The Dark Side

Of all the costs these are the least in your hands, and that is exactly why they need seeing, you cannot weigh a thing you refuse to look at. Back in the first series I said the anger at AI was worth taking seriously rather than waving off. This is where I make good on that, without the doom merchant's exaggeration or the booster's shrug. Some of what follows I could not soften if I wanted to. Some of it is smaller than the scary version you have heard, and I will say so, because getting this wrong in the frightening direction is still getting it wrong.

The art it was built on

Start where the fault line is deepest, because you met it in the first series. These models learned from a staggering amount of human work, text, art, photographs, music, most of it pulled off the internet without asking the people who made it and without paying them a cent. That is not a rumour or a fringe complaint. It is how the technology was built, and the makers have largely conceded the "how," they are now fighting over whether it was legal.

The courts are the clearest window into how unresolved this is. In 2025 one AI company, Anthropic, agreed to pay one and a half billion dollars to settle a case brought by authors, the largest payout of its kind in the history of copyright, roughly three thousand dollars for each of around half a million books. That sounds like a reckoning, but read the fine print. The settlement was for pirating the books, downloading them from illegal libraries, not for training on them. On the training question itself, the early rulings have mostly gone the other way, with judges calling it transformative enough to count as fair use, and no higher court has settled it. So far the pattern is early rulings that mostly help the companies on the core question and punish them only for how they got the material.

The promised fixes mostly never arrived either, the tool for opting out that was supposed to let creators pull their work out of training largely never shipped. So the practical situation for a working artist is this, your work may well be in there, you were not asked, you were not paid, taking it back out is difficult or impossible, and the model trained partly on your style is now competing for your job. Whatever a court eventually decides about the legality, you can see why that feels like theft to the people it happened to. There is a cleaner path, one image tool built its model only on work it had the rights to and made that its whole pitch, which proves the version built on consent was possible. Most of the industry just did not choose it.

Your data

Now turn the lens on yourself, because you are feeding these systems too. By default, the big consumer chatbots now use your conversations to train their next models unless you turn that off, and the retention windows have stretched, in one case to five years for anyone who allows training. Whatever you type into that friendly box, the plans you talk through, the draft you paste, the thing you would never say out loud, is in many cases not a private exchange, it is raw material, unless you found the setting.

And "delete" turns out to be a soft promise. During one lawsuit, a court ordered a chatbot company to preserve users' conversations, including ones people had deleted, overriding its own deletion policy. The order was later lifted, but the lesson stands, once your words are on someone else's servers, whether they truly disappear is not fully up to you.

Zoom out and it gets starker. One facial recognition company built a searchable database of billions of faces by scraping them off the public internet, yours very possibly among them, and was fined tens of millions in Europe for it, which it has largely just refused to pay. That last detail is the real story of AI privacy, the gap between the rules and the reach. Europe has grown some genuine teeth, a broad AI law now phasing in, real fines under its privacy regime. The United States still has no comprehensive privacy law at all, a patchwork of state rules instead, and the current federal push is aimed at overriding even those. So how protected you are depends heavily on where you live, and almost nowhere does it depend on your own consent as much as it should.

A commons full of fakes

The next cost is one everyone pays, whether or not they use AI, because it is a cost to the shared floor of what is real.

The tool that turns words into a picture, and now into a voice and a video, made convincing fakes cheap and fast, and the results are already ugly. A finance worker in Hong Kong paid out about 25 million dollars after joining a video call where every colleague on screen, including the chief financial officer, was a deepfake. Fake robocalls in a real election used a cloned presidential voice to tell people not to vote. And the single most common use of this technology is bleaker than any of that, by the best available count the overwhelming majority of deepfake videos online are pornography made of real people without their consent, almost entirely women, from ordinary people to famous ones, made to be humiliated. That is not a side effect. By volume it is the main thing the tools for fake images are used for.

The defence you would hope for, that people can just tell, does not hold. When researchers pool the studies, humans spot a deepfake at barely better than a coin flip. And underneath the outright fakes is a quieter flood, the internet filling with cheap filler made by machines, so much of it that "slop" became a word of the year. The tools meant to prove what is real, invisible watermarks and content labels, exist and are spreading, but most can be stripped by something as simple as a screenshot. The net effect is a world where seeing is no longer believing, and that erosion lands on everyone, including the people who never touched the tool.

The physical bill

Last, the cost that is literally physical, and here I have to correct the story in the other direction. You have probably heard that a single question to a chatbot burns wild amounts of energy or drinks a bottle of water. Those figures were overblown, and in the most famous case the researcher behind the water number has walked it right back. One prompt is small, a fraction of a watt hour, a few drops of water. Feeling guilty per question is not where this cost lives, and I would rather tell you that than scare you with a figure I know is wrong.

Where it does live is the total. The data centres behind all this are on track to roughly double their electricity use by 2030, heading toward a few percent of the world's entire supply, and their demand jumped sharply in just the last year. The companies building them have watched their own carbon emissions climb well above their targets, Google's are now far above where they were a few years ago, even as the promises to cut keep coming. Individual centres can drink millions of gallons of water a day, and a real share of them sit in places already short of it. And all that capacity is concentrated in a very small number of enormous companies spending hundreds of billions to own the infrastructure the rest of us increasingly depend on. The cost per prompt is trivial. The scale, and who controls it, is not.

Holding it

That is the dark side, and I did not enjoy writing some of it. Art taken without consent, privacy that depends on a hidden setting and stops at a server you cannot see, a shared reality getting harder to trust, and a vast physical machine in very few hands. These are not growing pains that a better model fixes next year. They are real costs, some falling on people who never signed up for any of it.

Naming the cost honestly is not the same as concluding the thing is all cost. I gave the upside its full weight in the earlier pieces, and I am giving this its full weight now, and the hard, adult move is to carry both at once without collapsing into the true believer or the cynic. There is one stranger thing to see before we reach your call, though, and it is the one place the technology stops being something you pick up or something aimed at you, and starts being something that acts on its own.