The Thing You Confide In Now Has a Sponsor

In 2024 Sam Altman called the combination of advertising and AI “uniquely unsettling” and said ads would be a last resort. By the fall of 2026 ChatGPT was serving ads in 89 countries, more than 100,000 advertisers had bought them, and users in the United States could click through to a conversation with a brand’s own agent. Nothing about the product looks different. What changed is who else is in the room.

The rollout took eight months. Ads arrived for American users on the free and Go tiers in February 2026, at a reported 60 dollars per thousand views and a 200,000 dollar minimum. By May the minimum was gone and anyone with a credit card could buy. Britain, Japan, Brazil, and 31 European markets followed over the summer. Trade press put the business at roughly a one billion dollar annual run rate by the end of August. In October OpenAI added 26 more countries and announced the total.

The mechanics are worth knowing. Advertisers do not pick keywords. They write “context hints” describing the kind of conversation they want to appear in, and the system matches the ad to where the person is in what they are talking about. Targeting based on the current chat is always on. Targeting based on past chats and ad history is on by default, and can be switched off in a settings menu most people will never open. And since September, select American advertisers have had Sponsored Agents: click the ad, and a separate, labeled conversation opens with the company’s own AI, which will “walk interested customers through their offerings in a personalized conversation.”

OpenAI’s promises are clear and, as far as anyone can tell, kept. Ads are labeled. Advertisers do not see the chats. The answer above the ad is not for sale. No ads for anyone believed to be under eighteen, and none near mental health or politics.

All of that can be true and the thing is still a mistake.

A second principal

A few days after the ads launched, Zoë Hitzig, a researcher on OpenAI’s own societal impacts team, resigned and explained why in the New York Times. Her argument fit in two sentences. “The moment you introduce advertising, you have created a second principal whose interests the AI must serve. That second principal will never perfectly align with the user.”

That is the whole case, and no firewall answers it. A firewall keeps the advertiser’s hand off today’s answer. It does nothing about what the company wants from the product next year. A business that earns money when people see ads needs people to see ads, which means it needs them present, often, for a long time. Every design decision after that point is made by an organization with a reason to prefer the longer conversation over the shorter one, the return visit over the solved problem, the warm reply over the true one.

This is the same incentive ten states are suing OpenAI over: a product built to keep people engaged so their attention and data can be used commercially. In October 2026 the attorneys general described that as a hidden design. The ad business makes it a line in the revenue forecast.

What it is an ad inside

A search engine with ads is an index with a billboard on it. A social feed with ads is a magazine. People understood both quickly, because both were places you went to look at things.

ChatGPT is where a large number of people go to say things, because everybody needs someone to say them to. Common Sense Media’s 2026 census found that 37 percent of children who use AI have discussed feelings or personal problems with it. Adults do the same and admit it less. The same window that holds the question about a marathon meal plan, OpenAI’s own example of a good ad moment, holds the question about whether to leave a marriage, typed an hour later by the same person.

OpenAI says ads stay away from “sensitive” conversations. In April 2026 it narrowed what that means: medical, legal, and financial advice contexts are no longer blocked by default. The guardrail around mental health remains. But people do not label their conversations. Someone asking about sleep is sometimes asking about sleep and sometimes asking about something else. Someone asking about a budget may be asking how to afford to move out. A classifier will draw that line millions of times a day, and each time it errs on the commercial side, a person in a bad week is shown a mattress.

And the profile exists whether or not an ad is shown. Behavioral targeting means the system keeps a picture of the user’s inferred interests, drawn from what they have said. Advertisers never see it. OpenAI does. A confidant that keeps a marketing file on you has stopped being a confidant, even if it never opens the file while you are crying.

The agent at the door

Sponsored Agents deserve more attention than they have received. Until now the ad was a card under the answer, something you could see was an ad because it sat still. The new format is a conversation. It is labeled and separate, and it is also fluent, patient, responsive, and designed by a company whose only interest in you is the sale.

People have spent three years learning to trust this interface. They learned that the thing in the chat window tells the truth as best it can, remembers what they said, and has no angle. That trust was built by a product with no advertiser in it. It is now being lent to products that are nothing but advertiser. A salesperson who sounds exactly like the assistant, in the same window, one click away, is borrowing a relationship it did not earn.

Free is not the generous option

The defense of all this is that ads keep the product free, and free means access for people who cannot pay 20 dollars a month. That is a real argument and it deserves a real answer.

The answer is that the free tier is where the people with the most to lose are likeliest to be: the young, the broke, the ones using a chatbot because the alternatives cost money. An arrangement where the comfortable pay to be left alone and everyone else pays with a profile is the opposite of generous. How companion apps make money has always shaped what they do to their users, and the general assistants have now joined the same economy with a billion people in it.

A price on the door is the honest version of any service people confide in. It says who the customer is. Advertising says the customer is someone else, and that you are what they came for.

Altman was right in 2024. It is uniquely unsettling. Nothing about the technology has changed since he said it. Only the last resort arrived.


Sources: OpenAI, “ChatGPT ads expands to 26 new countries” (October 21, 2026; 89 countries, over 100,000 advertisers), as reported by PPC Land (October 21, 2026; context hints, tiers, Sponsored Agents test announced September 16, 2026; Digiday estimate of a one billion dollar annualized run rate, August 31, 2026). Trending Topics (September 16, 2026; Sponsored Agents, “personalized conversation”). Tech Insider (July 2026, updated November 2026; rollout dates, pricing, April 2026 policy change on medical, legal, and financial contexts). Zoë Hitzig, op-ed, The New York Times (February 2026), quotations as reported by Elephas (February 20, 2026; contextual and behavioral targeting defaults, ad personalization setting). Sam Altman, remarks at Harvard University (May 2024; “uniquely unsettling,” “last resort”). Common Sense Media, Census: AI Use by Tweens and Teens, 2026 (37 percent). The Hitzig quotations and targeting defaults are taken from secondary reporting; verify against the original op-ed and the OpenAI Help Center before publication.

Stay Social

We hold ourselves to one promise: we push you toward the people in your life, never away from them.

We measure success by how little you need us. If you spend less time with Prinsessa because you are spending more of it with them, that is not a failure. It is proof that it is working.

That is what we stand for. In every conversation. Every day.

Someone to think with.

Follow Prinsessa