Where locked chats go
On October 20, ChatGPT shipped Locked chats. A locked conversation disappears from your regular chat history, pinned list, and search, but is not deleted. To open it you enter the four-digit account PIN or use device authentication. In plain terms it puts a door on certain conversations: hidden, but not lost.
A four-digit code
Locking is simple: open a chat's more-options menu and pick Lock chat. The thread then vanishes from the home page, search, and pins, returning only after the correct PIN. It is rolling out in batches to Free, Go, Plus, Pro, Business, Enterprise, and Edu on web, iOS, Android, and the new desktop app. For the I talked about something private but do not want to delete case, it beats exporting and wiping.
Where the privacy line sits
Locking is not the same as opting out of training or changing privacy settings. OpenAI says your retention and model-improvement preferences stay unchanged, meaning the lock is an interface-level hide, not a change to backend data handling. If you truly need a segment excluded from model improvement, use the existing data controls; do not assume a lock settles it.
What admins see
In a workspace, a locked chat does not bypass the organization's security policy. It still appears in the OpenAI Compliance Platform's conversation logs, and retention and compliance rules apply as before. Admins do not see the chat vanish; they see it flagged as user-locked with content still inside compliance scope. So for corporate data governance, locking is a personal convenience, not an audit bypass.
When to lock
Not every chat deserves a lock, but the ones that do should not wait. Sales threads with client names and contract details, medical chats with records and health questions, finance Q&A carrying bank statements and tax files, these especially on shared or public devices. Casual what should we eat tonight talk only adds friction if locked, you will tire of the PIN. A simple rule: if someone borrowing your device seeing it would embarrass you or cause trouble, lock it.
Lock is not archive or delete
Locking, archiving, and deleting are three different acts. Archive moves a thread to a secondary list still findable in search; delete is gone; lock keeps it in place but hidden, opened only by PIN. OpenAI shipped lock rather than delete, signaling it wants to preserve your ability to recover while adding a layer of interface cover. If your goal is disappear from the system, lock cannot do it; use data controls or export then delete.
Mobile and desktop details
This push covers web, iOS, Android, and the new desktop client, but rollout timing differs by surface, you may be able to lock on phone while the desktop lacks it. After locking, signing in on another device still asks for the PIN; switching machines does not auto-open. On desktop, if PowerPoint or Keynote is installed, lock and export abilities do not interfere. If an endpoint lacks the option, check whether the client is the latest version.
One counter-example
Some assume a lock means the content is safe, which is a mistake. Lock is an interface-level hide; it does not change backend data handling or opt out of model-improvement preferences. To keep a segment out of training, use the official data controls. Treat the lock as a drawer lock, defense against a casual glance, not against an insider or the platform. Knowing that boundary stops you overrating its protection.
Lock is not encryption
State it plainly: locking is an interface-level hide, not encryption, and not an opt-out of model training. Some assume a lock means no one, including the platform, can see it, which is wrong. OpenAI still handles locked content under its standard data terms, and it appears in compliance logs. The lock defends against someone borrowing your device, not against the platform or a lawful request. Worse, anyone who knows your PIN can still open it, so do not set the PIN to a guessable birthday. If your threat model is account theft or compliance access, a locked chat gives near-zero protection; you need two-factor auth and device encryption underneath. Know the defended party before trusting the tool, and before overrating its safety.
What this means for youIf you also sign in to ChatGPT on a shared or public machine, locking is a cheap self-protection move: lock threads with client data, health, or finance so a borrower of your device cannot browse them. But remember it hides without deleting and does not change training preferences; truly sensitive content should go through data controls. Treat it as a drawer lock, not a safe.