The First ChatGPT Ads Audit Finds an Income Gap—Not a Verdict on Intent

Researchers submitted the first empirical audit of advertising in ChatGPT on August 5. Their study created 91 controlled accounts in a three-by-three design, using location cues and prompts to represent three racial or ethnic groups and three income bands. Beginning in February 2026, the accounts generated more than 3,000 ads from 186 advertisers across 335 realistic prompts.

The headline finding is an exposure difference: lower-income test accounts were more likely to receive ads, regardless of the race signal used in the experiment. The researchers also found that ads began appearing 14 days after account creation, skewed toward consumer goods, usually promoted an advertiser rather than one specific product, and remained visibly separate from the assistant’s answer during the period studied.

An observed gap is a question, not an explanation

A sock-puppet audit is useful because it holds many behaviors constant while varying selected signals. It can reveal that otherwise similar synthetic accounts experienced the system differently. It cannot, by itself, identify which ranking feature caused the difference, show whether the system explicitly inferred income, or establish discriminatory intent. The accounts are experimental proxies, not a representative sample of every ChatGPT user.

That distinction should sharpen the response rather than soften it. OpenAI can test whether ad eligibility, inventory, location, language, prompt mix, or personalization produced the gap. Independent researchers can repeat the audit across more regions, account ages, devices, topics, and time periods. A finding that survives those changes would be stronger evidence of a systematic allocation effect.

Conversation context makes ad governance unusually sensitive

OpenAI says ad selection can use the current conversation and basic context such as general location or language. When personalized ads are enabled, it may also use past chats, memory, personalized responses, and ad interactions. The company says those signals remain inside ChatGPT, advertisers receive aggregated reporting rather than conversations or personal details, and ads do not influence answers.

Those boundaries matter, but users also need them to be observable. A useful ad explanation should distinguish current-thread context from longer-term personalization, identify the broad topic that triggered an ad, and make opting out immediate. OpenAI currently offers ad controls, deletion of ad data, paid ad-free tiers, and a free ad-free option with lower limits. Audits should test whether those controls reliably change outcomes.

The early policy perimeter will keep moving

OpenAI’s current policies permit a limited set of consumer categories and prohibit or restrict areas where persuasion can cause greater harm. Political advertising is disallowed; health services and claims face restrictions; and ads cannot exploit sensitive topics. The company also says ads do not appear around personal health conversations or for accounts identified as belonging to minors.

The paper captures an early system, not a permanent one. As inventory, targeting, measurement, and formats expand, the most valuable safeguard will be a public audit trail: exposure rates by broad demographic proxies, enforcement data, clear change logs, and repeatable researcher access. The study does not prove why lower-income accounts saw more ads. It does show that conversational advertising is already measurable—and should be measured before its incentives become harder to unwind.