same period, Cadbury’s India team ran a Diwali campaign that used voice synthesis rather than image generation: small business owners could record a hyperlocal ad in Shah Rukh Khan’s voice, with the actor’s consent, generated by Mondelez’s agency partner. The mechanic was unrelated to image AI, but the case lived in the same conversation because it was a brand using a generative model at scale, in market, with the celebrity-talent angle handled cleanly.
What both campaigns shared was that the AI was visible. The audience knew the work involved AI. The creative idea was about the AI being there, not despite it. This was a sensible posture for the moment — early-gen models were not good enough yet to disappear into a piece of brand creative without leaving fingerprints, and brands that tried to use them transparently fared better than brands that tried to use them invisibly.
The other end of the spectrum
The other thing the year proved, by contrast, was what did not work. The same period produced a wave of mid-tier brand campaigns that used AI to generate stock-style product photography, model imagery, or environmental shots, with the AI presence concealed. Most of these did not look right. The skin was too smooth. The hands were unreliable. The environmental detail was over-resolved in places and under-resolved in others. The work read, to anyone who cared to look, as generated.
The campaigns that ran in this mode tended to fall into one of two outcomes. Either they shipped quietly and were not discussed publicly. Or they shipped loudly, were noticed, and produced the first of what became a much longer running conversation about whether AI imagery was an honest tool for brand work or a corner-cutting one. Levi’s, in the months that followed, would put this conversation on the front page.
What this period set up
The period from late 2022 to mid 2023 set the conditions for everything that followed. Three positions emerged out of it, and most of the next eighteen months of industry conversation was a working-out of which one would dominate.
The first position was Heinz’s: AI is an idea generator and a topical instrument; use it for what it is, transparently, in campaigns that put the AI in the foreground.
The second was the stealth position: AI is a production tool; use it to generate finished work, undisclosed, hoping the audience does not notice the seams.
The third — the position CampaignsLive eventually built around — was that the right move was neither to foreground the AI nor to hide it, but to train tools narrowly enough that the output stopped being recognizable as AI and started being recognizable as brand creative. The seams went away not because they were hidden but because the model never produced them in the first place.
The next two years of the industry are the story of those three positions colliding. Heinz, in retrospect, was the cleanest expression of the first one. Everything that followed was the market figuring out the trade-offs of the other two.