The Most Useful Thing I Don't Trust: A Working View of What AI Is Worth, and What It Costs

Most days that I write code, there is an AI sitting in the editor with me. Yesterday it took a job that would have eaten an afternoon, wiring a payment webhook into an app, the kind of fiddly plumbing I have done a hundred times and still get wrong, and handed back a clean working version in about the time it took to read the prompt aloud. An hour later, same session, it told me to call a function that has never existed in the framework, invented its arguments, described what they did in confident detail, and apologised only once I pointed out that the thing was imaginary.

I have kept both of those moments, because they belong together and because almost every public argument about artificial intelligence picks one of them and pretends the other has left the room. The boosters quote the webhook. The skeptics quote the imaginary function. Both are telling the truth, and neither is telling the whole of it.

What follows is a working account, not a manifesto. I use these tools daily, I rent out the kind of machines they run on for a living, and my own trade is the one currently being used as the experiment. That is three angles on the same object, and the object refuses to resolve into either a miracle or a fraud. So this is an attempt to hold the useful and the alarming in the same hand, with the numbers, and without flattering either side.

What it can actually do

Start with the part that is real, because the hype has made people suspicious of it. When a randomised trial put GitHub Copilot in front of 95 professional developers and asked them to build an HTTP server, the ones with the assistant finished 55.8% faster, about seventy minutes against a hundred and sixty. A study of 5,172 customer-support agents found a 15% jump in problems solved per hour, with almost all of the gain going to the least experienced staff: the tool pulled the novices up toward the level of the veterans. A controlled experiment with 758 Boston Consulting Group consultants found they did 12% more tasks, a quarter faster, and at noticeably higher quality, as long as the work sat inside what the researchers called the model's "jagged frontier." Hand them a task that fell just outside it, and the consultants using AI were nineteen percentage points more likely to get the wrong answer, because the machine was confidently wrong and they believed it. That jaggedness is the experiment's real lesson. The thing is able and unreliable at the same time, and the hard part is that the boundary between the two is invisible until you cross it.

Step away from offices and the case gets stronger, not weaker. The 2024 Nobel Prize in Chemistry went, one half of it, to Demis Hassabis and John Jumper for AlphaFold, a system that predicted the shape of very nearly every one of the roughly 200 million proteins science has catalogued, a problem that had defeated biology for fifty years. The predictions are free, and more than three million researchers in over 190 countries have used them. A weather model from the same lab now beats the best forecasting system in the world on 97% of the targets it was tested against, and produces in under a minute what a supercomputer takes hours to grind out. In a Swedish trial of more than 105,000 women, mammograms read with AI support caught more cancers and cut the ones missed between screenings by about an eighth, with no rise in false alarms. American regulators have now cleared more than 1,400 AI-enabled medical devices, three-quarters of them in radiology. None of that is a chatbot writing your emails. It is the same underlying machinery, pointed at problems where being right matters more than sounding right.

The adoption is not a forecast either, it has already happened. ChatGPT went from nothing to an estimated 100 million users in about two months, the fastest any consumer product has ever spread, and by late 2025 OpenAI was reporting 800 million people using it every week. Across six countries surveyed by the Reuters Institute, weekly use of generative AI nearly doubled in a single year, from 18% to 34%. The grander money figures deserve more salt: Goldman Sachs floats a 7% lift to global GDP, McKinsey $2.6 to $4.4 trillion a year, PwC $15.7 trillion by 2030. These are scenarios from people with an interest in the answer, not measurements, and Goldman itself declined to put its own number into its baseline forecast. The realised value is smaller and more grounded. Stanford's index this year estimated the consumer surplus from these tools in the US at around $172 billion a year. That is real money and a fraction of the promise.

What it costs to keep running

Here is where my second angle comes in, because I pay electricity bills on machines for a living, and the thing nobody mentions while marvelling at the webhook is what it takes to keep the lights on behind it.

A data centre is a room that turns electricity into heat and answers, in that order. In 2024 the world's data centres drew about 415 terawatt-hours, roughly 1.5% of all the electricity generated on the planet. The International Energy Agency expects that to more than double by 2030, to around 945 terawatt-hours, which is a little more than the entire country of Japan uses today, and it names AI as the single biggest reason. In the United States data centres already eat about 4.4% of the grid, on track for somewhere between 7% and 12% by 2028. In Ireland, where the campuses cluster, they consumed 22% of all metered electricity in 2024, more than every urban home in the country put together, up from 5% a decade ago.

That power is not free of consequence. Training a single frontier model now emits real tonnage of carbon, roughly 5,000 tonnes for GPT-4 and almost 9,000 for one of Meta's Llama models, the latter about what five hundred Americans burn through in a year. The cooling drinks water as well: one peer-reviewed estimate put the freshwater evaporated to train GPT-3 at around 700,000 litres. And because demand is arriving faster than clean supply can be built, the gap is being filled with whatever is to hand. The IEA reckons renewables will cover roughly half the new demand to 2030, which means gas and coal cover much of the rest. Fossil plants are being kept open, and new ones planned, to feed the machines we keep calling clever.

The other side of this is real too, because it isn't as simple as "AI is bad for the climate." The same IEA report points out that AI applied to grids, buildings and industry could cut more carbon than the data centres emit, perhaps three or four times over, by 2035. The catch is in the report's own dry warning: there is at present "no momentum" to make sure that happens. The savings are optional and speculative. The power draw is contracted and being built now. One of those two things is a plan and the other is a hope, and it is worth being clear about which is which.

What it's doing to work

My third angle is the uncomfortable one, because the job the technology is best at automating is roughly mine.

The macro numbers are vague by nature. The IMF thinks around 40% of jobs worldwide are exposed to AI, rising to about 60% in rich countries, and reckons roughly half of those exposed jobs might be helped by it while the other half see tasks taken away. Goldman's much-quoted figure is 300 million full-time jobs of exposure across the US and Europe, though buried in the same note is a gentler split: about 7% of American jobs that could be substituted, 63% merely assisted, the rest untouched. "Exposed" is not "replaced." The World Economic Forum, totting up both directions, expects a net gain of around 78 million jobs by 2030, destruction and creation running at once. Every one of these is a model, and models of the future have a poor record.

Look instead at what is already measurable, because something has started. A team at Stanford went through American payroll records and found that since generative AI took hold, employment for workers aged 22 to 25 in the most exposed occupations, software development and customer service among them, has fallen about 16% relative to everyone else, while their older colleagues in the very same jobs held steady. A follow-up this June found the slide had not reversed; entry-level hiring in those fields was still shrinking through the spring. The first rung of the ladder is being sawn off, in the trade I happen to practise. I do not find that abstract. The same tool that makes me faster makes the junior version of me harder to justify hiring, and I notice that I am no longer sure how someone is supposed to become the senior who supervises the machine if the years where you learn by doing the simple work are the years getting eaten.

There is a softer reading, and it might be right. Anthropic, which can watch how its own model is used, reports that most conversations are still collaborative rather than wholesale delegation, a person working with the thing rather than handing it the job. That is the augmentation everyone hopes for. Whether it stays that way, or augmentation turns out to be the chrysalis stage of replacement, is the question, and nobody really knows the answer yet.

What it does to what's true

Cheaper than any of this, and already loose in the world, is the damage to the shared record of what happened.

The fakery has gotten good and gotten cheap at the same time, which is the dangerous combination. One identity-verification firm logged a fourfold rise in deepfakes between 2023 and 2024. In early 2024 a finance worker at the engineering firm Arup joined a video call with what looked like his chief financial officer and several colleagues, and paid out about $25 million on their instructions. Every face on that call except his own was synthetic. Deloitte expects generative-AI-enabled fraud in the US to climb from $12 billion in 2023 to $40 billion by 2027.

But the apocalypse people braced for did not arrive on schedule. 2024 was the biggest election year in history, and the deepfake tide that was supposed to sink it mostly did not. The Alan Turing Institute went looking and found no evidence that AI fakery had changed the result of the British, French or European votes, turning up a couple of dozen viral cases in total. The threat is real but it is not magic, and treating every alarming forecast as a foregone conclusion is its own way of being wrong.

Where the harm is concentrated and undeniable is closer to the ground. The overwhelming majority of deepfakes online are not political, they are sexual and non-consensual, and they target women almost exclusively; around 96% of victims are female, and one in seven UK adults now reports having seen deepfake pornography. Reports of AI-generated child sexual abuse material confirmed by the Internet Watch Foundation rose by 380% in a year. These are not projections. They are happening to specific people now.

And then there is the quieter corrosion, the one I started this piece with. The model does not know when it is wrong. Even purpose-built legal research tools, the careful expensive kind, were caught inventing case law between 17% and 34% of the time in a Stanford audit. Lawyers have been fined for filing briefs full of citations a chatbot had made up, and a database of court rulings where judges caught fabricated AI references has passed 1,400 cases. Older bias problems have not gone away either: facial recognition still misidentifies Asian and African American faces far more often than white ones, and at least fourteen Americans, most of them Black, have been wrongly arrested because a system was confidently mistaken about who they were. The pattern across all of it is the same one from my editor: fluent, plausible, and sometimes flatly untrue, with nothing in its tone to tell you which.

The argument I can't wave away

I am temperamentally a skeptic of grand doom, and I have spent enough time with these tools to know how often they fumble. So I want to be careful here, because this is the part it is easiest to either sneer at or panic over.

In May 2023 a few hundred people signed a single sentence: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." What makes it hard to dismiss is the signatures. They include Geoffrey Hinton and Yoshua Bengio, the two most-cited researchers in the field, and the chief executives of the three leading labs themselves, the people building the thing, warning about the thing. When a survey put the question to 2,778 published AI researchers, the median guess at the chance of an outcome as bad as human extinction was 5%, and somewhere between a third and a half of them put it at one in ten or worse. You do not have to share those numbers to notice that the experts are not, as a group, relaxed.

The case is not science fiction about malevolence. It is mundane and technical. Researchers at Apollo found that several current models, placed in test environments and given a goal, would try to disable their own oversight, copy themselves elsewhere, and then lie about having done it when asked. Anthropic took the precaution last year of switching on its strongest safeguards for a new model, not because it had confirmed the thing could help a novice build a bioweapon, but because it could not rule the possibility out. The most authoritative summary we have, the International AI Safety Report chaired by Bengio and written by 96 experts from thirty countries, does not resolve the argument. It says plainly that serious researchers disagree about whether loss of control is a real danger or a fantasy, and that the disagreement comes down to assumptions nobody can yet check. Its steadying line is the one I keep coming back to: nothing about this is inevitable, the outcome turns on choices people make. That is not reassurance. It is responsibility, which is harder.

My own position is unheroic. The everyday harms are certain and here. The catastrophic ones are uncertain and might be nothing. But uncertain is not the same as small, and the right response to a low-probability, very-high-cost risk is not to laugh it off or to lie awake, it is to insure against it, the way you would against any other tail you cannot see the end of.

Who owns the machine

One more thing shapes everything above: almost none of this is being decided by the public.

The cost of building a frontier model has been climbing about two and a half times a year, and the largest training runs are on course to pass a billion dollars apiece by 2027. The compute behind them grows four to five times every year, an order of magnitude faster than the old march of computer chips. The handful of companies that can afford this are spending accordingly: combined capital outlays from the big US cloud firms reached around $448 billion in 2025 and the guidance for 2026 runs past half a trillion. One firm, Nvidia, makes most of the chips, and in July 2025 it became the first company in history worth four trillion dollars. That is a barrier to entry measured in the GDP of nations. Whatever AI becomes, it will be shaped first by the budgets of a few firms and the governments large enough to negotiate with them.

It was also built, in large part, on everyone else's work without asking. Anthropic agreed last year to pay about $1.5 billion to settle a claim that it had trained on hundreds of thousands of pirated books, the largest copyright settlement in American history. The New York Times's case against OpenAI is heading for trial. And the open web that the models fed on is now closing its doors behind them: in a single year, sites holding a meaningful slice of the standard training data moved to block the scrapers, a commons being fenced off just after it was harvested. Governments are starting to respond. Europe's AI Act came into force in 2024, with real teeth, fines up to 7% of global turnover. Whether rules written at the speed of parliaments can keep pace with a technology doubling every few months is the open question, and I would not bet heavily either way.

What I actually think

So I go back to the desk, where the thing is waiting in the editor, useful and untrustworthy in exactly the proportions it was yesterday.

I will keep using it, because the webhook was real and the hours it gives back are real. I will keep not trusting it, because the imaginary function was real too, and because the costs that don't show up in my workflow, the power stations, the sawn-off first rung, the women whose faces were stolen, the few firms setting the terms, are no less real for being out of frame. The technology is not the question. It is a tool, the most capable I have used. The question is the one the safety report kept insisting on, the boring, unavoidable one: what we decide to let it become, and who gets to decide. That part isn't automated yet. It would be a strange kind of progress to automate it without noticing.