Advertising inside conversational AI raises a measurement problem as well as a trust problem. Users need to recognize when they are seeing a paid message, while researchers need to examine who receives those messages. A clear label addresses the first issue, but says little about whether advertising exposure differs systematically across groups.
Research discussed by Brookings on October 26 found that study accounts associated with lower income ZIP codes received ads more often and in greater numbers. The audit used 91 accounts and collected more than 3,600 advertisements. Geography served as a proxy for demographic characteristics, so the findings should not be read as direct proof that the system knew each user’s income or deliberately discriminated against them.
The limits of the findings matter. Researchers did not identify evidence that race or ethnicity affected whether accounts received ads, or obvious failures in the placement safeguards they tested. OpenAI’s published policies prohibit advertising in various sensitive contexts and require ads to remain distinguishable from the product experience. An early audit can test some outcomes, but it cannot certify every future campaign or explain every observed difference.
Earlier platform disputes show why delivery deserves scrutiny alongside advertiser intent. In 2022, the US Justice Department secured a settlement requiring Meta to change aspects of its housing advertising systems. That case concerned alleged discrimination under housing law, not ChatGPT. Its relevance is methodological: automated distribution can produce civil rights concerns that require examination beyond the wording of an advertisement.
Consumer understanding provides a second test. FTC guidance requires clear and prominent disclosure when necessary to prevent deceptive advertising. Applied to conversational interfaces, that principle supports labels that users can recognize while reading, rather than notices they must search for. Independent access to delivery data would complement those disclosures, allowing unequal exposure to be investigated without treating every observed difference as proof of unlawful conduct.
Original analysis inspired by Sorelle Friedler, Emma Lurie and Danaé Metaxa from Brookings Institution. Additional research and verification conducted through multiple sources.
