404 Media finds ChatGPT ad targeting uses prompts and uploaded contact details

ChatGPT advertisers can use conversation topics and uploaded contact details to reach users, according to an October 15 investigation by 404 Media. After running two ads itself, the outlet described a system that matches commercial messages to chats and advertiser-supplied audiences. Users who disable ad personalization can still receive ads based on their current conversation.

Joseph Cox’s investigation approached the system from the advertiser’s side. Rather than only checking which ads appeared during ordinary use, 404 Media bought two placements. Both promoted its earlier reporting about contractors reviewing real ChatGPT conversations, giving the outlet a way to examine the campaign controls available to an advertiser.

Buying an ad about ChatGPT privacy

The campaign setup lets advertisers enter “context hints”: descriptions of conversations, topics or keywords where their offering might fit. OpenAI’s interface says these descriptions guide matching rather than act as exact-match rules. That distinction means the tool offers topical guidance, not a promise that a particular phrase will trigger an advertisement.

For its ads, 404 Media supplied hints about people discussing privacy, the privacy of ChatGPT chats, and technology or AI topics. Its test therefore put an advertisement about the service’s handling of conversations into the same advertising system that uses conversation topics to help choose placements.

Contact lists become advertising audiences

The investigation also examined a different targeting route: uploaded audiences. Advertisers can submit raw email addresses and phone numbers, hashed versions of those details, or Android advertising IDs. Hashing digitally obscures the contact information. OpenAI checks the submitted identifiers against ChatGPT users so advertisers can target a matching audience.

OpenAI’s response set limits on what that matching enables. A spokesperson told Cox that inclusion filters and audience bid adjustments require at least 25,000 matched users. Advertisers cannot see which individuals matched or select particular people within an audience. Being on an uploaded list does not, by itself, cause an ad to appear.

The company said uploaded files are deleted after processing, usually within 24 hours of upload. But one handling detail remained unanswered: whether OpenAI hashes raw contact data after receiving it. Asked about that step, the company instead reiterated that advertisers can upload either raw information or already-hashed versions.

Allow ChatGPT to use your past chats, activity, and preferences to select ads. Ads may still be based on your current chat.

ChatGPT’s ads personalization setting, quoted by 404 Media

The opt-out boundary—and outside scrutiny

For users, the personalization switch is therefore narrower than an advertising opt-out. Its wording separates the use of past chats, activity and preferences from the use of the conversation happening now. Disabling personalization does not mean that the subject of a new exchange will be excluded from ad selection.

Outside observers are also mapping the connection between prompts and placements. Cox described ChatGPT Ad Library, which says it enters realistic consumer prompts and records the sponsored cards returned, including the advertiser and triggering prompt. It reported capturing at least 43,410 ads across 415,289 placements. Those figures describe its own sample, not OpenAI’s internal totals.

Zach Edwards, co-founder of advertising-research company DecryptAds, argued for a legal response. He told 404 Media that advertisers sharing sensitive personal information with ad-tech companies such as OpenAI should have to register as data brokers and give people a clear removal process. His proposed remedy focuses on control over the supplied data, beyond limits on individual targeting.