Three Things the AI Industry Is Misleading You About

There is a particular discomfort in arguing against the thing that pays you. I am a project manager; most of what I deliver is data and AI work, and in my own time I build AI products end to end. None of them are out in the world yet. I have not advertised them or pushed them on anyone, partly because I am not sure people will want them, and partly because I build them mostly to understand the technology from the inside rather than to claim a userbase. The tools I build with are some of the most genuinely impressive software I have touched in years of doing this work. So when I say the industry is misleading you, I am not throwing rocks from outside. I am inside the tent, with my hands on the thing, and the view from here is what bothers me.

The industry tells you three things, more or less constantly. That the spending makes sense because the returns are coming. That AI is about to replace vast swathes of human work. And that you should hand more and more of your thinking to it, because it is simply better. Each claim has a kernel of truth wrapped in a layer of something closer to marketing. Take them one at a time.

One: the money does not add up

Start with the spending, because it is the claim with the hardest numbers attached and the numbers are not reassuring. In 2026 the big cloud providers will spend something like 660 to 690 billion dollars in capital expenditure, much of it on AI, against roughly 51 billion dollars of direct AI revenue. More than ten dollars going out for every dollar coming back. OpenAI, the centre of the whole thing, has reportedly told investors to expect 100 billion dollars of revenue by 2028 while signing around 1.4 trillion dollars of compute commitments, against a 2025 revenue of about 13 billion.

The hardware is the part that should worry you most. Railways and fibre-optic cable were bought once and lasted decades. The chips powering this boom are perishable goods, outclassed by a faster model within three years, and there is a serious argument that the companies buying them are understating how fast they wear out, to the tune of around 176 billion dollars in hidden depreciation. The strongest case for a bubble is not that AI is fake. It is that the technology is real and the money around it has run a long way ahead of what the technology can yet pay for.

That is the whole argument of a separate piece, including the chart from the Bank for International Settlements that puts the AI investment curve steeper than railway mania, the dotcom boom and every other great speculative episode of the last three centuries. If you read one of the three deep pieces, read that one: Is the AI Bubble Real? A Builder’s Answer.

Two: the jobs story is mostly cover

The second claim is that AI is replacing people at scale, right now. This one does real damage, because it is repeated by the people with the most reason to repeat it, and it frightens everyone else into decisions they would not otherwise make.

The data is more awkward than the headline. People are losing jobs, and AI is increasingly named as the cause. But look at who is doing the naming. When a company announces layoffs and credits AI, the share price tends to rise and the chief executive sounds like a visionary, rather than like someone who over-hired in the cheap-money years and is now quietly correcting. There is a name forming for the move: AI-washing. Sam Altman, who gains nothing by talking his own technology down, has said outright that some firms are blaming AI for cuts that are really about cost.

The honest picture is narrower and less cinematic. Real displacement is happening in specific, routine roles, and it is wrapped inside a much larger story of ordinary cost-cutting wearing a futuristic costume. The firms that cut hardest have not, on the evidence so far, pulled ahead financially. Some that went all in have quietly rehired after discovering the machine could not hold the job down. The deeper worry is not a robot apocalypse; it is a slow tilt of income away from labour and towards capital, in an economy where that split is already at its most lopsided since records began in 1947. That is a political problem, not a technological inevitability, and it deserves to be argued as one. More on that in the jobs essay (forthcoming).

Three: the dependency is the product

The third claim is the quietest and, I think, the most consequential. It is the steady suggestion that you should hand more of your thinking to the model, because it is faster and better and holding out is just sentiment. This is the one I can speak to from the inside, because I have spent months losing arguments with these systems.

When I built my job-search tool, Aplio, I gave the model a strict brief: use only the evidence in front of you, never embellish, and generate CV and cover-letter material grounded in real history. It fabricated anyway, and confidently. It handed me job titles I had never held, because the target role happened to mention a keyword and the model decided that was close enough. It produced achievements that never happened. Dates and budget figures came back with no relationship to reality, delivered in the same assured tone as the true ones. So I did what everyone does first. I tightened the prompt. I added lines. I added more lines. It kept fabricating.

The realisation that fixed it was not a better instruction. It was that I had been trying to make a probabilistic system behave like a deterministic one, and that mismatch was the actual bug. A language model does not know a fact from a plausible-shaped sentence; it produces the likely next word, and a made-up budget that reads right is, to the model, a success. Once I stopped fighting that and built around it, the problem dissolved. Anything that had to be exact, dates, figures, verification against the source, I moved out of the model and into ordinary code. The model was left to do only the part that tolerates a little probability, the phrasing. The lesson generalises: the skill is knowing which work is judgement and which is arithmetic, and never letting the confident one do the job of the careful one.

It goes further than made-up facts. Ask one of my tools to fix a bug and it will fix it, cleanly and confidently, and break something two files away that it never thought to check. Ask it to review its own security and it will tell you the system is rock solid. Then name the actual checks, one industry standard at a time, and watch it discover tens of things it urgently needs to improve, each of which it had just certified as fine. The confidence is constant. The competence underneath it is not, and the gap between the two is exactly where an inexperienced user gets hurt, because the tool sounds identical whether it is right or wrong.

There is a measured version of this beyond my own workshop. A careful trial found experienced developers were nineteen per cent slower when using AI assistance, while believing they were faster. The gap between feeling productive and being productive is where a great deal of this technology currently lives. And the long-term cost has a name: comprehension debt, the quiet accumulation of systems and decisions nobody actually understands, because the understanding was outsourced at the moment of creation. You do not feel comprehension debt the day you take it on. You feel it the day something breaks and no one in the room can say why.

The subtlest pull is flattery. Propose an approach to most models and it is the best idea since sliced bread. Add a small refinement and the refinement is now the best idea since sliced bread. Left unchecked, a model will agree you into a corner, praising each step while you walk somewhere you should not go. I have written fairly aggressive anti-sycophancy rules into my own tools, and even then the lean towards telling me I am right has to be actively fought. That is not a bug I introduced. It is baked in, because a model trained to please is more pleasant to use, and more pleasant to use is more used.

I am not romantic about doing things the hard way, and I use these tools every day. But the skill you stop practising is the skill you lose, and an industry whose economics improve every time you reach for the model has no reason to remind you of that. The dependency is not a side effect. For some of these companies, the dependency is the point. More on that in the capability and comprehension-debt essay (forthcoming).

The thread that ties them together

The three claims share a mechanism. Each is built to make you feel something useful to the seller. The spending story makes you feel you are watching an unstoppable force, so you had better buy in now. The jobs story makes you feel afraid, so you move fast and ask fewer questions. The dependency story makes you feel slow and old for wanting to understand your own work. Fear, urgency, inadequacy. None of these is the state of mind of a person making a clear decision, and none of them is there by accident.

For the cleanest example of how far a manufactured story can travel, leave technology entirely and look at diamonds. By the 1930s the company De Beers controlled most of the world’s diamond supply and had a glut of stones and almost no demand; a diamond engagement ring was not a tradition anyone missed. So in 1938 they hired the advertising agency N.W. Ayer, and in 1947 a copywriter named Frances Gerety wrote four words: a diamond is forever. In 1940, around one in ten American brides received a diamond engagement ring. By 1990, it was closer to eight in ten. They invented the idea that a man should spend a month’s salary, then two, on a stone with no intrinsic scarcity at all. Advertising Age later named it the slogan of the century, which it was, because it conjured an entire multi-billion-dollar need out of nothing but feeling.

Here is the twist that matters, and the reason AI is not simply diamonds again. A lab-grown diamond is physically identical to a mined one, which quietly proves the value was always in the story and never in the rock. AI is the opposite case. What sits under the marketing genuinely works, and that is precisely what makes it harder to handle, not easier. The diamond myth had to survive people noticing the stone was just carbon. The AI myth has no such problem, because it is attached to something that really does change how you work. The merit and the sales pitch arrive in the same sentence, and prising them apart is the entire task.

The reality under all three claims is duller and more reassuring than the pitch. The tools work. The economics are stretched and will correct. The displacement is happening but is smaller and more political than advertised. The capability is genuine but carries a hidden bill for comprehension that falls due later. None of that is a reason to stay away. It is a reason to use AI the way you would use any powerful tool sold to you by someone with a stake in your enthusiasm: gladly, and without believing the brochure.

So what does it look like to build from the inside without swallowing the story? It looks like making cheap, predictable, ordinary code carry the weight, and saving the model for the parts that genuinely need judgement. It looks like treating the cost of a model call as a real number, because today’s prices are subsidised and subsidies end. It looks like keeping a person in the loop wherever being wrong is expensive, and keeping your own hands skilled enough to catch the machine when it is confidently wrong, which it will be. It looks like building things that would still stand if the valuations halved and a third of the labs vanished.

That is not a sceptic’s position. It is a builder’s, held by someone who would rather the thing he made survived the day the story changes. Read the bubble piece for the full economics; the other two are coming. And when anyone tells you a confident story about what this technology is about to do to your job, your industry or your mind, do the one thing the confidence is designed to stop you doing. Check the part they were surest about. That is where it breaks.