By Roman Aglaganov
ChatGPT now lets you lock a conversation behind a PIN or Face ID. It is a good feature, and OpenAI deserves credit for being honest about it: its help page says plainly that a locked chat is not end-to-end encrypted, is kept for just as long, and can still be used for training if your settings allow.
The trouble is that most people will not read the help page. They will see a padlock and assume their secrets are safe.
Two kinds of privacy
There is privacy from the people around you, and privacy from the company that holds your data. Locked Chats solves the first. That matters: people use chatbots for health fears, debts and relationships, often on shared devices. But the second is the one that decides what happens after a data breach, a subpoena or a change of policy.
A padlock icon has meant the second kind for years. In messaging apps it signals encryption that even the provider cannot open. Borrowing the symbol for a screen lock risks teaching people the wrong lesson.
The second kind is possible
Meta says it is building a “Muse Confidential VM” meant to keep even Meta from seeing users’ content. Apple runs some AI requests on what it calls Private Cloud Compute, servers it says even Apple cannot read. These approaches have limits and need outside checking, but they show the industry knows how to aim higher.
What would help
Two simple changes would make locks honest. First, locking a chat should switch off training on it by default, whatever the global setting. Second, the lock screen itself, not a help article, should say in one line what the lock does and does not protect. Then offer the stronger kind of privacy for those who want it. People are telling chatbots things they tell no one else. They should know exactly who else can read them.
