You Uploaded Sensitive Information to AI. Now What?

What happens to information you share with ChatGPT, Gemini, or Claude—and what you can do if you shared something you shouldn't have.

Imagine you're preparing a presentation for an upcoming client meeting. You have a spreadsheet containing customer information, sales figures, and a few internal notes. Rather than spending an hour organizing the data yourself, you upload it to ChatGPT, Gemini, or Claude and ask the AI to identify trends and create a summary.

Ten minutes later, you've got exactly what you needed.

Then you remember something.

That spreadsheet also contained customer contact information, confidential pricing, and some details you probably shouldn't have shared outside your company.

Did you just create a security problem? Can anyone else see that information? And if you delete the conversation, is the information actually gone?

The good news is that uploading sensitive information to an AI platform doesn't automatically mean you've exposed it to the public or caused a data breach. Major AI providers have security protections and privacy controls designed to protect the information users share.

Still, there are some things worth understanding and a few practical steps you can take.

First, where did your information go?

When you upload a document to an AI service, the information is transmitted to the provider's systems, where it is processed to generate your response.

Depending on the platform, your account, and your settings, several things may happen afterward.

• Your conversation may be saved. Most AI platforms maintain conversation histories so you can return to previous work. Uploaded documents may also remain associated with your conversation, a project, or a separate file library.

• The AI may remember certain details. Platforms such as ChatGPT, Gemini, and Claude offer personalization or memory features. These can retain useful information across conversations. Memory is different from conversation history, so deleting a conversation doesn't necessarily remove information that was remembered separately.

• Your information may be eligible for model training. Some consumer AI services use conversation data to improve their models, depending on your privacy settings. Business and enterprise services frequently provide stronger default protections against training on customer data.

• Limited human access may be possible. AI providers may permit authorized personnel or service providers to review certain content for safety, support, legal compliance, or model improvement under their policies.

None of this means another person can simply open ChatGPT and ask it to retrieve your company's spreadsheet. Private conversations are not ordinarily available for other users to browse.

However, the distinction between information being privately stored, retained for training, shared with another service, or publicly accessible matters when deciding what to do next.

Step 1: Determine what you actually shared

Before deleting anything, consider the information involved.

Returning to our spreadsheet example, perhaps you uploaded customer names, email addresses, and internal pricing. That's potentially confidential business information, but the severity depends on the circumstances.

Compare that with uploading a document containing active passwords, banking credentials, or private access keys. Those create more immediate opportunities for misuse.

Some useful questions include:

• Does the information identify customers, employees, patients, or other individuals?

• Does it contain passwords, financial account credentials, or other information someone could use to gain access?

• Is it confidential under a company policy, customer agreement, or legal requirement?

• Was the information uploaded using a personal AI account or a company-approved business account?

• Was the conversation shared publicly or with another application?

If you uploaded ordinary personal information or a document that doesn't contain anything particularly sensitive, you may not need to take any further action.

If it contains active passwords or API keys, however, change or revoke those credentials immediately. Deleting the conversation alone won't make the credentials safe again.

For confidential company or customer information, notify your IT or security team when required by company policy. They can help determine whether the upload represents a meaningful exposure and whether any additional action is necessary. If the event may require investigation, preserve the details your security team needs before deleting evidence.

Step 2: Delete the conversation and any uploaded files

Let's return to our employee. After reviewing the spreadsheet, they decide the information shouldn't remain in their personal AI account.

The most obvious response is to delete the conversation.

That's a sensible first step, but deleting a chat and permanently removing all associated information aren't necessarily the same thing.

For example, OpenAI states that deleted ChatGPT conversations are removed from account history immediately and generally scheduled for permanent deletion within 30 days, subject to certain legal, security, and other exceptions.

Anthropic provides a similar general 30-day deletion period for Claude conversations. However, data previously retained for model improvement, feedback, safety, or legal purposes can be subject to different rules.

Google Gemini has its own retention practices. Some conversations reviewed by human reviewers may be retained for up to three years, even after the user deletes their Gemini activity.

It's also worth checking for copies outside the original conversation. A file might still be saved in an AI file library, project, or custom assistant. Some platforms retain uploaded files separately from the chat in which they were used.

In other words, delete both the conversation and any separately saved copies of the sensitive document. If the conversation was shared through a public link, remove that link as well.

Deleting information from your account reduces unnecessary retention and exposure, even when the provider cannot guarantee that every associated record disappears immediately.

Step 3: Check what the AI remembers

Our employee has deleted the conversation, but what if the AI retained information from it?

AI memory features allow platforms to personalize future conversations. For example, an AI might remember your job, your company's customers, or the types of projects you regularly work on.

This information isn't necessarily stored as a complete copy of the original conversation. Instead, it may be part of a memory or personalization system that draws on previous interactions.

In ChatGPT, you can review memory controls under Settings → Personalization → Memory. Depending on your account, you may be able to review a memory summary, request corrections, or turn memory off.

Claude and Gemini also provide memory-management options, although the specific controls differ.

If an AI has retained confidential information, review and remove the relevant memory using the available controls. Then make sure you've deleted the original conversation and any other sources containing that information.

You can also ask the AI what it remembers about you, but don't treat its response as a complete inventory of all the information stored in your account.

Step 4: Review your model-training settings

There's one more important setting to examine.

AI companies improve their models partly through training. Depending on the service and account type, eligible conversations may contribute to this process.

That doesn't mean a model simply saves your spreadsheet in a database that anyone can search. Training generally involves learning patterns from data.

Nevertheless, researchers have demonstrated that models can sometimes reproduce memorized training information, which is one reason sensitive data deserves additional protection.

If you regularly work with personal or confidential information, it's worth reviewing whether you've allowed your conversations to be used for training.

Here's where to start:

AI platform	Privacy setting to review
ChatGPT	Settings → Data controls → Improve the model for everyone
Claude	Settings → Privacy → Help Improve our AI models
Gemini	Gemini Apps Activity → Keep Activity

Turning off these settings generally prevents eligible new consumer conversations from being used for model training, although the details vary by platform. Gemini also restricts certain features when Keep Activity is disabled.

Importantly, changing your settings today doesn't guarantee that previous uses of your data in completed model training can be undone.

For businesses, the type of account matters. ChatGPT Business and Enterprise, Claude commercial offerings, and eligible Google Workspace AI services generally don't use organizational customer content to train their foundational models by default. Those protections should not automatically be assumed for personal accounts or third-party AI tools.

Step 5: Check whether anyone else could access the conversation

A private conversation stored with an AI provider is different from one shared through a public link or transmitted to an external application.

Our employee should check whether they shared the conversation with colleagues, created a public link, or used an AI integration that sent information to another service.

Public conversation links can allow other people to view shared content. And removing a link doesn't necessarily remove copies that someone else has already saved.

Third-party tools deserve similar attention. Some AI applications, browser extensions, and integrations process information outside the original AI provider's environment. Deleting a conversation in one platform may not delete information already passed to another service.

For organizations, an additional consideration is access through workplace administration and compliance tools. Depending on the platform, authorized administrators may be able to access or export employee conversations under company policies.

These are all reasons to review where information was shared rather than assuming every potential exposure comes from AI model training.

What if you don't remember what you've uploaded?

You might have hundreds of AI conversations, making it difficult to remember which ones contain sensitive information.

Fortunately, you can perform a basic privacy audit.

Start by searching your available conversation history for terms such as confidential, customer information, password, financial records, or project names associated with sensitive work.

Then examine your uploaded files, projects, shared links, connected applications, and memory settings.

ChatGPT, Claude, and Gemini also provide data-export options for eligible accounts. These allow you to download available account information and review past conversations more systematically.

For a more thorough review, exported conversation files can be scanned locally for potentially sensitive details. Keep in mind that no automated scan is perfect, and exported histories may not contain every uploaded file or previously deleted record.

Be careful where you store these exports. A collection of your AI conversations may contain more personal and business information than any individual conversation.

And avoid uploading your complete conversation history into another unapproved AI service simply to have it search for sensitive information.

How to prevent the same problem in the future

The simplest protection is to examine what you're uploading before sending it.

Our employee probably didn't need customer names, email addresses, or confidential pricing to identify trends in the spreadsheet. Removing those fields beforehand would have allowed them to get the same useful analysis with considerably less risk.

You can use this approach with many kinds of information. Replace customer names with anonymous identifiers, remove unnecessary personal details, and share only the portions of documents relevant to the task.

For especially sensitive one-time tasks, ChatGPT offers Temporary Chat and Claude offers Incognito chats, which avoid ordinary conversation-history retention and model training. These features can still involve temporary provider-side retention and are not a substitute for company security requirements.

Businesses should also establish clear guidance about which AI tools employees can use, what kinds of information are appropriate to share, and when approval is required.

An employee using an approved enterprise AI tool under company controls is not necessarily creating the same risk as someone uploading confidential documents to an unapproved personal account.

You can use AI without sharing everything

Back to our employee and that spreadsheet.

If the document contained confidential information, the appropriate response might be to notify IT, delete the conversation and associated files when appropriate, review memory and privacy settings, and verify that nothing was shared externally.

If there were active credentials in the document, those would also need to be changed.

But if the spreadsheet contained nothing more than ordinary, nonconfidential business data, there might be very little to worry about.

The point isn't that sharing information with AI is inherently dangerous. It's that not all information should be handled in the same way.

AI can be enormously useful for working with documents, analyzing information, and solving business problems. Understanding where your information goes—and knowing what controls are available—helps you use those capabilities responsibly.

The best protection is a combination of good judgment before uploading, appropriate privacy settings, and knowing what to do when something sensitive slips through.

---

Disclaimer

The information contained in this communication is intended for limited use for informational purposes only. It is not considered professional advice, and instead, is general information that may or may not apply to specific situations. Each case is unique and should be evaluated on its own by a professional qualified to provide advice specifically intended to protect your individual situation. Kosh is not liable for improper use of this information.