What will actually kill us is the rubbish you are doing with AI

Sarah Connor would not find this funny at all. The AI industry has spent a while running the best end-of-the-world marketing campaign anyone can remember: resignations with extinction warnings, chief executives asking to slow down, and half the world talking about cartels and hidden agendas. While we argue about whether the machines are going to kill us, something is already doing damage, and it isn’t Skynet waking up: it’s what we let AI do with no oversight and no professionalism.

There’s an unwritten rule in communications: the more dangerous you make your product sound, the more powerful it seems. It isn’t new and AI didn’t invent it. In the seventies, Wilson Bryan Key, an academic and author specialising in subliminal messages, built a career claiming that advertisers airbrushed skulls into the ice cubes of whisky ads. That suicidal pull, that attraction to death, was supposedly what drew the consumer in. Like almost any good story, it had more falsehood than truth in it, but it has been taught in advertising and communications faculties for decades. And the underlying intuition wasn’t far off: an experiment published in 2010 in the Journal of Experimental Social Psychology found that “smoking kills” warnings pushed smokers whose self-esteem depended on smoking to smoke more. The death warning, turned into an advert for smoking.

The artificial intelligence industry is taking that rule all the way, and the underlying debate concerns all of us even if there’s little we can do about it. Where we can act is on the small AI zombie apocalypses we are breeding inside our own organisations. Death by AI starts slowly. And then, if it feels like it, all at once.

The artificial stupidity grenade

Dumping the mess on someone else. Passing the buck. Call it what you like. We can all picture what Tobias Lütke, chief executive of Shopify, means when he put a name to what is happening in the small print of the AI revolution. “We call those slop grenades that people toss at each other,” he said, as Fortune reported. “You don’t really read it, and now it has to be reviewed by your colleagues. You’re just letting AI do the work for you.” Slop, in this context, is the swill you tip into a pig trough: junk, time-stealing botch jobs. And it comes from someone, Lütke, who told Shopify that using AI was “a baseline expectation”: he forced adoption all the way through and is now swallowing the bill for that unprepared push.

A slop grenade is letting AI produce the work and passing it along without adding anything to it, no judgement, no supervision. Sometimes not even a quick read. Whoever throws it saves time and looks productive. The cost is paid by the next person down the line, and eventually the whole process collapses like a house of cards. And the organisation’s reputation with it.

Who tends to assemble this kind of corporate guerrilla ordnance? The same people who used to make everyone’s life miserable with endless emails, lovingly crafted PowerPoints and a bedsheet of a spreadsheet with fourteen tabs. They have now turned working with them into something worse: prettier, more interactive, more bureaucrap.

This AI apocalypse inside organisations isn’t free, of course. BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 US desk workers and published the result in Harvard Business Review: 40% say they received, in the previous month, what the authors call workslop, content that looks polished but is empty. Each piece costs an average of one hour and 56 minutes to redo, around 186 dollars a month per employee, according to an estimate the authors built on self-reported data. There is a second finding that matters more than the money: about half of those who receive one of these pieces start to see the sender as less creative, less capable and less reliable. The grenade also costs reputation, and it’s paid by whoever threw it without realising.

40%%

of surveyed desk workers received AI-generated content in the past month that looked polished and was empty

BetterUp Labs y Stanford · 1.150 trabajadores, EE. UU.

1 h 56 min

is the average time it takes to redo each of those pieces, paid by whoever receives it and not by whoever sent it

Harvard Business Review, 2025

186 $ $

is the estimated invisible cost per employee per month, calculated by the authors from reported time and average salary

Harvard Business Review, 2025

97%%

of Spanish SMEs already use AI in their regular work, according to the study commissioned and published by the vendor

OpenAI · 1.000 directivos de pymes

39%%

use it with no formal policy on what it is for, who answers for it and what gets reviewed before it goes out

OpenAI · 1.000 directivos de pymes

The tool is already inside almost every organisation: what hasn't arrived is the decision about what it's for.

Don’t look up, the apocalypse is in your inbox

And yes, the debate that blew up in September 2026 matters and is worth keeping in mind. Cut to the bone: a researcher resigned accusing OpenAI and Anthropic of racing towards superintelligence while “gambling with our lives”, their chief executives called for slowing the race and for a regulator to give them cover to coordinate with each other, and within days the word cartel was on the table.

Who says what and why matters to us, and a lot. Manel Pérez laid it out well in La Vanguardia: the warnings about risk to humanity come mostly from those holding the most advanced and most closed models, and the regulation they propose “could clip the wings of those who, lagging behind, are catching up fast”.

Sometimes, think the worst and you’ll be right. We have become cynical. Beaten into it.

Andrew Sharp, who covers the sector from Sharp Text, on analyst Ben Thompson’s network, goes further: “the people driving AI progress are the same ones auditing its risks to humanity, and they are all basically arriving at their conclusions by talking to one another”. And he asks what concrete harm AI has caused so far to justify pre-emptive regulation.

Good question. No plane has fallen out of the sky yet, and no nuclear plant has gone dark because of a swarm of unsupervised AI agents. But the mess it is starting to make inside a lot of companies is very much worth regulating within those organisations.

The big-picture debate concerns all of us. But perhaps there’s less need to look up, as in the film about the comet, and more need to look at the bureaucratic, pointless AI mess that is eroding your working life.

Grow up, damn it

One of the sharpest contributions to this slightly hysterical debate came from Steven Sinofsky. He ran Windows and Office when Microsoft stopped development to close the holes that made viruses possible, and he recalled how it was done: “We did not appeal to Congress to make us pause or to get the other companies to pause first. Grow up. Add logging, checking, debugging, and more to your products. It needs it.”

Swap “products” for “deliverables” and it works just as well for an agency or a content team: the domestic version of the frontier debate isn’t whether AI should be slowed down, it’s whether anybody is looking at what comes out the other end.

In Spain, gulp, almost nobody. A study commissioned by OpenAI covering 1,000 SME managers claims that 97% already use AI in their regular work. The figure is published by the company selling the tool, so read it with that label attached, but the two numbers beside it are the ones that could keep a manager awake: 39% use it with no formal policy at all, and 70% have employees using it on their own initiative. The tool is inside almost every organisation, and in four out of ten nobody has decided what it is for, who answers for what it produces, and what gets reviewed before it goes out.

That has its own name in security. Ralph Echemendia, the expert they call “the Hollywood hacker”, calls it shadow IT, as he explained in this interview: the use of external, unapproved tools, each one widening the attack surface. And he explains the loop that feeds it: “most of the things we do to have more security make us more vulnerable”, because “fear is what we mostly use when an incident happens”. People respond with more rules, the rules make the work harder, and everyone goes around them. The apocalypse loop again, in miniature, in your office.

That’s the most interesting part and the one least talked about. AI doesn’t make anyone careless: it amplifies what was already there. Whoever had method now produces more and better. Whoever never reread what they signed now fails to reread ten times faster, and every piece goes out, alas, with their organisation’s logo on it. Someone who was already a bad professional has now, on top of that, become dangerous with AI.

Someone who was already a bad professional has now, on top of that, become dangerous with AI.

The noise has turned into cacophony on headed paper, and it even looks like work. You don’t fight that by banning the tool, you fight it with order, with limits and with process, which sounds like bureaucracy and is exactly the opposite: it’s the only thing that stops it.

The whiteboard test

The best formulation of judgement I’ve read on this also comes from an engineer, Mitchell Hashimoto. He calls it “the whiteboard defense”: any organisation should be able to pull you aside at any moment and ask you to explain anything you have delivered to a client, and you should be able to defend the decisions you made. Nobody expects you to remember implementation detail, only to know why you did A instead of B. What reaches a client cannot be something you don’t understand, or a decision you can’t argue for and defend.

Echemendia says it from a security perspective and it works for any business project: “you can’t trust everything that comes out of a computer”. With AI, anyone can operate tools that used to take years to understand, and they do it “without understanding anything at all”. Capability arrives before judgement. Always.

And judgement is still what professionals are for.

Maybe AI will end civilisation as we know it. Or keep giving us scares until the final horror. It could also be that we find a cure for cancer and for diseases that were incurable until now, that it helps us fight climate change and brings prosperity and knowledge to everyone. Or that it shrinks our cognitive capacity to a minimum and we turn into the kind of society the prophetic film Idiocracy already showed us. We’ll see.

What AI is certainly doing is that in many organisations rational thinking has been delegated and an uncritical use of AI has been chosen, and the result follows: hours, credibility and clients lost to something entirely mundane, passing off as finished a piece of work that nobody has thought through, reasoned out, developed or checked.

Hell, half the time nobody has even read it.

The apocalypse has started in the office and the grenades are already flying towards your desk.

Are you going to keep looking up, at the debate about regulation, Terminators or Trump’s AI Force, or are you going to look down and get to work preventing the apocalypse in your own patch?

Make Sarah Connor proud.

Sources

• Wilson Bryan Key, Subliminal Seduction (1974) and Media Sexploitation (1976), where he claimed advertisers hid retouched images, skulls among them, in the ice cubes of drinks ads. His theses are discredited: Timothy E. Moore and Anthony Pratkanis rebutted them in Skeptical Inquirer (1992). The most cited popular rebuttal is Cecil Adams’s in The Straight Dope. It is used here for what it is, a legend.
• Jochim Hansen, Susanne Winzeler and Sascha Topolinski, «When the death makes you smoke», Journal of Experimental Social Psychology, 2010. Small sample: 39 student smokers. Mortality warnings produced more favourable attitudes to smoking among those whose self-esteem was based on smoking. Plain-language summary at the British Psychological Society. Cited as an indication, not as proof.
• Fortune, «Shopify CEO says employees’ ‘slop grenades’ are making more work for everyone else», 17 September 2026. Quotes from Tobias Lütke on The Knowledge Project podcast, and context on his 2025 internal directive about AI use at Shopify.
• Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano and Jeffrey T. Hancock, «AI-Generated “Workslop” Is Destroying Productivity», Harvard Business Review, 22 September 2025. Research by BetterUp Labs and the Stanford Social Media Lab. Sample: 1,150 full-time US desk workers, August and September 2025, self-reported. Figures also covered by Axios. The 186 dollars per employee per month is an estimate the authors built from reported time and average salary.
• Study on AI use in Spanish SMEs commissioned by OpenAI, presented on 17 September 2026 at the Digitaliza Madrid event with the Community of Madrid. Sample: 1,000 SME managers. Data reported by Infobae/EFE and Revista Byte. It is commissioned and published by the vendor of the tool whose use it measures.
• Manel Pérez, «¿Quién dice qué y por qué en la pugna sobre la IA?», La ventana indiscreta, La Vanguardia, Economics section, 20 September 2026. A map of positions between closed and open models, and the line about clipping the wings of those lagging behind. It is a signed analysis column: it contributes ordered judgement, not new data. It is also the source for the summary of the September episode that opens the section. In Spanish.
• Andrew Sharp, «Some of All Fears», Sharp Text, 18 September 2026. Sharp covers technology and media from Sharp Text, part of Ben Thompson’s Stratechery network, and co-hosts the Sharp Tech podcast with him. A declared opinion column with no data of its own, at the most sceptical end of the debate. Cited for the idea and the question, not as evidence.
• Francesc Peirón, «Cualquiera puede ser un hacker con la inteligencia artificial», interview with Ralph Echemendia, La Vanguardia, Vida section, 19 September 2026. Quotes on shadow IT, the security paradox, fear as the reaction after an incident, and distrust of what comes out of a computer. Echemendia is a cybersecurity consultant and speaker, not a researcher: he contributes professional judgement, not a study. In Spanish.
• Donald Trump, Truth Social post of 19 September 2026 announcing the “AI Force” and a future AI “czar”, reported by Fox News and NBC News. Both outlets stress that he gave no detail on structure, authority or membership. Cited as a fact, not as a political assessment.
• Steven Sinofsky, post on X, 17 September 2026.
• Mitchell Hashimoto, post on X, 16 September 2026, on “the whiteboard defense”.