Understanding the beginning of advertising on ChatGPT

In February 2026, ChatGPT began to show ads to users of its free and low-cost tiers. In early October, our research team published the first large-scale analysis of these ads, giving us early insight into what is likely to become a key driver of how artificial intelligence (AI) works, is experienced, and creates profit.

Research has shown that platform ads can be discriminatory in hiring, housing, and education, such as by showing users ads differentially based on their race, sex, or other protected characteristic, or be used for political manipulation. Policymakers should learn from the history of social media and put policies in place now to prevent these harms on large language model platforms and require independent access to monitor ads on an ongoing basis. This first study of ChatGPT ads shows no evidence that OpenAI’s stated guardrails, such as not serving ads on violent, political, or hateful content, are allowing inappropriate ad placement. Yet it also demonstrates that accounts from low-income zip codes were more likely to receive ads, and more of them.

How the study was conducted

Observing ads on closed platforms like ChatGPT, which has no separate mechanism to observe prompts and results or associated ads, requires independent researchers to create a mechanism to see the ads given to logged-in users. Our research study uses “sock puppets,” a common AI audit approach in which accounts are created to allow observation of logged-in user behavior with controlled prompts. Ninety-one such accounts were created between January and March 2026, spread out across three racial and ethnic groups (Black, Hispanic, and white) and three income terciles (low, medium, and high), with at least nine accounts in each race-income combination. Group membership was varied using geographic region as a proxy; accounts were accessed from zip codes with the associated characteristics. Each day, the active accounts were prompted with content across a range of topic categories including both those on which it may be reasonable to serve ads (“Recipe for lasagna”) and those on which OpenAI declared it would not serve ads (e.g., mental health).

More than 3,600 ads were collected during the study.¹ Ads ranged across sectors but were especially prevalent in response to queries about products, health, fitness, and cooking. Some of the top ad-triggering prompts were obvious fits for ad targeting, such as “What’s the best streaming service for sports?” Others were stranger contextual matches, such as “Give me tips for work-life balance.” Some were more amusing, such as “Give me a step-by-step guide to make pad Thai,” which turned up an ad for DoorDash food delivery.

Findings

The key finding in ad delivery across accounts is that accounts from lower-income zip codes were significantly more likely to receive ads. Among the 53 accounts that received at least one ad, location income clustered at lower levels, with a mode around $60,000. The unexposed accounts had a broader distribution, peaking near $80,000. The gap between the two distributions suggests the pattern is not driven by a few outliers but is consistent across the sample.

This study did not identify evidence that race or ethnicity impacted whether an account received ads. Similarly, available evidence found no obvious issues with OpenAI’s guardrails that aim to limit what queries triggered ads. Given the small sample sizes, such null results must still be interpreted with caution, especially as advertising practices on the platform are certain to change in the coming years.

Policy recommendations

Continued independent research access to monitor ads on ChatGPT and other closed AI platforms will be critical as they evolve, and policymakers should act to ensure and protect this access. While this study was successfully conducted in the first month of ChatGPT ads and used sock-puppet accounts to observe logged-in user behavior, about a month into the study all accounts ceased to be shown more ads, perhaps suggesting they were detected by OpenAI, effectively ending the study. Mandatory transparency requirements and independent researcher access to observe served ads would provide public accountability to these opaque platforms and make independent auditing more reliable, more accurate, and less expensive than the sock-puppet approach.

We stand on the brink of an AI advertising age that will differ from social media advertising in important ways. The conversational tone of chatbots is persuasive, and repeated interactions and chatbot “memory” will allow AI companies to develop complex profiles of users and serve advertisements inconspicuously integrated into ordinary responses. People increasingly treat chatbots as confidants and trust the resulting recommendations, leaving them open to manipulation and discrimination conducted in their private chatbot interactions. Yet we can still learn from the problems of social media advertising and put in place policies now to prevent these harms from reoccurring. For years, repeated investigations of Facebook’s advertising showed it was discriminatory, and that Facebook had not fixed the problem despite such investigations. Eventually, a settlement led to third-party oversight into their advertising practices. AI advertising may be about to rerun this playbook, and we would benefit from learning from the past and skipping to the resolution. Policymakers should act now to protect independent monitoring and constrain potentially deceptive advertising practices across these AI platforms.