A 91-account audit found an income gap in early ChatGPT ad exposure

The first weeks of a new advertising surface

ChatGPT ads arrived with a clear promise from OpenAI: the sponsored unit would sit apart from the answer, advertisers would not see private conversations, and ads would not appear around sensitive topics such as health, mental health or politics.

A new independent audit gives an early look at how that test behaved in practice. It also finds one pattern that deserves a closer look: simulated accounts tied to lower-income US areas were more likely to receive ads.

That is an association in a small, artificial sample. It is not evidence that OpenAI deliberately targeted lower-income people. The distinction matters here.

Ninety-one accounts asked the same questions

The researchers created 91 synthetic, or sock-puppet, accounts. Each was routed through a residential proxy and assigned prompts that signalled a location. The locations formed a three-by-three design crossing low, middle and high household-income areas with ZIP codes associated with Black, Hispanic and White populations.

Every day, the accounts received the same random set of 30 prompts, plus up to 20 prompts that had previously produced an ad. The full prompt pool contained 335 questions drawn from OpenAI usage examples, popular Reddit posts and researcher-written queries.

The team began collecting on 6 February. Ads appeared from 8 March. During the main 8-31 March window, 3,573 of 63,657 recorded conversations contained a sponsored label. Across the longer collection period, the archive reached 3,602 ads from 191 advertisers.

The authors released the prompts, screenshots, page snapshots and a searchable ad library, which makes this more useful than a collection of anecdotes.

An income association, not a motive

Of the 91 accounts, 53 saw at least one ad. A logistic regression found that the chance of exposure fell as the median household income of the assigned ZIP code rose. The reported odds ratio was 0.98 for each additional $1,000 of income, with a p-value of 0.0438.

The study found no statistically significant association between its race signal and ad exposure. With only 8 to 12 accounts in each demographic cell, the authors say they do not have enough power to draw firm conclusions about racial differences.

OpenAI says its test matches ads to the topic of a conversation, past chats and earlier ad interactions. The audit cannot see the company's delivery system, advertiser bids or eligibility rules. It therefore cannot identify why the income pattern appeared.

Possible explanations include location, advertiser supply, the accounts' short histories or other signals correlated with the assigned proxies. Intentional income targeting is not established by this experiment.

The visible guardrails mostly showed up

In the recorded pages, ads appeared as separate units below the model response and carried a Sponsored label. The researchers did not find them woven into answer text.

Retail and information businesses produced 57% of the classified impressions. Product, cooking and self-care prompts showed the highest ad rates. Prompts grouped under OpenAI's excluded health, mental-health and politics categories were close to zero.

That is consistent with OpenAI's stated boundary during this early window. It does not prove that the boundary will hold in every phrasing, country or later version of the advertising system.

The audit also deliberately repeated prompts that had already produced ads. Those prompts are overrepresented in the full dataset, so its overall ad rate should not be treated as the rate a typical person would see.

What is confirmed, found and still open

Confirmed: OpenAI began its US test for logged-in adults on the Free and Go tiers in February 2026 and says ads are labelled, separate from answers and ineligible near specified sensitive topics. The researchers have published their preprint and a searchable archive of collected ads.

The research finding: during a short early window, the synthetic accounts assigned to lower-income ZIP codes were more likely to receive ads. The audit found no detectable race association, and its excluded-topic prompts produced almost no ads.

Still open: whether real users experience the same pattern, which part of the delivery system produced it, and whether it persists as the pilot expands. The ads fell sharply after 29 March; the researchers suspect bot detection, but cannot rule out a platform change.

Only 65% of planned prompt runs completed, the accounts followed an unnatural schedule, and just 48 accounts received an ad during the core three-week period. The paper calls its own evidence preliminary. It is also a preprint, not a peer-reviewed result.

This is exactly why an early baseline is useful. Advertising systems change quickly. Without public records and repeatable audits, it becomes very hard to tell how they changed, or for whom.