Schrödinger’s AI: simultaneously too stupid to write an email and clever enough to end the world.

There are all sorts of genuine moral and environmental arguments against machine learning and data centres, but as an AI advocate, one particular whine from a particular group of people really grinds my gears. These arguments question the utility of large language models, and invariably the comment shines a light on the unfamiliarity of the author with the topic of their observation.

I’ve also noticed these comments often share two conflicting viewpoints that are frequently held by the same observer. The person who claims AI is useless because it writes sloppy text or computer code will also warn against the same fumbling technology taking our jobs, or taking over the world and killing us all in a ‘Skynet becomes self-aware’ doomsday scenario from the Terminator movies.

I can’t help thinking of Schrödinger’s cat. For the uninitiated, this is a thought experiment cooked up by the physicist Erwin Schrödinger in 1935. A cat is sealed in a box with a vial of poison rigged to a radioactive atom that may or may not decay. Until somebody opens the box and looks, quantum mechanics insists the atom is in both states at once, and therefore so is the cat: simultaneously alive and dead. Schrödinger intended it as a mockery of that interpretation, which makes it all the more fitting here. The AI sceptic keeps a similar cat in a similar box. Their AI is simultaneously too stupid to draft a coherent email and clever enough to enslave the species, and its state only collapses into one or the other depending on which rhetorical point the observer needs at that moment.

A meme appeared on social media this week that immediately triggered the writers in whose Facebook group it appeared.

“No matter how much we may tell ourselves that AI is just a tool like spell-check, it isn’t. When we use AI to flesh out ideas, we lose the most important part of the writing process: thinking.” Eve Fairbanks

The source of the meme appears to be an article from The Atlantic (The Biggest Tell That Something Was Written by AI, May 2026, https://www.theatlantic.com/technology/2026/05/how-to-tell-ai-writing/687345/).

It’s an elegant essay. It’s also wrong in about half a dozen distinct ways, and since I have what you might call a professional interest, let me take them in order.

The central claim is the one in the meme: the struggle of drafting, the backtracking and the dead ends, is where thinking happens, so outsourcing the draft means outsourcing the thought. This confuses one particular workflow with cognition itself. Plenty of people think by talking, by outlining, by arguing with someone across a table. A writer who spends an hour interrogating a model, rejecting its suggestions and sharpening a premise against its pushback has done a great deal of thinking. It just didn’t happen in solitary confinement with a blinking cursor, which is apparently the only venue Fairbanks recognises.

Then there’s the sycophancy charge, backed by a study showing AI models agree with their users 49 per cent more than humans do, and therefore cannot challenge your premises the way your inner critic can. Fair enough, as a description of default behaviour. But sycophancy is a setting, not a soul. Ask a model to be adversarial and it will happily savage your argument, which is more than can be said for most colleagues, who flatter and hedge for social reasons and cannot be reconfigured with a sentence.

The article’s showpiece is an interrogation of ChatGPT about why market delusions wear a “little” tech-futurist hat rather than a big one, which ends with the model rambling about raccoons at VC afterparties. This is offered as proof that AI has no reasoning underneath its word choices, only confabulation. Well, yes. So do humans. Ask any comedian to explain from first principles why the third word of a punchline is funnier than its synonym and you’ll get the same escalating nonsense, because jokes aren’t built that way. Psychology settled decades ago that people invent post-hoc justifications for their own choices without realising it. The demo was rigged, and when pressed the model agreed the metaphor had collapsed, which is precisely the self-correction the article claims it cannot perform. Fairbanks reads that agreement as servility, which is convenient, because it means her thesis wins whatever the model says. The cat, you’ll notice, is both alive and dead again.

The claim that AI text is uneditable because every element is “equally off” describes the output of a lazy two-sentence prompt, generalised to an entire technology. The worry about homogenised language, that we’re all starting to say “delve”, is real but ancient. Language has always converged on dominant media, from the King James Bible to corporate email-speak, and nobody wrote elegies for the sentences lost to “per my last email”. And the closing anecdote, in which an unsendable email accidentally saved the author from marrying the wrong man, credits friction with a judgement she made herself. Friction is sometimes wisdom. It is more often just friction.

What the article gets right is the economics. Competitive pressure rewards volume and polish over substance, and AI supercharges both. But that is an argument about incentives, not about the instrument. Trust attaches to the claims in a piece of writing and the person willing to stand behind them, which is why disclosure matters and always has. You’ll notice I disclose. It’s on the tin.

Because here is the boring, unglamorous truth that every one of these objections dances around, and the thing that finally lets the cat out of the box: AI is a tool. Like the word processor, the spell-checker and the printing press before it, the interesting question was never whether the tool exists but how it’s wielded. Hand it a two-sentence prompt and a deadline and you’ll get exactly the beige sludge the Atlantic’s editor keeps receiving, and she’ll be right to bin it. Bring it structure, argument, taste and the willingness to throw away most of what it offers, and you’ll get something worth reading. This is the oldest rule in computing and it has survived every revolution since the punch card: garbage in, garbage out. The machines haven’t changed that. Neither, I notice, have the people.