A meeting. One of the executives says: “The neural network read the contract and decided there are no risks.” Everyone nods. I’m stuck on the word “decided.” The neural network didn’t read anything and didn’t decide anything. It calculated which words most often follow words like “contract,” “risks,” and “absent.” But our language already has verbs for it from the human register: it thinks, it understands, it makes mistakes, it fantasizes. Nearly every disappointment in AI starts with those verbs. Let’s chase them out of our heads and honestly sort out what AI is not.

It’s not intelligence

Under the hood of any large language model there is one and the same activity: predict which word is most likely to come next. Billions of times per response. Out of this activity grew a skill that amazes even me, and I’ve been looking at these systems professionally since 2017: the model reasons brilliantly. It will explain why the sky is blue, build a chain of logical steps, give examples. It will explain with the same brilliance why a blue sky is bad for the economy — if you ask politely.

There’s no understanding behind that brilliance. There’s statistics, which from the outside is nearly indistinguishable from understanding. The test is simple: ask the model if it’s sure. It will gladly apologize and rewrite the answer to the opposite — with the same elegance, the same examples. Ask again — it will agree with that one too. Ask it to explain why two plus two is five, and you’ll get a beautiful lecture about the fifth dimension. Engineers call these systems stochastic parrots — parrots because the words aren’t theirs, stochastic because each word choice is a bit of a lottery. The term is unkind but accurate, and it helps: you stop expecting from a tool what was never put into it.

If you want to dig deeper — the paper that coined the term is freely available: On the Dangers of Stochastic Parrots.

Neither artificial nor real

The name “artificial intelligence” appeared at a conference in Dartmouth in 1956 — and has been doing double duty ever since. The word “artificial” promises: look, a made-by-human mind. But there’s no mind inside, artificial or otherwise — there’s mathematics. So “artificial” is a lie. And “not real” is also a lie, in its own way: the texts a model produces are real, alive, some of them good. Only the meaning in them isn’t real — it isn’t anyone’s, not hers, not yours.

This is not a creature and not a trick. It’s a machine. I like comparing it to the steam engine: that one turned the labor of hands into the labor of steam; this one turns hours of work with text into minutes. Nothing mystical. Nothing human. Immensely useful — and with completely different rules of handling than a colleague.

Useful but unreliable

Both halves matter, and the order matters. Useful: drafts, summaries, options, translations, first sketches where a blank page used to stand. This genuinely takes hours off the work — I see it in every project. Unreliable: facts, figures, links, quotes — the model delivers all of it in the same even voice, with the same degree of confidence in either case. The beautiful-and-true sounds exactly like the beautiful-and-invented.

Hence the main conclusion I carry into every rollout: reliability is not a property of the model. It’s a property of the process around it. A mixer beats dough beautifully, but it doesn’t choose the recipe and isn’t responsible for the salt. A model writes beautifully, but it doesn’t check facts — we do. In detail, how control degrades into a rubber stamp when people miss this, I wrote in the piece on personal AI assistant risks.

And it’s not going anywhere

One hope remained: wait for the fashion to pass. It won’t. We’ve been through this with electricity, with the internet, with email: every such machine first scared people, then got banned, then became the background of life — for those who learned to use it. Banning neural networks in the office today is like banning spreadsheets in accounting in the nineties: the ban exists, and so do the spreadsheets.

So the only grown-up path left is learning to coexist. Like with a car: it’s not evil and not kind, but in clumsy hands it’s dangerous, and in capable hands it multiplies a person’s strength by orders of magnitude. Learning to drive this machine for words means three things: understand its boundaries (this article is the first step), build a habit of checking the result, and teach the people around you. The ten most common mistakes in that teaching I broke down in another piece — on employees’ AI mistakes. And if you want the short path — come by, let’s look at your case together: from the first step to a working setup.

In short

AI is neither a mind nor a trick. It’s a machine for words: wonderful and heedless. It does the wonderful part on its own. We answer for the vigilance.

What to read