AI amplifies whatever you plug into it. Most of us are plugging in a deaf kazoo.
There is that remarkable scene in This Is Spinal Tap that every slightly overweight air guitarist over fifty remembers, and that every CEO should be made to watch at least twice a year, preferably before budget season. Nigel Tufnel, lead guitarist, hair of a startled spaniel, gives the filmmaker Marty DiBergi a tour of his gear and stops, with the tenderness of a drunken man showing you his firstborn, at one particular amplifier. The dials on this one go to eleven. Everybody else’s stop at ten. So when you’re already at ten, all the way up, and you need that extra push over the cliff, you go to eleven. One louder. Marty, who appears to be the only person in the film with a functioning frontal lobe, asks the obvious question: why not just make ten louder and call that the top? Nigel looks at him for a long, pained, genuinely wounded moment. “These go to eleven.”
I keep thinking about that long pause. We have probably built the most powerful amplifier in human history, and most of the planet is standing around it admiring the dials.
An amplifier has a very simple, beautiful moral philosophy, which is… that it has none. Plug in Jimi Hendrix and you get Hendrix, louder, bigger. Plug in Bach, and the vicar comes down the cathedral stairs with a nosebleed. Plug in the neighbors’ nephew, who received an electric guitar for his communion and has spent the two years since discovering the same three wrong chords, and you get those, bigger, with reverb, through the bedroom wall at eleven at night. Generative AI is that amp. It takes whatever you bring to the session, your knowledge or the lack of it, your curiosity or your hurry to get to the pub, your taste or the tragic, deplorable absence thereof, and turns it up. Ignorance, biases, laziness and mediocrity go into one input jack. Curiosity, contrarian thinking and the odd flash of pure genius go into the other. The machine doesn’t check which cable is which, and that is both the very point of an amplifier and the entire problem with this one.
So the eternal conference-panel question I keep getting, will AI make us smarter or dumber, deserves the long, cold Nigel stare. A pause. A sigh. Yes, it will. Both. Sadly, often to the same person, before and after lunch and a quarter liter of cheap rosé. I can’t help noticing that at the moment the dumb channel is winning on volume by a marathon.
A quarter of a million euros of kazoo
In July 2025, Australia’s Department of Employment and Workplace Relations published a whopping 237-page review it had bought from Deloitte for roughly a quarter of a million euros (or one fairly serious Patek Philippe). It quoted a federal court judgment that had never said what the report said it said, and it cited research that did not exist, including a book attributed to a Sydney law professor who had never written it. Ouch. Deloitte agreed to refund the final instalment, slipped in a cheeky line admitting that Azure OpenAI GPT-4o had lent a hand, and assured everyone that the corrections “in no way impact or affect the substantive content, findings and recommendations in the report.” Which is roughly what the chef says about the hairy rat in the soup. Bon appétit.
I would love to laugh louder than I do. I can’t. I work in the business of being paid to think on other people’s behalf, and I use these tools every day, so I know the moment as well as anyone: it is late, the deck is due at nine, the machine hands you a perfectly plausible paragraph with a perfectly plausible footnote, and the little voice that is supposed to say go and check that has already put on its pajamas and is pouring itself a smoky whisky. Seen from inside the chat window, laziness feels exactly like efficiency.
Ignorance is worse, because ignorance can’t hear the wrong notes. When Harvard and BCG gave GPT-4 to 758 consultants, the ones using it on tasks inside what the researchers called AI’s “jagged technological frontier” finished 12.2% more tasks, 25.1% faster, at more than 40% higher quality. Then the researchers slipped in one task that sat just outside the frontier, where the machine sounds every bit as confident and happens to be wrong. Consultants with AI were 19 percentage points less likely to get it right than their colleagues without it. Nineteen points. Same people, same tool, same nice offices. One big fuck-up. The amplifier made them brilliant right up to the edge of what it could do, and then played them straight over that edge, and they followed, like sheep, with the serene confidence of lemmings on a team-building day. If you know the terrain, you can feel where the ground ends. If you don’t, eleven is exactly what Nigel promised it would be: that extra push over the cliff.
Mediocrity got a raise
Fairness, briefly, because I am grumpy and not blind. In that same study the below-average consultants improved by 43%, against 17% for those above average. In a story-writing experiment published in Science Advances, access to AI ideas made stories more creative, better written and more enjoyable, “especially among less creative writers.” I am all for that. Plenty of people have spent their working lives held hostage by the blank page, by a second language, by dyslexia, by having nobody around to bounce an idea off, and they finally have somebody. Sneering at that is cheap, and I try to keep my sneering expensive (I am, after all, a consultant).
The same paper, a few lines further down, notes that the AI-assisted stories “are more similar to each other than stories by humans alone.” The floor went up and the ceiling didn’t move, and the space in between is filling with what I can only call hotel-lobby prose, pseudo-intellectual muzak: clean, pleasant, faintly scented, and impossible to remember once you’ve left the building. Beige and slop. Spend four minutes on LinkedIn (I would advise against five) and you will meet a thousand people who all learned the same three leadership lessons from the same marathon, in the same cadence, under the same rocket emoji. Most of these people did not even bother to think.
The howl
Anybody who has ever walked onstage holding a microphone knows what happens when you point it at the speaker. The amp hears itself, amplifies itself, hears that, and within a second the whole room is clutching its ears. In April 2025 OpenAI shipped an update to GPT-4o that turned out, in the company’s own words, “noticeably more sycophantic,” given to “validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions.” Users posted screenshots of it cheering on a business plan for, quite literally, shit on a stick. OpenAI rolled it back within days, which is to its credit, but the incentive behind it never left the building. A machine that learns from our thumbs-up learns very quickly that we like being agreed with, which is the most reliable finding about human beings since the invention of the court jester (who, let’s remember, was the one person at court allowed to disagree). Point that at the ignorant and you get certainty. Point it at the lazy and you get contentment. Point it at your brother-in-law in the do your own research T-shirt and he now has a research department. God help us all.
The other input jack
The other channel exists, though, and when somebody plugs something real into it, the sound is extraordinary. In 2024 Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for AlphaFold, which cracked a problem chemists had been banging their heads against for fifty years and has since predicted the structure of virtually all 200 million known proteins, for more than two million researchers in 190 countries. Tucked into the committee’s plain-language explainer is a phrase I would happily have engraved above the door of every AI steering committee in Europe: the improved model was “coloured by Jumper’s knowledge of proteins.” Decades of carefully deposited protein structures, a team that knew its biology cold, and a stubborn curiosity about a problem most people had filed under impossible, try again in 2080. That is what you have to bring to the session before the machine can make it louder: experience, grit, conviction, craft, an opinion.
Contrarian thinking is the input the machine is worst at supplying itself, for a reason that is almost embarrassingly technical. A large language model is trained, at its core, to predict the probable next word. A genuinely contrarian idea is improbable by definition (that is why nobody has had it yet). Ask for the consensus and you get the consensus, beautifully formatted, by return of post. Walk in with a heretical hunch, on the other hand, and you get a superb sparring partner that will steelman you, poke holes in you and dig up every inconvenient counter-example on the internet. It will very rarely be the one to start the fight. Somebody has to walk in already holding the hunch.
I happen to have a brain that was wired to start that particular fight, and it is a slightly surreal experience to watch the very trait that got me asked to “keep things constructive” for most of my career quietly become the scarcest input in the economy. Vindication would be a stretch: contrarians are wrong a great deal, often loudly, and I have the scar tissue to prove it. But the habit is worth more now than it was a decade ago, and we are doing almost nothing to produce it.
Fire the devil’s advocate and see what happens
The Catholic Church worked this out centuries before Silicon Valley. From the late sixteenth century, every candidate for sainthood had to get past the promotor fidei, better known as the devil’s advocate, a canon lawyer paid to argue that the candidate was not, in fact, quite as saintly as advertised. In 1983 John Paul II reformed the process, cut the number of required miracles and effectively retired the adversary. Between 1900 and 1978 the Church canonized 98 saints. John Paul II alone canonized 482. Remove the one person whose job is to say hang on, and saints roll off the line like Volvos in Ghent. (I say this as an atheist who has already needed one Pope this year, so I am hardly in a position to throw stones at the Vatican.)
Most companies keep a contrarian the way most Belgian households keep a fire extinguisher: one, somewhere in a cupboard, never inspected, and deeply resented the day it goes off. Every strategy deck I have ever seen has a slide about “encouraging challenge.” Very few have a budget line for it, a seat for it at the decision, or a career path that survives it. And now we have handed everyone a machine that agrees with them fluently, at scale, for twenty dollars a month. Hallelujah.
A school for Martys
So yes, I want classes. Actual classes, with a timetable and homework, in contrarian and critical thinking: in secondary schools, at university, in onboarding, and in every leadership program that currently spends three days teaching people to “manage stakeholders.” The obvious objection is that you’ll breed a generation of flat-earthers. The flat-earthers are doing splendidly without a curriculum, thank you. What they lack is method, and method is what you can teach. The good kind of contrarian starts by being contrarian about their own conclusions, which is a skill, which means it can be practiced, which means it can be graded.
We even know it works, and fast. Sam Wineburg and Sarah McGrew at Stanford watched PhD historians, Stanford undergraduates and professional fact-checkers try to judge the credibility of websites. The historians and the students read “vertically,” staying on the page, admiring its logo, its footnotes, its About section, and they were regularly taken for a ride. Doctorates and all. The fact-checkers glanced at the page, left it almost immediately, opened new tabs and checked what the rest of the world said about the source, and reached better conclusions in a fraction of the time. In a later randomized study with around 500 high school students, those who got six 50-minute lessons in this kind of reasoning were twice as likely to spot a questionable website as they had been before. Six lessons. That is less time than the average company spends every year on its compulsory cybersecurity e-learning, the one with the cartoon hacker in a hoodie that everybody clicks through with one hand while eating a sandwich with the other.
Now imagine the version built for the machine age. You would teach people to leave the chatbot’s answer the way the fact-checkers left the website, and check it somewhere that has no interest in pleasing them. You would make students argue the opposite of what they believe once a week and grade them on how well they lose. You would have them test the machine on the one subject they know intimately, a sport, a recipe, a dead language, the gearbox of a 1987 Golf, until they have personally watched it lie to them with a straight face and can remember exactly how that felt. And you would start grading the question instead of the answer, since the answer is now free and the question is the only part the machine can’t write for you, because it has no idea what you don’t know.
None of it costs more than a decent coffee machine. All of it requires teachers, managers and boards who reward the person who says wait a minute more generously than the person who says looks great, ship it.
Marty DiBergi asked the only critical question in the whole of Spinal Tap, and he asked it politely, of a man holding a very expensive amplifier. We are currently pouring an indecent fortune into building louder Nigels. I would settle for a few more Martys in every classroom, and a boardroom willing to sit through the pause.
