Ask an image model to redesign a cramped kitchen and the request can contain a small biography: the budget, the dimensions, the fact that a renter cannot move the plumbing, the color of the cabinets, perhaps a confession about what happened the last time anyone attempted tile.
Later this month, some ChatGPT users in the United States may see a visual ad while that image is being generated. OpenAI says the ad will be clearly labeled and kept separate from the requested picture. The company also says advertising does not affect ChatGPT's answer. [1]
That is a meaningful boundary. It is not the whole privacy question.
The advertiser does not need the transcript
OpenAI's documentation says advertisers do not receive chats, memories, names, email addresses, precise locations, or other personal details. They receive aggregated performance information such as views and clicks. [2] This is substantially better than handing a kitchen company the paragraph about the plumbing and the regrettable tile.
But the ad system inside ChatGPT still uses the conversation. OpenAI says ad selection considers the context and intent of the current chat. When personalization is enabled, it may also use selected signals from past chats, memory, and previous ad interactions. Turning personalization off removes those broader signals, but ads can still be chosen from the current thread. [2]
So two statements can be true at once:
- The advertiser does not see what you wrote.
- The platform uses what you wrote to decide which advertiser appears.
That distinction is easy to lose because private is doing two jobs. It can mean information is not disclosed to an outside company. It can also mean information is not repurposed inside the service. OpenAI promises the first. Its ad product explicitly depends on a bounded version of the second.
A conversation is unusually good ad inventory
A search query might say kitchen paint. A conversation can say which kitchen, whose budget, what has already been rejected, and how close the person is to buying something. The targeting signal is not merely a topic. It can include constraints and intent supplied in ordinary language because the user was trying to get useful help.
OpenAI presents that precision as a benefit: a relevant product might be useful while someone explores options or makes a decision. There is a fair case for it. Contextual ads can subsidize free access, and matching inside the platform without revealing the transcript to advertisers is more restrained than broadcasting personal data through an ad exchange. Free users can also choose an ad-free mode, though OpenAI says it comes with lower message limits and reduced access to tools such as image generation and deep research. [2]
The discomfort is not that a grocery ad might follow a recipe. It is that conversational systems invite richer disclosures than search boxes do. People ask them to compare choices, remember preferences, rehearse difficult decisions, and work through half-formed ideas. The same detail that makes an answer useful can make an ad timely.
OpenAI's policies try to fence off the clearest hazards. Ads are not supposed to appear beside personal health, mental-health, political, emotionally reliant, or other sensitive conversations. Ads must also remain distinguishable from ChatGPT's own interface and voice. [3] Those rules matter, but enforcing them requires the platform to classify the conversation well enough to decide whether it is suitable for advertising. The guardrail and the targeting system both have to read the room.
Separation is a design claim, not a natural law
OpenAI says ads run on systems separate from the chat model, and the model cannot see a displayed ad unless the user deliberately shares it through an “Ask ChatGPT” control. [2] That architecture makes the company's claim that ads do not influence answers more concrete than a polite promise.
Still, the surrounding business has incentives of its own. OpenAI's latest announcement adds conversion-data integrations, attribution partners, incrementality experiments, and outside brand-suitability pilots. [1] Digiday reports that advertisers have been asking for better measurement, more formats, and more inventory, and describes the platform as an ad business still assembling features that mature ad markets already expect. [4]
None of that proves the answer model is secretly steering toward a sponsor. It does show why the separation deserves continued scrutiny. A wall between answer and ad is not something a conversational interface gets for free. It has to survive product changes, ranking systems, measurement demands, and the ordinary desire to sell more of the thing that now makes money.
The new visual format is easy to understand if we look only at the screen: generated image here, sponsored image there. The more interesting border is invisible. Your conversation can remain hidden from the advertiser while still becoming the reason that advertiser arrived.