We trusted the machine and stopped looking
Starbucks Korea asked an AI for ideas, trusted what came back, and approved a campaign nobody actually checked. The bill: a fired CEO, a national boycott, and a quarter-billion in refund demands.
Last month Starbucks Korea launched a promotion for a new line of “tank” tumblers. They named the campaign “Tank Day” and set it live on May 18. If you don’t know what May 18 is in Korea, that’s the whole story in miniature. It’s the anniversary of the 1980 Gwangju massacre, when paratroopers crushed pro-democracy protests against the military regime and, by victims’ groups’ count, hundreds were killed. One of the campaign slogans, translated as “thwack on the desk,” landed close to the cover story police used after the 1987 torture death of student activist Park Jong-chul, when authorities claimed he died because an officer “hit the desk with a thwack.” So you had a tank, on a massacre anniversary, with a slogan that rhymed with a state cover-up.
The backlash was immediate and brutal. Videos of people smashing Starbucks tumblers spread online. Protests outside stores, boycott calls, government ministries cutting ties, the defence ministry suspending a partnership. Card spending at Starbucks fell 26% in a single week, and customers began demanding refunds against an estimated 400 billion won, around 260 million dollars, held in prepaid cards. The CEO of Starbucks Korea was fired the same day the campaign launched. The billionaire chairman of Shinsegae, which runs Starbucks in Korea under licence, bowed three times at a televised press conference. Both men were apparently also later booked as criminal suspects by police. Starbucks’ global headquarters in Seattle called it “an unacceptable marketing incident.” All of this over drinkware.
Then came the explanation. Shinsegae said the marketers had chosen the slogan after consulting an AI tool for suggestions, and that the May 18 anniversary had simply never crossed anyone’s mind. The investigation found no evidence the campaign was deliberate. And some of the managers who approved it signed off without ever opening the email attachments showing the marketing material. Speed was the priority. The AI gave an answer, and the answer went out the door.
That last part is what I want to talk about, because it’s the part most companies are walking straight into.
The easy version: blame the machine
The easy read is “AI did it.” A tool with no sense of history generated something offensive, and a brand paid the price. Cue the think-pieces about the dangers of artificial intelligence.
I don’t buy it, and reaching for it lets the actual problem off the hook.
The AI behaved exactly as designed. You ask a language model for slogan ideas, it gives you slogan ideas. Fast, fluent, confident, and with no idea what a given date means to fifty million people. That isn’t the model failing. That’s the model doing precisely what it does. Expecting it to flinch at “May 18” is like blaming a calculator for not noticing your budget is immoral. Wrong tool, wrong question.
If the tool did what it always does, the failure lives with the people who decided its output was good enough to ship without checking.
The real failure: trust without a check
Here’s the uncomfortable mechanism. AI output doesn’t arrive looking like a rough draft. It arrives looking finished. Clean sentences, confident tone, the surface polish of something that’s already been thought through. And the better it looks, the less anyone feels the need to verify it.
That’s the trap. A scrappy first draft from a junior marketer gets scrutinised, because it looks like a first draft (although, to be honest, that junior now would have likely also drafted it with AI, come to think of it). The same idea, laundered through an AI tool into fluent, professional copy, gets waved through, because it looks done. The polish buys trust it hasn’t earned. Nobody opening the attachment isn’t laziness so much as misplaced confidence. On some level the assumption was: the tool handled it, it reads fine, ship it.
We’ve spent the last couple of years being told AI will make us faster. It does. What nobody puts on the slide is that the speed comes from removing steps, and the easiest step to remove is the human check, precisely because the output looks like it doesn’t need one. Fast and unverified is a great combination right up until the one time the machine is confidently, catastrophically wrong about something it was never equipped to understand.
The thing AI cannot do is care whether it should be saying something. It has no stake. It doesn’t know it’s about to walk a global brand onto a massacre anniversary, and it will never feel the cold drop in your stomach that a Korean marketer would have felt the instant they saw “Tank Day, May 18.” That feeling is the check. Outsource the work and keep the check, and you’re fine. Outsource the work and trust the output enough to skip the check, and you’ve automated your way past your own judgment.
The part that should worry anyone planning headcount
This connects to the conversation every leadership team is actually having, which is how many people you still need now that the tool can “do the work.”
The pitch goes: the model drafts the campaign, writes the copy, generates the assets, all without the salaries. For the production, that’s often true. The model can absolutely generate “Tank Day.”
But look at what failed in Korea. The campaig generation didn’t fail. The judgment did. The missing function wasn’t “make a campaign.” It was “know that this campaign, on this date, in this country, is a disaster.” That function is a person. Usually several. The local marketer who carries the context the tool can’t hold, and who’s willing to say “wait, what day is this?”
Those are exactly the roles that look most cuttable when you’re staring at an impressive AI demo. They don’t produce the obvious artifact. They produce the absence of disasters, which is invisible right up until it isn’t. On a spreadsheet they look like cost and friction. In reality they’re the only reason the friction, the slowing-down, the second look, exists at all. Cut them to let the AI run faster and you haven’t removed overhead. You’ve removed the part of the system that knows when to stop.
So the question isn’t “should companies use AI?” Of course they will, and they should. It’s a genuinely useful tool. The question is what you keep around it. Keep the people, and make their job to check what the machine produces, and AI makes them faster and sharper. Cut the people because the output looks finished, and you’ve kept the thing that generates confident mistakes and fired the only thing that catches them.
What I’d actually take from this
Less advice, more what I’d be checking if this were my client.
Treat AI output as a draft, not a decision. The more finished it looks, the more deliberately someone has to verify it against the things the tool can’t see. Context, history, who’s going to be in the room when it lands. The polish is the risk, not the reassurance.
Put the check somewhere it can’t be skipped for speed. At Starbucks Korea, speed was the stated priority and the review didn’t happen. If your only safeguard is “someone will probably look,” then on the day it matters most, the day everyone’s moving fast, nobody will.
And be honest about which roles you’re cutting. “AI can do this now” is true for the production. It is not true for the judgment, the context, and the willingness to stop the line. Confuse the two and you’ll save money for three quarters and lose a CEO in the fourth.
Starbucks Korea asked a machine for an answer and trusted it more than they checked it. The lesson isn’t that AI is dangerous. It’s that AI is happy to hand you a confident answer and let you skip the part where a human decides whether it should ever see daylight.
The machine can write it. It still can’t care whether you should run it. That part is still your job.
Sources if you want to go down the rabbit hole: The Guardian, Raphael Rashid, “How a Starbucks marketing stunt spiralled into mass boycotts in South Korea,” 6 June 2026; PRovoke Media, “Starbucks Korea’s Tank Day Firestorm: Did Speed Outrun Common Sense?”; Bangkok Post, “Starbucks Korea says AI suggestions led to ‘Tank Day’ campaign flop”.