Sponsored, Still Speaking

OpenAI has given the advertisement a reply button.

Its new Sponsored Agents let someone click an ad in ChatGPT and begin a conversation with an agent sponsored by the business. The person can ask whether a dining table will fit, how many people it seats, or what the finish needs. OpenAI says the exchange will be clearly labeled, distinct from ChatGPT's independent answers, and separate from the conversation the person was already having.

That boundary is important, but is not the same as honesty.

A label can identify who paid for the room. It cannot tell you what the voice inside the room has been instructed to do. Is the agent trying to help someone decide, including deciding not to buy? Is it trying to keep the brand in consideration? Is it rewarded for a lead, a click, or a sale? When accuracy and conversion pull in different directions, which one wins?

The public material does not answer those questions.

OpenAI's Ad Tools Terms make the gap more concrete. They treat the advertiser as the builder responsible for a Sponsored Agent's configuration and output. The materials an advertiser can supply include prompts, instructions, business logic, sales or support scripts, and tone or brand guidelines. None of that proves the agent will mislead someone. It does establish that the conversation is shaped by a commercial principal, not merely surrounded by one.

Then the conversation starts.

The label stays behind as a visual element. The person moves forward through language: Will this fit through my door? Is there a cheaper version? What happens if the finish scratches? I have two children and a dog; what would you choose? Each answer can feel less like an ad impression and more like assistance shaped around a life.

This creates a disclosure problem deeper than recognition. Commercial identity is no longer only a property of one card. Commercial alignment is a condition that has to remain legible across time.

An advertising label has traditionally lived near a bounded object: above a search result, beside an image, before a video, or on the page opened after a click. The Federal Trade Commission's native-advertising guidance treats format as part of the claim. People need to know when material that resembles ordinary content comes from a sponsor, and the disclosure should appear before they enter the advertisement and again where the complete advertising content lives.

A conversation does not hold still like a page. Its focal point keeps moving to the newest answer.

And that is the key distinction because chat has trained us to understand the voice as responsive. A banner cannot hear an objection. A sponsored agent can answer it, reframe it, ask a better question, and adjust the pitch to the reason the first pitch did not work. The useful qualities of conversation—memory, patience, specificity, adaptation—also make commercial influence harder to locate in any one sentence.

OpenAI's advertising principles say advertisers cannot shape, rank, or alter ChatGPT's independent answers. That is a meaningful protection, but it should not be borrowed as a description of the sponsored agent itself. The point of entering that separate room is to speak with an AI-generated representative of the advertiser, and knowing that is a representative of the advertiser, not objective truth.

OpenAI's ad policies prohibit misleading claims, false endorsements, deceptive ads, and imitation that could make an ad seem like part of ChatGPT. However, a prohibition is not a description of behavior. The policies do not show whether a Sponsored Agent must volunteer a product's weaknesses, present a competitor fairly, distinguish a merchant claim from independent evidence, or resist using a person's hesitation as another sales cue. A system can avoid a plainly false statement while still being relentlessly selective about the truths it offers. There is no scorecard for honesty or truthfulness in sponsored content.

A study published before Sponsored Agents were announced offers a resistant case. In Ads that Talk Back, researchers built an experimental chatbot that inserted personalized product promotions directly into its answers. That is not OpenAI's announced design: the study mixed ads into an assistant's replies, while OpenAI says its sponsored conversations will be separate from independent answers. The difference is consequential.

The user behavior is still instructive.

Even with a “Sponsored” link attached to advertising responses, 49.15 percent of participants in the disclosure conditions did not realize they had been served an ad. Only four of sixty clicked the disclosure. Eighteen participants instead questioned the chatbot about the ads or asked it to remove them. They tried to manage the commercial relationship in the same medium where they encountered it: conversation.

One person asked the bot to stop generating the advertising portion of each response. Another asked it to remove a product reference. The static control had said learn why. The person said stop.

That distinction should shape sponsored agents from the first turn. A useful commercial conversation needs more than an entrance label. It needs a sponsorship posture the person can inspect and change without leaving the exchange.

At minimum, the room should keep answering five questions.

Who is speaking? The sponsor's name and the agent's commercial role should remain close to the active response, not only in the header someone stopped noticing eight turns ago. The visual treatment can become quieter after entry without pretending the relationship has ended.

What is this agent trying to optimize? “Help you explore” is not an objective someone can audit. The interface should say whether the agent is designed for product education, support, lead collection, or conversion, and whether it is allowed to recommend no purchase or a competitor. If those goals change during the conversation, the person should not have to infer the change from a suddenly firmer pitch.

What constrains its claims? Product catalog data, store policy, available inventory, advertiser instructions, and independent evidence are not interchangeable. The conversation should identify which claims come from the merchant, cite sources that can be checked, and make uncertainty and unfavorable information as easy to surface as benefits. A policy against deception is a floor. The person still needs a way to see how the system handles omission, comparison, and disagreement.

What does it know about me? OpenAI's ads FAQ says ad selection can use the current chat and basic context such as general location or language. With personalization enabled, it can also use signals from past chats and memory. OpenAI says advertisers do not receive those chats, memories, or personal details. If a person chooses to message an advertiser through an ad, however, the advertiser sees the messages sent directly to it.

That leaves an important interface question. Before the first sponsored reply, the person should be told which context selected the ad, which context is available to the Sponsored Agent, and which new messages will become visible to the advertiser. “The advertiser cannot see your memory” and “the system used your memory to bring you here” are different facts. Both belong at the boundary.

Can the conversation hear no? “Stop showing sponsored recommendations,” “compare this with a competitor,” and “return me to the independent answer” should behave like controls, not topics for persuasion. If a preference changes, the interface should confirm what changed and whether it applies to this turn, this agent, or advertising more broadly.

This moves earlier than Shopping Agents Need A Receipt. A receipt can show what influenced a purchase after an agent acted. Sponsored conversation asks what must remain visible while influence is still being formed, before there is a transaction to explain.

Separation is a useful boundary, but it does not settle the question of trust. A distinct sponsored room can still be tuned to flatter, withhold, redirect, and persist. It can also be designed to admit limits, surface alternatives, and let the person leave. The label alone does not tell us which room we have entered.

After ten turns, the person should still be able to answer four plain questions without opening a policy page: Who is this agent working for? What is it trying to get me to do? What did it know and learn about me? What happens when its commercial interest conflicts with mine?

Conversation can blur the line between a library, a bookstore, and a dealership. The room should still tell us which one we are standing in.