Chapter 10 of 11 · 9 min read
The Machine Learns to Talk
From The Market for Your Mind by Wrotebook
At the end of November 2022 a research company in San Francisco released a chatbot to the public, describing it modestly as a research preview. You typed a question in ordinary language, and it answered in ordinary language: not a list of links, not ten blue lines to choose between, but paragraphs, in complete sentences, in a tone that was patient, articulate and faintly eager to please. It could explain a tax return, draft a letter of complaint, suggest a recipe for what was left in the fridge and then write a poem about the recipe. ChatGPT found an enormous audience faster than almost any consumer product before it, and within a couple of years asking a machine a question and receiving an answer in prose had become, for hundreds of millions of people, an ordinary thing to do.
The search companies understood at once what was at stake, because they had spent twenty years selling the moment at which someone typed what they wanted into a box. If that moment moved from the search box to the chat window, the money would move with it. In May 2024 Google began showing what it called AI Overviews at the top of its results in the United States, short machine-written answers that summarised what the web had to say before the reader reached the links. The answer began to replace the list, and the effect on what people did next showed up quickly. In July 2025 the Pew Research Center published an analysis of the browsing of 900 American adults during March of that year. When a Google results page carried an AI summary, people clicked through to an ordinary result on about 8 per cent of visits; when it did not, they did so on about 15 per cent. The links cited inside the summaries were clicked on about 1 per cent of visits. A separate measurement by the research firm SparkToro found that in the first four months of 2026 about 68 per cent of Google searches in the United States ended without a click to anywhere at all, though its comparable figure for 2024, about 60 per cent, came from a different panel, so the two are not a clean comparison.
Neither study shows that the web is dying, and both point the same way. More of the answer is given on the page or in the chat, and fewer people make the trip to the places the answer was assembled from. That matters, because the answers are made of other people's work. They are built from what journalists, reviewers, hobbyists explaining how to bleed a radiator, small firms describing what they do and volunteers editing encyclopedias wrote down, and for two decades that work was paid for, directly or not, by the visits it attracted from search. When the answer arrives without the visit, the visit and the money attached to it do not happen, and the people who made the answer possible are invited to go on supplying it for less. Some large publishers have signed licensing agreements with the companies building these systems. A great many smaller ones have simply watched their traffic decline. It is the familiar pattern at one further remove: the intermediary that gathers the attention keeps more of the value, and the people who produced what the attention was for keep less.
Now try it from the inside. Suppose the boiler that was fixed one January morning in Leeds, the one the engineer said would not see many more winters, has finally given up, and you are the person in the house deciding what to replace it with. One evening you open a chatbot on your phone and begin to explain. You tell it what kind of house you live in and when it was built. You tell it roughly what you can afford, that the running costs worry you, that your partner thinks you should wait a year, that you have read contradictory things about heat pumps and do not know whom to believe. You ask what it would do. It answers thoughtfully and at length, with advantages and drawbacks, a rough sense of the costs, and a few questions back to you, which you answer too. Twenty minutes later you have had a more useful conversation about heating than you have ever had with a human being, and you have told a machine more about your house, your money, your worries and your household than you would tell most of your friends.
The first chapter of this book began in a room where eight women were paid in coffee and cash to talk about margarine, while people on the other side of a mirror wrote down what they could not quite say and decided afterwards what it meant. The chatbot needs no mirror, no moderator and no envelope. You come to it of your own accord, unpaid, and explain your hopes and anxieties in detail and in your own words, because that is how you get a good answer. Every focus group ever convened, taken together, has not gathered as much, as intimately, about what people want and why. The notes are still kept on the far side of the glass. The difference is that the people being studied now write the notes themselves, at length, in full sentences, and hand them over with thanks. What is gathered in a conversation like yours is the most valuable form of attention the market has yet found: not intent expressed in four words in a search box, but intent explained, with the reasons attached.
For the first few years, all this was not sold in the old way. The chatbot companies earned money from subscriptions and from selling access to their systems to businesses, and their free versions were paid for by investors betting on the future. But the machines are expensive to run, and the history of this market does not contain many free services that stayed free of advertising for long. On 16 January 2026 OpenAI announced that it would test advertisements in ChatGPT for logged-in adult users in the United States on its free tier and its low-cost Go plan. Testing began on 9 February. The advertisements were labelled as sponsored, appeared beneath the chatbot's answers rather than inside them, and were matched to the subject of the conversation. Subscribers to the paid tiers would not see them. The test spread quickly. Advertisements were piloted in Britain from June 2026 and formally launched there in August, and by late October they were running in about 89 countries, still only on the free and Go tiers. OpenAI said it had more than 100,000 advertisers and that its advertising revenue was running at about $1 billion a year, both of them the company's own figures, which in this history is a familiar kind of figure to be offered.
There was a brief attempt to go further. In September 2025 OpenAI launched Instant Checkout, which let people buy certain products without leaving the conversation, and for a few months it looked as though the assistant would not only recommend things but buy them on your behalf. In March 2026 the company scaled the feature back, saying the first version had lacked the flexibility it wanted, and moved the purchase itself to merchants' own websites while keeping the browsing and comparing inside the chat. Only a few dozen merchants at most had gone live. The episode is a useful corrective to some of the more breathless predictions about machines that shop for us. For now the assistants mostly read, compare and recommend, and the attention they gather is being priced in the oldest way there is, with an advertisement placed next to the answer.
The companies running these assistants have every reason to keep the advertisements well away from the answers, and say that they will, since an assistant suspected of quietly selling to you is worthless as an assistant. The promise is sincere and the incentive behind it is real. It is also the promise every earlier medium in this history made at the start. Paid search listings began in a shaded box down the side of the page, plainly apart from the results, and the shading faded by degrees over the following years, each change defensible on its own and none of them announced as a change of principle. Feeds began in chronological order. Notifications began as a convenience. None of the drift was planned in advance, and all of it was the work of reasonable people responding to numbers, which is exactly why a promise about where the line will sit in ten years is less informative than a look at who will be measuring the results of moving it, and what they will be paid to find.
Here is your boiler conversation again, as a receipt, with notes in the margin:
You: Our boiler's sixteen years old and finally packed in. We're worried about running costs and my partner thinks we should wait a year. What would you do? [Given: the age of the heating, your anxiety about money, a disagreement at home.]
Assistant: That's a sensible thing to weigh up. Here are the main options… (four careful paragraphs) [Paid out: a clear, patient explanation that would once have cost an evening or a consultant's fee.]
You: We can probably stretch to about four thousand. Is a heat pump realistic? [Given: your budget, to the nearest thousand pounds.]
Assistant: At that budget, the realistic choices are… (three more paragraphs) [Paid out: options you would have spent a weekend finding.]
Sponsored: A local installer, with a discount on new boilers this month. [Sold: a moment of trust from someone who has just been helped, matched to everything above.]
[Measured: conversations, messages, return visits, sponsored taps. Left behind: a decision made easier, and a detailed record of what you wanted and why.]
The paid-out column of that receipt is genuine. A good assistant can save people hours, explain things patiently that no busy professional has time to explain, and give people who have never had access to expert advice something close to it. That is exactly why the market has arrived. Attention follows usefulness, and money follows attention, and the more useful the thing, the more there is to sell.
The taken-in column is new in kind. Television took evenings, the web took clicks, search took intent in a few words, the feed took time and content and data, and the phone took the day, one interruption at a time. The talking machine takes the reasons behind what people want, offered voluntarily, in conversation, to something that sounds as if it is on their side. And it combines in one place three roles that every earlier medium kept at least partly separate. It is the meter, because it knows more precisely than anything before it what each person is looking for. It is the shop window, because it chooses which options to describe and in what order. And it is the trusted voice, because it speaks in the patient, even tone of help. A banner is plainly an advertisement. A search result marked as an advertisement is, at least in principle, visibly apart from the rest. A sponsored line beneath an answer from a machine that has just spent twenty minutes being thoughtful about your boiler is read in the warm light of the conversation above it.
It also arrives with a choice that no earlier medium put quite so plainly. The paid tiers carry no advertisements; the free ones do. For the first time the terms of the attention market are printed on the price list: you can pay for the answer with money, or you can pay for it with your attention and the conversation that produced it. Most people, as with every free service before it, will choose the second, often without noticing that there was a choice to make.
After the conversation you put the phone down, go downstairs and explain to your partner, at some length and rather persuasively, what you think you should do about the boiler, and what the realistic options are at your budget, and why the running costs need not be as frightening as they seemed, and somewhere in the middle of the part about the running costs you hear yourself, fluent and reasonable and entirely convinced, and wonder, before the installer with the discount rings back: whose reasons are these?