ChatGPT Privacy: How Ad Targeting and Government Access Work - FINCHANNEL
An advertiser may be able to reach someone on ChatGPT by using an email address or phone number collected from another source. A government agency may also obtain records of what a user shared with the service if it follows the required legal process. Both practices are documented, but they give access to personal information in different ways.
Privacy on ChatGPT came under new attention after an investigation published by 404 Media on Oct. 15. The report said advertisers could target groups of people using email addresses, phone numbers and mobile advertising IDs. The publication said it based its findings on tests of OpenAI's advertising system.
OpenAI says advertisers cannot read users' conversations. Separately, the company's government transparency report records cases where it shared user content in response to legal requests. This content can include text and files that users submitted to its services or that the services created.
These records describe several ways information shared with an AI assistant may be used or shared. These include choosing ads, improving AI models, using outside applications, software security problems and legal requests. Each has different rules.
How ChatGPT advertising uses personal information
According to 404 Media, the available audience-targeting options included email addresses and phone numbers converted into hashed versions, as well as Android advertising IDs. Hashing turns an identifier into a digital form that can be matched with another record without uploading the original identifier in readable form. In this case, the purpose is to match records, not to make the person impossible for the system to recognize.
OpenAI's Ad Tools Terms clearly allow advertisers to provide audience data so the company can create custom audiences. Advertisers must promise that the information comes directly from their own records, rather than being bought or obtained from data brokers or other outside suppliers. They must also have the required notices, permissions and legal reasons to provide the information.
The terms forbid advertisers from using these tools to identify or identify again OpenAI users, guess sensitive information about people, or provide banned types of information, such as health information, political beliefs or financial hardship. These are rules in the contract. Their existence does not prove that every advertiser always follows them.
Ads can also be matched to the conversation itself. OpenAI's information for advertisers says its system looks at the topic and purpose of a conversation, the ad's content and destination, and "context hints" provided by the advertiser. These hints help the system match ads to conversations, but they do not guarantee that an ad will appear in any particular conversation.
OpenAI says its advertising system works separately from the chat model and cannot change or influence its answers. Advertisers receive combined performance data, not access to chats, memories or conversation histories.
Under OpenAI's current policy, ads may appear on Free and Go plans. Plus, Pro, Business, Enterprise and Edu accounts do not have ads. If users turn off ad personalization, they may still see ads based on the conversation they are having. However, other chat threads and advertising history will no longer be used for that purpose. OpenAI says it does not show ads in sensitive conversations, including those about personal health, mental health and politics.
In comments provided for this article, Jurgita Lapienytė, editor-in-chief of Cybernews, said advertising might make people rethink how personally they relate to an AI assistant.
“People turn to ChatGPT for personal advice because it feels like talking to a supportive friend,” she said.
Her concern is about where the ad appears: a commercial message may show up alongside a conversation that the user felt was private.
Governments already obtain user records
OpenAI's transparency report for January through June 2026 documents government access to user information. The company reported receiving 146 requests for content and sharing data in response to 88 of them. These disclosures involved 130 accounts. Content can include users' text and files, as well as responses created by the service.
OpenAI also received 620 requests for information that was not the content of conversations. It shared data in response to 441 of these requests, covering 679 accounts. This type of information can include account names, email addresses, payment details and transaction histories. Of 28 emergency requests, 10 led to disclosures involving 13 accounts.
Under its policy for government requests for user data, which took effect on Jan. 1, OpenAI's U.S. company requires a valid warrant or an equivalent legal document before giving law enforcement the content of conversations, except in emergencies. Information that is not conversation content may be shared in response to a subpoena, court order, warrant or equivalent legal process.
The emergency exception applies when there is a risk of death or serious physical injury and the company believes sharing the information is necessary to prevent harm.
Foreign authorities also have ways to request information. Depending on the situation, OpenAI's policy allows certain foreign legal demands or requires international procedures that obtain the records through the U.S. legal system. The policy sets out separate requirements for users served by OpenAI Ireland.
Requests related to national security are reported separately. For the first half of 2026, OpenAI said it received between 250 and 499 national-security requests involving between 250 and 499 accounts. The published ranges combine National Security Letters and requests made under the Foreign Intelligence Surveillance Act. They do not identify individual cases.
These reports show that governments can obtain records held by OpenAI through legal processes. They do not support the claim that governments have unlimited access to everyone's conversations.
Where information can escape
Public reports do not provide enough evidence to reliably rank the most common ways information leaks from ChatGPT. However, they document several different ways information may be exposed.
One example was a software bug. On March 20, 2023, a bug allowed some users to see titles from another user's chat history. It may also have exposed the first message of a newly created chat. OpenAI's report on the incident said payment-related information might also have been visible for 1.2 percent of Plus subscribers who were active during a particular nine-hour period. Full credit card numbers were not exposed, and the company fixed the bug.
Another example involved sharing conversations publicly. In an interview with The Verge in August 2025, ChatGPT head Nick Turley confirmed an experiment that let users who shared conversations choose to make them discoverable through Google. OpenAI later withdrew the feature. The exposure involved shared conversations that could be found through search, not a general opening of private accounts.
Outside applications create another point where information may be shared. OpenAI's GPT documentation says that when a GPT uses apps or external APIs, relevant parts of a user's input may be sent to third-party services. People who build GPTs cannot see individual conversations just because they created the GPT. However, an outside service may receive the information it needs to carry out an action.
Connected tools can also face prompt-injection attacks. These attacks hide harmful instructions inside material that an assistant reads. The instructions may try to make the assistant reveal information. OpenAI describes unauthorized sharing of data as one possible result and says protecting against these attacks remains a challenge across the industry.
Normal data processing is different from an accidental leak. OpenAI says prompts, responses and uploaded material from consumer accounts may be used to improve its models, depending on account settings. Content from Business, Enterprise, Edu and API services is not used to improve models by default. Temporary chats are not used for model improvement while they remain temporary, but copies may be kept for up to 30 days for safety reasons.
How U.S. media companies use the technology
Advertising projects run by publishers show several ways OpenAI technology is being used. However, these business arrangements do not prove that publishers receive private ChatGPT conversations.
Vox Media's partnership announcement from May 2024 combined a deal to license content with plans to develop products for audiences and advertisers. OpenAI's announcement about the partnership described plans to improve advertising content and target specific audience groups. These were announced business goals, not results confirmed by an independent review of advertising campaigns.
BuzzFeed developed sponsored AI experiences on its own websites. During its earnings call for the second quarter of 2023, the company described two sponsored quizzes for mattress brand Serta: a dream interpreter and a bedtime-story generator. Its investor materials also named Scotts Miracle-Gro as a sponsor connected to the launch of its AI-powered Infinity Quizzes. These were advertising products hosted by the publisher, not paid ads placed inside the consumer version of ChatGPT.
Advertising agencies provide a more direct way for businesses to buy ad space on ChatGPT. In an announcement in May 2026, OpenAI named Dentsu, Omnicom, Publicis and WPP as partners helping businesses buy ads, manage budgets and create advertising material. OpenAI said its own advertising system would continue to control decisions about where and when ads appear.
The business relationship can involve more than showing a sponsored message. OpenAI's current advertising policy says that if a user chooses to message an advertiser through an ad, the advertiser can see the messages sent directly to it.