ChatGPT privacy: what human review means

Many people use ChatGPT as if it were a private notebook, a quick assistant, or a semi-confidential conversation partner. That is convenient, but it can also create privacy risks. OpenAI’s official help pages, together with media reports from recent days, suggest that it is worth being more deliberate about what goes into each chat.

The most important question is not whether people should use a chatbot, but what kind of data they enter into it, and which settings can reduce their exposure. According to the official description, for consumer accounts it is possible to switch off data use related to model improvement, but it is important to understand exactly what that option covers.

What does it mean that people can also see conversations?

According to a report by 404 Media, OpenAI has external contractors review real ChatGPT conversations in an internal process referred to as Project Lily. The publication says the purpose is to improve response quality and evaluate the system.

Based on Tom's Hardware’s reporting, reviewers may in some cases see not only short excerpts but also broader context. According to the media reports, even alongside automatic filters, personal or business-related information may still end up in such review processes.

An important distinction is that this part is based on media reporting. OpenAI’s official help pages mainly explain how data is used to improve models and which settings can be used to limit this.

What does OpenAI say officially about this?

According to OpenAI’s official guidance, content from consumer ChatGPT accounts may be used to improve model performance. The same page also states that users can switch off this data use in the settings.

According to the same guidance, for Team, Enterprise and Edu plans, as well as for data handled through the API, this training-related use is not enabled by default. This is an important difference for those using these tools in workplace, client-facing or educational settings.

In short: what is the most important difference?

For a consumer ChatGPT account, it is worth specifically checking the data-handling settings, because according to the official guidance, use related to model improvement may be enabled by default. In workplace or institutional use, however, it matters a great deal which plan an organisation uses, because according to the official description, the default setting is different for business and education offerings.

• For consumer accounts, it is worth specifically checking the Data Controls menu.
• Keeping chat history and allowing data use related to model improvement are not the same thing.
• The strongest protection is still what gets entered into the chat.

Where can you switch off data use related to model improvement?

According to OpenAI’s official help pages, the setting is available in the ChatGPT interface under Settings > Data Controls. There you can disable the use of new conversations for improving the models.

On another official OpenAI page, the company separately highlights that preserving chat history and switching off model training are two separate settings. This means users do not necessarily have to give up the convenience of history in order to limit data use for training purposes.

• Open the ChatGPT settings.
• Find the Data Controls menu item.
• Check whether data use related to model improvement is switched on.
• If you do not want your new chats to be used for this purpose, switch this option off.

What does this switch not solve?

This part is particularly important because the role of the setting can easily be misunderstood. According to PCMag’s article, switching it off does not apply retroactively, meaning it does not automatically mean that earlier conversations are removed from every previous process.

According to Malwarebytes’ analysis, disabling model training is not the same as ending data retention for security and abuse-prevention purposes. The practical consequence is that this setting is more about reducing risk than completely cutting off all data-handling processes.

What is the difference between training, retention and human review?

These three concepts are often blurred together, even though they do not mean the same thing. Use for training purposes means that conversations may contribute to model improvement. Retention refers to the fact that some data may be stored for a period of time for operational or security reasons.

Human review, by contrast, means that some conversations or excerpts may also be seen by people for evaluation or quality-assurance purposes. The details about this process in the article come from media reports, while OpenAI’s official pages mainly describe the training-related use side and the user settings.

What kind of data is best not entered?

As a general precaution, it is best not to provide information whose disclosure could be inconvenient, harmful or legally problematic. This can include, for example, an identifier tied to a full name, home address, telephone number, personal document details, banking information, client databases, internal company documents or detailed health information.

According to reporting by 404 Media and Tom's Hardware, sensitive and personal content may also have reached human reviewers. This does not mean that every conversation is read by people, but it does suggest that the service should not be treated as a completely private space.

• Do not enter a password or access code.
• Do not paste in a contract or non-public business material in full.
• Do not provide easily identifiable details about a health or mental condition.
• Do not upload client data or internal employer information without permission.
• If you want to use an example, anonymise the text first.

Who should be especially careful to check the settings now?

Primarily those who use ChatGPT for daily work and regularly enter prompts containing business, legal, financial or personal data. It is just as important for students, teachers, freelancers, consultants and anyone who tends to use a chatbot as a quick second notebook.

It also matters that anyone working in a company or institutional setting checks exactly which plan is being used. According to OpenAI’s official description, Team, Enterprise and Edu subscriptions have different default data handling from consumer accounts.

What is worth doing now?

The best privacy strategy is not a single switch but several precautions used together. The first step is to review the settings, the second is to recognise that keeping chat history and allowing use related to model improvement may be separate settings, and the third is to make sure only the minimum necessary personal or business information enters the system.

In short, anyone using a consumer ChatGPT account should review the privacy options. The settings can reduce exposure, but the most important protection is still caution: if something is genuinely confidential, it is better not to type it in.