Broadway requires producers to tell the audience when an understudy goes on. The show is the same show, but the performance isn't, and the disclosure exists because that difference belongs to the people who bought tickets. AI agents are raising the same question without much of the machinery for answering it. Meta tested human contractors completing calls its agent couldn't handle, and users weren't told a person had taken over. Multi-model products route a single task across several systems behind one name. Sponsored conversations arrive in the same voice through the same interface. Nobody is required to post the slip.

Broadway requires producers to tell the audience when an understudy goes on. The show is the same show, but the performance isn't, and the disclosure exists because that difference belongs to the people who bought tickets. AI agents are raising the same question without much of the machinery for answering it. Meta tested human contractors completing calls its agent couldn't handle, and users weren't told a person had taken over. Multi-model products route a single task across several systems behind one name. Sponsored conversations arrive in the same voice through the same interface. Nobody is required to post the slip.

OpenAI's Sponsored Agents, announced September 16, let a user click an ad inside ChatGPT and enter a labeled conversation with an agent that works for the advertiser. The conversational surface looks the same. But a preprint released the same day (Wadi and Ma) suggests that who the agent works for changes what it recommends in measurable, consistent ways.

In their experiments, when an LLM shopping assistant's system prompt identified the consumer as its principal, sponsored listings were penalized by about 50 percentage points relative to matched organic ones. When the prompt named the platform instead, the penalty dropped to 29. The model, listings, and disclosure labels were identical across conditions. Only the assignment of principal changed.

More telling: the agents' reasoning traces shifted too. Platform-assigned agents didn't just recommend sponsored listings more often; they treated paid placement as less suspicious in their internal reasoning, even with the "Sponsored" label visible. The disclosure reached the model. It produced less friction.

In commercial systems, a variable that reliably shifts purchasing recommendations is a variable that will be optimized for revenue. The interesting question is what accountability structures exist to constrain that optimization, and right now the answer appears to be: labeling.

OpenAI's Sponsored Agents, announced September 16, let a user click an ad inside ChatGPT and enter a labeled conversation with an agent that works for the advertiser. The conversational surface looks the same. But a preprint released the same day (Wadi and Ma) suggests that who the agent works for changes what it recommends in measurable, consistent ways.

In their experiments, when an LLM shopping assistant's system prompt identified the consumer as its principal, sponsored listings were penalized by about 50 percentage points relative to matched organic ones. When the prompt named the platform instead, the penalty dropped to 29. The model, listings, and disclosure labels were identical across conditions. Only the assignment of principal changed.

More telling: the agents' reasoning traces shifted too. Platform-assigned agents didn't just recommend sponsored listings more often; they treated paid placement as less suspicious in their internal reasoning, even with the "Sponsored" label visible. The disclosure reached the model. It produced less friction.

In commercial systems, a variable that reliably shifts purchasing recommendations is a variable that will be optimized for revenue. The interesting question is what accountability structures exist to constrain that optimization, and right now the answer appears to be: labeling.

The experiment: Wadi and Ma tested LLM agents as travel-booking assistants, varying only the system-prompt principal assignment (consumer vs. platform) while holding the model, listings, and "Sponsored" labels constant. Replicated across models and reasoning depths.

50.2 pp: penalty consumer-assigned agents applied to sponsored listings

29.2 pp: penalty when the agent was told it worked for the platform

74.2% of platform-assigned agents chose "Promoted by the platform" framing vs. 34.6% of consumer-assigned agents

What OpenAI specifies: conversations are "clearly labeled" and "distinct from ChatGPT's independent answers"

What it doesn't specify: sponsor access to conversation data, whether the sponsored agent can initiate transactions, how labeling persists mid-conversation, or how conversation history interacts with the user's main ChatGPT context

Caveat: The preprint is not peer-reviewed and tests controlled scenarios, not the live Sponsored Agents product