When something goes wrong with an organization's AI-driven venture, who is responsible? The human? The machine? Or the organization that brought them together? This question isn't new; but what's new is that it has stopped being hypothetical. AI has dragged this question out of the philosophy seminar and onto the org chart, and the people who actually have to answer it don't agree.

In the Thomson Reuters 2026 Future of Professionals Report, 47% of respondents said they put final responsibility for an AI-assisted error on the individual professional. Not on the firm, not on the tool. The plurality lays it on the person who pressed the button.

This is the question the whole professional services world is debating right now, from the people theorizing about accountability to the engineers building the tools that force the issue onto the table. Everyone has a stake in the answer, and almost no one agrees on it. So when a story surfaced a few months ago that ran straight into this question, the answer felt buried somewhere inside it.

But to find it, we need to take a bit of a trip.

“Slap it on the desk”

In May, Starbucks Korea launched a promotion for a line of coffee tumblers, one of them called the Tank. The company built a day around it, branding May 18 "Tank Day," discounting the sets, and pushing the campaign across its app and social channels under a cheerful slogan about slapping the desk with the Tank tumbler. On any other day, attached to any other product, none of it would have mattered.

But May 18 is not any other day in South Korea. On that date in 1980, the military sent tanks and paratroopers into the city of Gwangju to crush a pro-democracy uprising, killing hundreds of civilians. Forty-six years later, a tank on that date is not a neutral image. For the country's far-right internet culture, it summons "Jeon Ttangkeu" (전땅크), roughly "Chun the Tank," the sometimes affectionate, sometimes mocking nickname for the general who led the crackdown.

The slogan landed just as badly. Older Koreans hear "slap the desk" and think of 1987, when police claimed a student activist named Park Jong-chul had simply dropped dead in custody after an officer slapped a desk. In fact they had tortured him to death. Exposing that lie helped topple the dictatorship the same year.

Neither reference is obscure. To a Korean reader, both are unmistakable.

Someone to blame

Starbucks Korea's post-mortem arrived a week later, and it is worth taking seriously, because it is not far-fetched. Simply put, there was no plot. A team inside the company's commerce division had been racing to ship the campaign, hunting for a slogan to rhyme with an existing tumbler tagline, and somewhere in the rush they asked an AI for options and ran with one. May 18, they said, never crossed anyone's mind. The work cleared a four-stage approval process and seven separate signoffs, several of which, the company admitted, were granted without anyone opening the attached file. The legal review was skipped to save time. The catastrophe was a Friday accident: a rushed slogan waved through a review process that turned out to be theater.

That word, theater, hides the sharpest fact of all. The same technology blamed for writing the slogan would have caught it in a heartbeat. Point a model at the finished campaign, ask whether anything on the page is historically radioactive, and Tank Day, May 18, and Gwangju light up instantly. The team reached for AI to write the copy and never thought to reach for it to read the copy. The machine did not fail them. They wielded it in its most dangerous function and left aside the one that might have protected them.

The same technology blamed for writing the slogan would have caught it in a heartbeat.

Then Starbucks did what organizations do when the story is told that way: it went looking for the people to blame. Members of the commerce team were removed, the division head was left on the hook pending the rest of the inquiry, and the head of Starbucks Korea was gone inside the week. The "who" got answered, and answered hard. Somebody moving too fast on a Friday became somebody without a job.

The story worth having

Which brings me back to the human, the machine, and the organization. Given three places to lay the blame, the story Starbucks Korea was telling chose the humans, and the humans paid the price.

However, the safety researcher Sidney Dekker draws a distinction that fits here. After every failure there are two accounts. The first story looks back, finds the people closest to the mistake, and punishes them. The second story, on the other hand, assumes those people were rational, doing the sensible thing given the training and instruments in front of them, and asks what made the catastrophe feel sensible in the moment. The first story ends with a signature on a dismissal letter. The second asks how the letter came to be needed at all.

Starbucks Korea's whole defense is that this was negligence, not malice. Sit with what that concedes. If nobody meant it, then a product called Tank, launched on the anniversary of a massacre carried out by tanks, with a torture cover-up echoed in the copy beneath it, moved through seven signoffs and a legal review and struck no one as worth a second glance. Malice would at least require someone in that room to understand what the words meant. Negligence on this scale means something worse. It means the vocabulary had gone ambient enough inside the building to read as ordinary marketing.

That was echoed in the most revealing moment of the whole affair, when a staffer's first honest reaction to the backlash was bewilderment that anyone was upset at all. That is not the sound of someone hiding a crime — that's the sound of an organization that had no idea what it was bringing to market.

Tolerance for risk

And what the company did next makes the point better than any critic could. Alongside the firings, Starbucks Korea closed its stores early and put all of its people, some 24,000 of them, through mandatory history and cultural-sensitivity training. On paper it looks like contrition at worst, and an honest attempt to address the second story at best. In practice it was the first story wearing the second story's costume.

The campaign was built in a commerce division; the baristas crafting mochas in Busan and Daegu had nothing to do with it, and the public saw that at once. Training the store floor for a decision made in headquarters did not read as accountability, it read as misdirection — a company reaching for the nearest available bodies rather than the decision that produced the copy. The blame it was trying to disperse snapped back, harder, onto the culture that let the marketing plan ship.

When the people who set the tone treat something as survivable, everyone below them learns to stop flinching. No memo required.

Because that is where these failures live. Tolerance for risk is set at the top, and it rarely has a single author. When the people who set the tone treat something as survivable, everyone below them learns to stop flinching. No memo required. The reflex simply never forms, and sooner or later someone walks face first into a crisis that a different tone would have taught them to avoid. There is usually no villain here, no provocateur at the top, only a tone that drifted until nobody in the room could hear it anymore.

I went in wanting a name. Starbucks Korea handed me a fistful and it changed nothing, because if you punish the hands on the keyboard without repairing the ideas that guided them, then everyone watching can tell the difference.

I write about AI accountability for a living, and I sit next to the engineers who put that button on everyone else's desk. I have no illusion we are immune, because nobody on autopilot is. Somewhere in your organization, work is clearing every checkpoint you built and telling you nothing, and that silent output is the most honest statement your organization will ever produce about itself, precisely because nobody meant it as a message.

The 47% from our Future of Professionals Report who say they would assign the failure to the person who pressed the button are reaching for the first story. The second story is harder, and worth more. It asks not who pressed the button, but what made pressing it feel normal.

You can find more by this author here