Read What an AI Feature Doesn't Protect

Originally written in Traditional Chinese: AI 產品最該讀的，是它沒保護什麼

Last week ChatGPT got a lock. Many people will assume locked means safe. The fine print lists several things it does not protect.

On October 20, OpenAI began rolling out Locked Chats. You can lock selected conversations with a four-digit PIN, or Face ID or Touch ID. Locked chats leave your regular history, pinned chats, and search results, and they get their own separate memory; new messages there don't update your regular memory. The rollout is gradual, so not everyone has it yet.

Now the other half. It is not end-to-end encryption or a separate encrypted vault. It does not change how long chats are retained. It does not change the "Improve the model for everyone" setting; if that was on, locking does not turn it off. What ChatGPT remembered before you locked a chat is not removed. And resetting a forgotten PIN permanently deletes everything that was locked.

None of this makes the feature bad. Keeping sensitive conversations out of sight on a shared computer is useful, and the feature does what it says. The problem is our minds: the word "lock" invites us to fill in promises it never made.

So here is the line I want to leave with you: the most important part of any AI feature is not what it protects. It is what it doesn't.

In my course I teach a three-layer verification habit. The most dangerous thing about AI is being confidently wrong, so I ask two questions of every answer: "What is your source?" and "Under what conditions is this not true?" Those same questions work on product announcements: "Under what conditions does this feature not protect me?"

For parents: OpenAI's parental controls FAQ states that linking accounts does not give a parent access to a teen's conversations. Whether Locked Chats applies to teen accounts has not been addressed in the coverage I found. Either way, scrolling through history was never the best approach. A better one is a weekly invitation: "Show me one conversation where AI helped you most."

A three-minute exercise for anyone: open ChatGPT settings, check whether "Improve the model for everyone" is on, and review what Memory has stored about you. Then finish this sentence: "I thought it ___, but actually it ___." Educators can turn the feature into an AI literacy task: three things it protects, three it doesn't. Professionals on Enterprise or Edu plans should know: admins can't unlock your chats, but locked chats still appear in the conversation logs available through OpenAI's Compliance Platform.

The lock is real and useful. Behind that locked door, a few other doors are still open.

Get the Human-First practice card

Four steps, ten minutes, with copy-ready AI prompts. The card and my weekly letter are currently written in Traditional Chinese.

👉 Subscribe free to 碩學苑 (Traditional Chinese)

Follow in English

I also post in English on LinkedIn.

👉 Follow Haishuo Lee on LinkedIn

References