Published July 24, 2023 · By CampaignsLive · Case Study
In March 2023, Coca-Cola launched “Create Real Magic,” a platform that let users generate AI imagery using a curated set of Coca-Cola brand assets — the bottle silhouette, the Spencerian script logo, the polar bears, the Santa illustrations — combined through DALL·E and GPT-4. The campaign was a partnership with OpenAI and Bain, the first time a brand of Coca-Cola’s stature had handed elements of its visual identity over to a user-facing AI generation tool.
The reception was mixed. Some of the user-generated work was striking; some of it was forgettable; some of it stretched the brand identity in directions Coca-Cola’s marketing leadership probably would not have approved internally. The campaign attracted predictable criticism — that it commoditized the brand, that it gave away creative control, that the outputs were not consistently on-brand.
Most of that criticism, in retrospect, missed the structural thing the campaign got right.
What was actually unusual about it
Most brands that tried AI in 2023 used it as a content-generation tool. Coca-Cola used it as a permissions architecture.
The platform was, at its core, a question about which parts of the brand could be shared with users and which could not. The answer the company arrived at was specific: certain iconic assets were available in the toolset; certain ones were not. Within the available set, users could generate; the brand retained control over what entered the canon by curating which submissions surfaced as official campaign artifacts.
This is a structurally different proposition from “AI generates ads for our brand.” It is closer to “AI generates community work, anchored in our brand vocabulary, with us as the curator.” The campaign was less about generation and more about co-creation with brand-identity guardrails baked in at the tool layer.
That framing pre-empted most of the failure modes brands ran into in the next eighteen months. Coca-Cola was not handing the brand to a tool to produce finished commercial work. It was running a community engagement program that happened to use AI as the canvas.
The training-data question, handled implicitly
The other thing the campaign did, almost in passing, was answer the question that has dominated the next two years of brand-AI conversation: what is your tool trained on, and what does it owe rights-holders?
Coca-Cola’s tool was trained on Coca-Cola’s own archive — its bottles, its logos, its illustrations, its decades of campaign work. The legal posture was straightforward because the IP was the brand’s own. The conversation about whether the underlying model had ingested someone else’s copyrighted work was avoided by limiting the user-facing tool to brand assets. The general DALL·E model sat underneath, but the user-facing surface was constrained.
The structural lesson here — narrow the surface, anchor it in proprietary visual material — is one that brand-AI tools have been working their way toward in fits and starts since. Most have not gotten as far as Coca-Cola did in March 2023.
What didn’t work
The campaign was not a clean success. Three problems showed up in the post-mortem.
The outputs were uneven. Even with curated brand assets, the underlying generation quality was constrained by what DALL·E 2 and GPT-4 could produce in early 2023, which was good enough for novelty and not yet good enough for production-grade brand creative. Some of the user-generated work read, visibly, as AI. The platform leaned into this — the campaign’s tone framed it as a creative collaboration with users, where rough edges were part of the texture — but at the level of brand-equity output, the work was uneven.
The signal-to-noise ratio on user submissions was low. As with any open-platform creative campaign, most of what users generated was undistinguished. The standout work surfaced through curation; the bulk did not.
The campaign’s commercial outcome was unclear. Coca-Cola never published a full performance breakdown. Whether the platform moved equity metrics, brand consideration, or any commercial KPI — the kinds of numbers that justify the line-item — was not made public. The campaign was widely read as a forward-positioning move rather than a performance-driven one.
The thesis it left behind
What “Create Real Magic” left behind, for the brands that followed, was a specific thesis about how branded AI tools should work.
The thesis had four parts:
- Anchor in proprietary visual material. A tool trained on your archive produces work that lives inside your brand world. A tool trained on the open internet produces work that lives in the open internet’s visual register.
- Constrain the surface. What users can do, and what the tool will not do, is part of the design. An infinite-canvas tool dilutes the brand; a constrained one anchors it.
- Position the AI honestly. The audience knows AI is involved. Pretending otherwise creates risk; foregrounding it appropriately neutralizes it.
- Curate the output. Generation alone produces noise. Curation — by humans, with brand judgment — is what turns generation into a canon.
These four points, more or less, are what the brand-AI tooling market has been working its way toward since. The tools that are succeeding in 2026 share all four. The tools that are struggling tend to violate at least two.
Coca-Cola did not invent this framing; it operationalized it earlier than most. For an industry that spent 2023 mostly debating whether AI imagery should be allowed in brand work at all, the campaign was a useful demonstration that the question was already the wrong one. The question was how to do it, not whether.