Case study

Starbucks Korea’s “Tank Day”: Where Labels Only Partly Help

In May 2026, Starbucks Korea ran a “Tank Day” promotion on the anniversary of a deadly 1980 military crackdown on a pro-democracy uprising in the city of Gwangju. Staff said one slogan came from AI. Most of the harm came from human choices that labels can’t catch.

What are the labels? The Evidentiality Framework asks an AI to mark each claim it writes: (g) generated, its own work; (u) given, passed to it by someone else; or (m) checked against a named source. The marks stay on a claim while people work with it. How the labels work.

What Happened

On 15 April 2026, the “Tank Day” name and the 18 May launch date were set. No source says AI chose them.
On 8 May, the online sales team added a slogan, “Tak on the desk!”, without telling management. They told investigators they had asked AI for the wording.
A memorial group said the phrase recalls the military government’s cover-up of the 1987 death of the student activist Park Jong-chul, which police explained away with a story about banging on a desk.
The promotion launched on 18 May, the anniversary, and was pulled the same day. The head of Starbucks Korea was dismissed within a day. Starbucks’ US headquarters later apologised in writing.
An investigation by Shinsegae, the group that runs Starbucks in Korea, found that seven approvers across four stages raised no objection, some without opening the design file, and that the usual legal review was skipped. Its own investigation found no clear evidence of intent.
In August 2026 police searched the company’s headquarters, after a complaint from a civic group, on suspicion of insult. The investigation is ongoing; that is not a finding that anyone did wrong.

Follow the Claim

This is an illustration of how the labels would have worked, not a test. It only holds if the conditions under “What Would Have Had to Be True” held.

Follow the slogan, and the parts labels can’t reach

Key: as it happened, the type gets bigger as the claim sounds more certain. With labels: red (g) generated: written by the AI; green (u) given: passed on, with who said it.

As it happenedWith labels

15 April 2026 · Marketing team

As it happened“Tank Day”, set for 18 MayA human choice. No AI involved, as far as reported.

With labelsNo label applies: people chose this.Labels track where words came from. They can’t spot a bad idea.

8 May 2026 · AI tool → online sales team

As it happened“Tak on the desk!”Wording staff said came from AI, added without telling management.

With labels(g)Tak on the desk!(/g)If the tool used the labels: marked as written by the AI.

8–17 May 2026 · Team → seven approvers

As it happenedApproved at four stages, with no objectionsNow it’s signed off.

With labels(u)Tak on the desk!(/u: AI tool, not reviewed)Approvers can see this line came from an AI and needs a careful human read.

18 May 2026 · Launch

As it happenedLaunched on the anniversaryThe harm is done.

With labelsMaybe the slogan is cut. “Tank Day” on 18 May still goes out.This is why it’s only a partial fit.

The first row is the main problem, and labels don’t touch it. The slogan in English is a translation of the Korean (책상에 탁!).

How the Labels Could Have Helped

The AI-written line stands out. Approvers see which words came from an AI tool.
A second look, at the right line. Seven approvers might have stopped on a flagged slogan.
That’s all. The labels touch one line of a campaign whose main problem was a human idea.

What Would Have Had to Be True

The four conditions every case shares: the AI tool used the labels; it labelled its own work correctly (the weakest link: in our tests, AI sometimes mislabels its own work); the label stayed on when the text was copied; and someone owned a rule that unchecked claims don’t go further. In this case:

The rule is owned by the approvers, and only applies if the staff account that the slogan came from AI is right.

What Already Existed

A four-stage approval chain. Seven people signed off. Some didn’t open the design file.
Legal review. Used on earlier campaigns, skipped this time to save time.
A simpler check would have caught it: a calendar of sensitive dates. Labels can’t do that job.

What the Labels Wouldn’t Have Caught

The name and the date. “Tank Day” on 18 May was a human decision. No label would mark it.
Knowing the history. The team said they never thought about 18 May. Labels don’t give people context they lack.
Rubber-stamp approvals. A label only helps if someone opens the file.