Privacy controls have different purposes
A chat disappears from the sidebar. A memory is no longer mentioned. A voice assistant is disabled.
It can feel as though the information has been dealt with. Several questions need separating: is it merely hidden, is it still stored, can it improve a model, and are copies or derived information held elsewhere?
There is a current reason to ask: OpenAI introduced locked chats on 20 October 2026, and Google updated its Gemini Apps Privacy Hub on 19 October. Both show why one privacy setting cannot govern every use of data. [1][4]
Rather than asking which provider is private, I would ask which particular use of data I want to prevent or limit.
Five actions that should not be confused
OpenAI describes training, history and personalisation as separate controls. Removing access to a connected service should not be confused with deleting chats that already contain its information. [3]
The table helps choose a measure that fits the aim, not just define terms.
Hiding a confidential chat during screen sharing is different from removing information no longer needed.
| Setting or action | The important question |
|---|---|
| Hide or lock a conversation | Who can see and open it in the interface? |
| Switch off model training | Can new content be used to improve models? |
| Manage memories | What information can influence later answers? |
| Disconnect an application | Can the assistant access that service in future? |
| Delete content | Which stored information is actually removed? |
What locked chats in ChatGPT do
Locked chats leave regular history and conversation search. Opening them requires an account PIN or supported device authentication.
OpenAI makes clear that locking does not provide end-to-end encryption, alter retention or training settings, or remove previously saved memories. Existing shared links need separate attention too. [2]
The feature can help prevent unwanted glances at the interface. It is not a private vault stored exclusively on your own device.
Another warning matters: resetting the account PIN permanently deletes locked chats and content protected only by the previous PIN. A PIN reset is not a harmless troubleshooting step. [2]
Read the stated consequences carefully before taking that action.
Turning training off is not deletion
When Improve the model for everyone is disabled, OpenAI says new conversations are not used for model training. They may still appear in history. Voluntary feedback is an exception to consider: its associated conversation may be made available for training. [3]
That separation can suit someone who wants to keep their chats. Removing information requires checking stored conversations and potentially other sources too.
Deleting the original chat does not always remove a separately saved memory. OpenAI also identifies files and connected apps as sources to consider, depending on where information came from. [3]
Start with the aim, then check the relevant controls instead of expecting one universal privacy switch.
Temporary chats are not necessarily traceless
While they remain temporary, OpenAI says these chats do not appear in history, create memories or improve its models. They may still be retained for up to 30 days for safety.
Before starting, users can choose whether existing personalisation is used. Saving a temporary chat makes it a normal conversation governed by the usual settings. [3]
A conversation can therefore create no new memories while still using existing personal information.
For a deliberately unpersonalised question, I would check that starting option rather than rely on the temporary label alone.
Gemini: distinguish activities, training and retention
Google's Privacy Hub, updated on 19 October 2026, lists an adjustable default automatic deletion period of 18 months for Gemini activities.
When activity saving is disabled, future chats are not used for model training unless feedback is submitted. Google still describes 72 hours of storage for those conversations and temporary chats. Content reviewed for training can be retained separately from the account for up to three years. [4]
That does not mean every type of Gemini data always disappears after 72 hours. The data type and its use matter.
Google also says deleting Gemini activities does not automatically delete information kept by other Google services under their own rules. [4]
When an assistant connects several applications, those services belong in the review as well as the conversation history.
The Gemini app and Gemini for Google Home are not identical
Do not automatically apply a smartphone or browser application's privacy details to a household voice assistant.
Google publishes separate Home information. Its introduction says personal data is not used to train generative AI models outside Google Home. [6]
That is a specific, limited statement, not a promise that no processing or improvement takes place within Home.
First identify the service: Gemini app, Google Home, a connected application or a managed business account.
Alexa+: review recordings, attachments and operation separately
Amazon describes controls in its privacy portal and the Alexa app for interactions, including voice recordings, shared attachments and recording retention. [5]
There is also a separate help page for reviewing and deleting uploaded attachments. [5]
A voice recording, an uploaded document and an action in a connected application are different kinds of content. Audio alone is not enough to assess the service.
My decision to disable Alexa+ concerned which assistant I used. I would not infer that all earlier data was removed too.
If I want both outcomes, I need to check both the assistant's status and the stored information.
A document may contain more information than the task needs
Suppose I want an invoice summarised. Explaining its line items may require neither a private address nor bank details.
I would first supply a redacted copy or only the relevant sections. That reduces the information whose later use and deletion I need to consider.
For one appointment summary, I would also ask whether connecting a full calendar is necessary. Convenience alone would not justify broad access for me.
These decisions can be more effective than trying to trace every later use of the data.
Before the next confidential task
I would check the account first. Personal and managed work accounts can have different rules. OpenAI describes organisational controls and no model training by default for managed products. [3]
Then consider the task: which information is needed, which functions are authorised, should the chat remain findable, and may it contribute to personalisation?
Afterwards, review unnecessary files, conversations, shared links and connections separately. A connection created for one task does not have to remain indefinitely.
Do not forget your own backup. Before deleting documents, make sure any original still needed exists outside the AI service.
One switch does not make a provider the winner
A useful chat lock does not establish superiority on every privacy issue. Another service might offer short retention for one mode while connected applications retain their own records.
Compare the same concrete task: what is transmitted, which settings apply, and what steps are required afterwards?
Only then can we judge which product fits our conditions better.
A polished padlock icon is no more sufficient than a broad statement that inputs are not used for training.
Conclusion: privacy needs understandable states
New features can be useful. Locked chats restrict unwanted views, temporary conversations limit certain uses, and controls for training, memories and connections offer additional choices.
But they solve different problems.
Hidden is not deleted. No training does not mean no storage. Disabling an assistant does not automatically clear the account.
Understanding those distinctions makes the controls more useful. For me, good privacy starts with a clear answer to what a setting changes, and what it does not.
Scope of the information
This article reflects documented features and provider statements, not an independent technical audit of internal processing or individual legal advice. Account type, region and organisational policies can make a difference.
Sources and further reading
Original sources checked on 20 October 2026. Manufacturer statements are not independent comparative measurements; availability and settings can depend on the account.
[1] OpenAI: ChatGPT release notes (locked chats, 20 October 2026)
Read source ↗
[2] OpenAI: using locked chats in ChatGPT
Read source ↗
[3] OpenAI: data controls and memory
Read source ↗
[3] OpenAI: data controls and memory
Read source ↗
[4] Google: Gemini Apps Privacy Hub (19 October 2026)
Read source ↗
[5] Amazon: Alexa+ privacy and uploaded attachments
Read source ↗
[5] Amazon: Alexa+ privacy and uploaded attachments
Read source ↗
[6] Google: getting started with Gemini for Home
Read source ↗
Related articles
ChatGPT and Gemini for everyday work
Read source ↗
Alexa+ and Gemini for Google Home
Read source ↗