After reading this, you should be able to find the AI facts stored about you, delete the ones you do not want, and shut off training permissions separately in ChatGPT, Gemini, Claude, and GitHub Copilot.
The trap is simple: memory, chat history, and model training are not the same control. Deleting one does not always delete the others, according to Notebookcheck. That matters if your prompts include customer records, internal strategy, source code, contracts, financials, employee data, or anything your company would not want sitting in a third-party account.
Set the target: stop AI tools from retaining sensitive facts
Your goal is not to make AI useless. It is to stop it from keeping facts it does not need.
Start with this split:
| Control | What it affects | Why it matters |
|---|---|---|
| Memory | Saved facts the assistant may reuse later | A deleted chat may not remove these facts |
| Chat history | Conversation transcripts in your account | These may remain until deleted under the provider’s rules |
| Training / model improvement | Whether prompts can be used to improve systems | This is controlled separately from memory in some tools |
Notebookcheck’s central warning is the one most users miss: turning on Memory does not automatically turn on training, and disabling training does not automatically stop an AI assistant from saving memories.
That means your cleanup has to hit both doors.
Before you start: list the accounts that can hold your prompts
Do this before deleting anything. Write down every AI account you use for real work.
Focus on the tools covered by the source material:
- ChatGPT
- Gemini
- Claude
- GitHub Copilot, if you use it for coding
Then mark whether each one is a personal account, work account, Business plan, Enterprise plan, or API setup. Notebookcheck says consumer apps are not the right place for real customer data. For data protection compliance, it points to a Business or Enterprise plan or the API, plus a signed data processing agreement.
Set a hard internal rule before you audit:
- Credentials: never paste passwords or secrets.
- Customer data: do not paste complete records into consumer chat apps.
- Company data: treat contracts, unreleased financials, source code, employee data, and client communications as restricted unless your company has approved the tool for that use.
If your use case touches health data, pair this checklist with our coverage of how regulators are scrutinizing AI-related health data flows in 180-Day Clock Puts AI Health Data Sales on Notice.
Step 1: separate memory, history, and training before you delete
Do not start by mass-deleting chats. First decide which bucket you are cleaning.
Memory is the assistant’s stored profile of you: preferences, recurring projects, work context, and other facts it may carry into future answers.
Chat history is the transcript: the actual conversations stored in your account.
Training permission is whether your content can be used to improve the provider’s systems.
The mistake is assuming one action handles all three. It may not.
“Whether the AI remembers things about you through Memory and whether it learns from your inputs through training are controlled by two different settings.”
That sentence is the operating principle for the rest of this guide.
Step 2: delete saved memories in ChatGPT separately from chats
ChatGPT needs special attention because saved memories are stored separately from chat history.
Notebookcheck says deleting a ChatGPT conversation does not erase saved memories from that conversation. Those memories remain until you delete them individually in Memory settings.
Use this sequence:
- Open ChatGPT settings.
- Go to the Memory controls.
- Review the saved memories.
- Delete individual memories you no longer want.
- Turn Memory off if persistent personalization is not worth the risk.
Use Temporary chats for one-off sensitive topics where available. Notebookcheck says temporary chats are deleted within 30 days, but that does not make them a safe place for passwords, complete customer records, patient data, or confidential client material.
For another ChatGPT-specific privacy angle, see MLXIO’s 11 Missing Images Expose ChatGPT’s ‘Secret Archive’. Treat it as adjacent context, not a substitute for checking your own settings.
Step 3: turn off ChatGPT training in Data Controls
Now handle training separately.
In ChatGPT, go to:
Settings → Data Controls → disable “Improve the model for everyone.”
That setting is about model improvement. It is not the same as deleting memory.
Notebookcheck says ChatGPT keeps chats until you delete them, and after deletion they are generally removed within 30 days. Temporary chats are also deleted within 30 days.
Watch out for a legal wrinkle. Notebookcheck cites a US legal dispute in 2025 that temporarily forced OpenAI to retain even deleted chats. The order did not apply to users in the EU and has since been lifted. The practical lesson: deletion policies are not immune from court orders.
Step 4: use Gemini Apps Activity as the master switch
Gemini works differently. Notebookcheck says Gemini does not have a separate training control. Its data use is tied to Gemini Apps Activity.
Use this sequence:
- Open your Google account controls for Gemini Apps Activity.
- Review stored activity.
- Delete old activity you do not want retained.
- Turn off Gemini Apps Activity if you do not want data used for training.
- Adjust auto-delete if you prefer a shorter or longer retention period.
The tradeoff is clear: disabling Gemini Apps Activity also means Gemini will no longer save your chats.
Retention is also different. Notebookcheck says Gemini deletes your activity after 18 months by default, but you can change this to 3 or 36 months.
Watch the exception: chats reviewed by a human are disconnected from your account and retained for up to three years, even if you delete everything.
Step 5: clean Claude chats and disable improvement use
For Claude, go to:
Settings → Privacy → disable “Help Improve Claude.”
Notebookcheck says deleted Claude chats are removed within 30 days. It also says Claude and Gemini handle associated memories differently from ChatGPT: deleting chats also removes the associated memories.
That makes Claude cleanup simpler than ChatGPT, but not automatic privacy. If you allow training, Notebookcheck says Claude may retain anonymized data for up to five years.
Use this checklist:
- Delete old conversations with personal or business-sensitive context.
- Turn off Help Improve Claude if you do not want chats used for service improvement.
- Use work-approved arrangements for customer or confidential data rather than a regular consumer chatbot.
Step 6: keep customer and company data out of prompts
Settings reduce exposure. They do not turn consumer chatbots into approved systems for confidential data.
Use this prompt rule:
If the raw data would be risky in email, a public document, or an external support ticket, do not paste it into an unapproved AI assistant.
Notebookcheck is direct on this point: anyone entering confidential client or patient data into a regular chatbot has a problem, no matter how many settings are disabled.
For sensitive workflows, use:
- Business or Enterprise plans
- API access
- A signed data processing agreement
- Manually enabled EU data processing, where required and available
Notebookcheck adds one more constraint: True Zero Data Retention is effectively available only through APIs, not in the ChatGPT, Claude, or Gemini apps.
Step 7: check GitHub Copilot if you code with private repositories
If you use GitHub Copilot Free or Pro, check this now.
Notebookcheck says that since April 2026, use of inputs for training has been enabled by default, and it can cover code from private repositories while you are actively working with it.
To reduce that exposure:
- Open your Copilot account settings.
- Find the training or input-use control.
- Disable it if you do not want your coding inputs used for training.
This is especially important if your private repo contains customer logic, unreleased product work, or proprietary infrastructure code.
Step 8: run a quarterly AI privacy reset
Set a recurring calendar reminder. Five minutes per tool is enough for most individual users.
Review:
- Memories: delete stale facts, old projects, locations, client names, or preferences.
- Chat history: remove conversations you no longer need.
- Training controls: confirm opt-outs did not change.
- Account type: check whether you are using a consumer app for work data.
- Copilot settings: confirm defaults still match your risk tolerance.
For teams, assign this to a real owner: IT, security, legal, or operations. Individual habits are not a policy.
Your next move: delete the facts, then close the training loop
Start with ChatGPT Memory settings, because deleting chats there does not erase saved memories. Then check Data Controls for training. After that, review Gemini Apps Activity, Claude Privacy, and GitHub Copilot settings.
The practical rule is simple: delete what the assistant does not need, disable training you do not want, and keep sensitive customer or company data out of consumer AI tools unless your organization has approved the setup.
Key Takeaways
- Deleting a chat does not necessarily delete saved memories or stop model training use.
- Sensitive work prompts can remain exposed if memory, history, and training settings are not managed separately.
- Users should review ChatGPT, Gemini, Claude, and GitHub Copilot settings across personal, work, Business, Enterprise, and API accounts.
