Starbucks Korea Didn't Have an AI Problem. It Had a Content Governance Problem.
- David Golub
- Jun 22
- 6 min read
Updated: Jun 22
That hissing sound you hear from Starbucks headquarters is the frothing wand of crisis control, cleaning up a marketing catastrophe in Korea where an AI-generated campaign ran straight into some of the country's most painful recent history.
The ad was written by AI, but this failure was all-to-human. This is a story about flawed marketing governance and a timely illustration of what can happen when no one minds the store between machine output and national audience.
In brief: on May 18, the anniversary of the 1980 Gwangju uprising, Starbucks Korea launched a "Tank Day" promotion for a line of oversized tumblers, paired with the slogan "thwack it on the desk." The date called up the tanks of the military crackdown that killed hundreds of pro-democracy protesters. The slogan echoed the lie police told after the 1987 torture death of student activist Park Jong-chol, that he had simply collapsed when an interrogator "hit the desk with a thwack."
The backlash was instant. The Starbucks Korea chief executive was fired within a day, Chairman Chung Yong-jin of the Shinsegae Group (the local owners) apologized on national television, sales fell sharply, government ministries pulled Starbucks from official events and on June 22 every store in the country closed early for the first time since 1999 so staff could sit through mandatory training in modern Korean history.
Then the internal investigation surfaced the detail that turned a cultural blunder into a story about the operating risks to company reputation in the LLM era. The slogan was generated by an AI chatbot and, according to the reporting, pushed into production while seven reviewers did something else besides look at it.
Pass me some New Coke, but don't blame the AI. The campaign worked in the sense that matters to an algorithm: the phrase was catchy and it spread. What broke was the human layer that is supposed to ask whether catchy is the same as correct.
Another Misfire from the Coffeehouse
This isn't the coffeehouse's first brush with PR disaster. While Starbucks has spent decades selling itself as a socially conscious brand that stakes its identity on social progress, in fact the company has badly missed that mark more than once on political, cultural and labor grounds.
In October 2023, after a Starbucks Workers United account posted "Solidarity with Palestine", the company sued the union for trademark infringement. The union countersued for defamation, and the dispute hardened into a worldwide boycott by consumers who read the lawsuit as siding with Israel. Starbucks shed roughly $11 billion in market value over about two weeks as the boycott spread, though analysts noted other forces were weighing on the stock too.
That sat on top of a union-busting fight running since baristas began organizing in late 2021. The National Labor Relations Board has found Starbucks committed hundreds of labor-law violations, including illegally firing pro-union workers, and the standoff escalated into the longest barista strike in the company's history on Red Cup Day in November 2025.
In 2018, two Black men were arrested at a Philadelphia store for sitting without ordering, a viral video that pushed Starbucks to close more than 8,000 U.S. stores for an afternoon of racial-bias training for 175,000 employees and to open its cafes and restrooms to anyone, purchase or not. In 2015, then-chief executive Howard Schultz's #RaceTogether campaign asked baristas to write the slogan on cups to start conversations about race, and was ridiculed so widely that the cup-writing was gone within a week.
Every one of these reputation-shredding decisions was made by a human: a lawsuit, a labor strategy, a manager's 911 call and a marketing brainstorm. Tank Day is the first where an AI was in the room when the bad idea was born.
The AI Did Exactly As Asked
Strip away the outrage and the machine's behavior is mundane. Asked for a snappy, rhyming line to sell a tumbler called the Tank, a language model returns the highest-probability pattern. "Tank" pairs with strength and durability. "Tak," the Korean onomatopoeia for a thump on a table, sounds good in ad copy. Both associations are statistically dominant, and both are harmless 98 times out of 100.
Cultural trauma lives in the other two. As a crisis-communications specialist quoted by PRovoke Media put it, a model optimizes for the 98% of its training data where "Tank" plus "tumbler" is innocent, while the hyper-local, highly charged anomalies sit in the sliver it never takes into account.
Ask the same model point blank what happened in Gwangju on May 18 and it will recite the history accurately. Generating a slogan, it has no reason to stop and check the calendar. It can reach the facts. It does not know when to be afraid.
That is not a defect to be patched in the next release. It's the nature of the tool. A statistical prediction doesn't come with situational awareness, no instinct that whispers "are you sure about that one?" Expecting otherwise leads to tears.
The Failure Was the Users, Not the Tool
Here's something to think about: the AI did not choose May 18.
According to the company's account (per Korea Times), the e-commerce team picked the Tank Day theme and the date for ordinary commercial reasons. Weekday online sales run hotter than weekends, and the tumblers needed to move. A human chose the date. A human chose the product name. Then AI was handed a suicide brief.
And the work did not go out unseen. Seven people reportedly signed off across a four-stage approval process. The investigation found no evidence anyone meant to mock the dead, which is the most damning finding of all, because it means the disaster required no malice. It required only that no one do their jobs.
Astonishingly, some of the managers who approved the campaign, by several accounts, did not open the attachments that held the materials they were signing for.
AI did not break the process. It just removed the last bit of friction that might have saved face. The output looked finished, so it got rubber-stamped, which is a polite word for approving what you never bothered to look at.
Governance Is Judgment, Not Sign-Off
There's a distinction we keep returning to at Agentic Foundry to think about how companies related to AI: the critical difference between human-in-the-loop and human-on-the-loop oversight. On paper, Starbucks Korea was the picture of human-in-the-loop. Seven humans, four stages, signatures all the way down.
Yet what the chain lacked was a single human exercising judgment at the loop.
A loop is not the same as judgment. A signature is not the same as a review. The Gwangju echo is something millions of Koreans felt in their bones. Just because AI seems clever does not mean governance can always be set to autopilot. Humans need to supply judgment on purpose.
For a marketing organization, that turns verification from a nice-to-have into a requirement: a gate, with a named owner who has both the authority and the time to abort a risky launch. It means treating AI output the way you would treat a draft from a talented junior, useful and fast but not shippable without scrutiny.
As AI pushes content volume past what any approval chain can red-line, it means encoding that judgment into the system itself: the brand rules, the do-not-touch list and the dates that should trip a mandatory human check, not hoping a tired manager catches it in semi-sleep plowing through screens of content.
Building that judgment into an AI content system is the harder and more interesting engineering problem, and it is the one that actually protects the brand. AI is strong at execution and weak at the origination and judgment that carry real context and conviction. We have made that argument before, and Tank Day is perhaps the most expensive illustration of it yet.
Now is the part where you ask yourself: if an AI handed your team a catastrophic marketing idea tomorrow, who is positioned to catch it, and do they have the authority and the attention to say no?
Agentic Foundry: AI For Real-World Results
Learn how agentic AI boosts productivity, speeds decisions and drives growth
— while always keeping you in the loop.
