新智元报道
10月20日，ChatGPT聊天框里多了一把锁。
OpenAI推出「锁定聊天」功能：设置4位账户PIN码后，用户就能把指定对话藏起来。
在支持的移动设备上，还可以刷脸、按指纹，或使用设备解锁方式访问。
那些舍不得删、又不想被外人撞见的对话，这回有了一个专属空间。
不过，聊天藏起来之后，还有几个问题需要搞清楚：
AI还会记得你们聊过的秘密吗？这些内容会不会被拿去训练模型？万一忘记PIN码，聊天记录会不会被清空？
对话上了锁，并不意味着这些问题都解决了。
在ChatGPT里，隐藏对话、管理记忆、控制训练用途，各有各的规则。
所以，先别急着把所有秘密都往里塞，弄清这把锁的边界后再用。
很多人的AI聊天框，其实已经变成了半个日记本。
方便是方便，隐私也更容易泄露了。
想象一下：投屏开会，侧边栏里挂着你昨天咨询健康问题的记录；把手机递给朋友看照片，却不小心切到了辞职信草稿；想给家人准备节日惊喜，聊天列表里却全是礼物清单和比价记录……
上面这些私密对话，你还想留着继续用，所以不能一删了之。
但共享屏幕或借出设备时，一个不经意露出的聊天片段，又可能让你当场「社死」。
「锁定聊天」给出的刚好是这样一个折中方案，而且它的操作并不复杂。
打开单条聊天的菜单，选择「Lock chat」。
按提示设置并确认4位PIN码。
锁定后的对话可在「Locked chats」中找到。
上锁后，该对话会从常规历史列表、置顶区和普通搜索结果中隐藏，转而收进侧边栏「项目」下的「锁定聊天」专区。
后续只要验证身份，随时可以继续往下聊天。
根据OpenAI帮助中心，锁定聊天正面向Free、Go、Plus、Pro、Business、Enterprise和Edu套餐逐步开放，免费用户也能用。
目前功能正逐步覆盖网页端、iOS、Android，以及新版macOS和Windows桌面应用。
当然，这把锁也不是万能的。
临时聊天、归档聊天、群聊，以及带有计划任务、会议录音等属性的对话，目前不支持上锁。
通过他人分享链接创建的聊天、Codex和本地工作对话也有相应的限制。
而医疗专用的ChatGPT for Healthcare则直接不提供该功能。
把聊天锁进列表，所有设备上的画面就会瞬间消失吗？
没那么简单。
OpenAI在文档中提到：
网页端在点击上锁后，当前打开的标签页仍可能继续显示该对话框；如果您同时在平板或另一台电脑上登录并打开了该页面，也必须手动刷新，才能彻底隐藏标题和已加载的内容。
因此，当你准备投屏或把手机递给别人之前，千万别盲目相信「已上锁」三个字，顺手关掉当前页面、刷新一下列表，才算真正完成保险。
还有一个容易踩坑的细节：上锁不会让此前创建的分享链接失效，拿到链接的人仍能查看。
锁定期间新增的内容不会同步到旧链接；如果想撤回访问权限，还需单独删除这条分享链接。
OpenAI也说明，上锁只是增加一道查看验证，并不提供端到端加密，也不会改变数据留存和模型改进设置。
它更像一张「屏幕防窥膜」：能减少旁人看到聊天的机会，内容仍按原有规则处理。
看到「上锁」二字，很多人的第一反应是既然锁了，这些对话肯定不会被AI记住，也不会被拿去训练模型吧？
并非如此：把「界面隐藏」直接放在「数据隔离」上，是最大的认知误区。
先看「记忆」功能。
如果你开启了记忆功能，锁定聊天仍可能调用普通聊天中已保存的记忆，让你不用重新介绍自己。
同时，你在锁定聊天里了解到的新信息，可能被AI记住，并共享给其他锁定聊天使用。
好消息是，在锁定状态下产生新的记忆，不会泄露到普通聊天里。这意味着你的争论话题有一个独特的记忆自留地。
ChatGPT的「记忆摘要」页面，展示已保存的个人背景信息；选中内容后，可点「Make a correction」（更正）进行修改。
但这里还有一笔「旧账」：
锁定前已经被AI记住的信息，不会因为对话上锁而自动清除。
比如，某件私事此前已在普通聊天中被保存为记忆，之后即使锁住那段对话，这条记忆仍可能用于普通聊天。
如果某天你把对话永久解锁，整段聊天内容又可能按照账户和工作区设置，被用于形成普通记忆。
但锁定期间单独形成的记忆，不会随着解锁自动并入普通记忆。
一句话总结：旧账不会因为补锁而记勾销，新账则严格实行「专账专用」。
再看「训练」数据。
锁定聊天不会改变原有留存政策，也不会自动关闭模型改进；对话是否用于训练，仍由账户的数据控制设置决定。
这里必须厘清一个概念：记忆影响AI如何个性化回复你；模型改进设置则决定你的对话是否会用于改进模型。
想停止新对话用于模型训练，需前往「设置—数据控制」，手动关闭「为大家改进模型」。
数据控制中的模型改进开关。
关闭后，新的对话才不会用于模型训练。
但官方补充：即使关闭了模型训练开关，如果你通过点赞、点踩等方式主动提交反馈，与这条反馈相关的整段对话，仍可能被用于训练模型。
如果你想让对话不进入历史、不创建或更新记忆，也不用于模型训练，可以使用「临时聊天」。
出于安全审查目的，对话仍可能最多保留30天。若之后转存为普通聊天，则按普通聊天规则处理。
说白了，在ChatGPT中，这几项职能各司其职：
怕熟人扫一下屏幕看到，用「锁定聊天」；怕AI在日常对话里哪壶不开提哪壶，管好「记忆设置」；怕数据被拿去喂大模型，去关「数据控制」里的训练开关。
在使用锁定聊天的过程中，有一个按钮的杀伤力极大：重置「PIN码」。
如果你只是想换一下旧PIN码，可以选择「修改PIN码」，聊天记录会保留。
即使忘记了PIN码，也不用立刻慌张：忘记密码本身，不会自动删除聊天或文件。
真正触发删除的，是你选择「重置PIN码」，并确认执行。
根据OpenAI的规则，重置会永久删除账户中的全部锁定聊天（不只是当前这一条）、仅保存在锁定聊天中的记忆，以及仅受旧PIN码保护的文件。
这是一次不可逆的清仓，连OpenAI客服也无法恢复被删除的内容。
因此，按下确认键之前，先看清删除范围，确认其中是否有需要保留的重要对话和文件。
如今，我们和AI聊的，早已不只是工作。
你用它写周报、查代码，也会和它讨论健康、感情，以及那些难以做出的决定。
有些问题问完就过去了，有些对话却留下了重要的思考，值得以后反复回看。
锁定聊天，可以让这些内容可以留着，同时减少被旁人看到的机会。
但我们也需要知道：聊天锁住了，旧记忆还在吗？分享出去的链接还能打开吗？重置PIN码，会删掉哪些对话和文件？
这些后果，应该在操作前讲清楚，不能等用户丢了内容，才让他们去帮助中心找答案。
参考资料：
编辑：元宇
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
On October 20, there was a lock in the ChatGPT chat box.
Open AI launched the "Lock Chat" feature: with a 4-bit account PIN code, the user can hide the specified conversation.
Supported mobile devices can also be accessed by painting, fingerprinting or unlocking devices.
There is an exclusive space for dialogue that cannot be deleted and that does not want to be encountered by outsiders.
However, there are a few questions that need to be clarified after the conversation is hidden:
Does AI remember the secrets you talked about? Will these be used to train models? If you forget the PIN code, will the chat be emptied?
The fact that the dialogue is locked does not mean that these problems are resolved.
In ChatGPT, there are rules for hiding dialogue, managing memory, controlling the use of training.
So, don't rush all the secrets in, find out where the lock is.
Many of the AI chat boxes have actually become half a diary.
Accessibility is convenient and privacy more easily disclosed.
Imagine: a meeting on the screen, with records of your consultations on health issues yesterday in the sidebar; handing over the cell phone to a friend for a photo but accidentally cut a draft letter of resignation; trying to surprise family with a holiday, the chat list is full of gifts and price records...
You can't just delete these private conversations.
However, when sharing screens or lending devices, a non-intuitive chat clip may cause you to "scramble."
"Lock Chat" offers such a compromise, and its operation is not complicated.
Opens a single chat menu and selects "Lock Chat".
Set and confirm 4-bit PIN code by hint.
The post-locking conversation can be found in Locked Chats.
When locked, the dialogue is hidden from the general historical list, tops and general search results, and is then placed in the " lock-and-talk" section under the "item" sidebar.
Following identification, you can continue to talk down.
According to the Open AI Help Centre, the locked chat is gradually open to Free, Go, Plus, Pro, Business, Enterprise and Edu packages, which are also available to users free of charge.
The current functionality is gradually covering the end of the web page, iOS, Android, and the new MacOS and Windows desktop applications.
Of course, this lock is not everything.
Temporary chats, archived chats, group chats and conversations with characteristics such as planned tasks, audio recordings of meetings, etc., do not currently support locking.
There are corresponding limitations to chats created by sharing links, codex and local work dialogues.
That function is not provided directly by ChatGPT for Healthcare, a medical specialist.
Lock the chat in the list, and all the images on the device disappear?
It's not that simple.
Open AI mentions in its document:
The currently opened tab may continue to display the dialogue box after the page is locked by clicking; if you log in and open the page on the tablet or another computer at the same time, you must also refresh the page manually in order to completely hide the title and loaded contents.
So, before you're ready to go on screen or hand your cell phone over to someone else, don't blindly believe the word "locked" and turn off the current page and refresh the list.
There is also an easy-to-penetrate detail: locking does not invalidate previously created shared links, which can still be viewed by those who get them.
What is added during the lockup will not be synchronized with the old link; if you want to withdraw access to permission, you need to remove the share link separately.
Open AI also states that the locking only adds a single view to the validation, does not provide end-to-end encryption and does not change data retention and model improvement settings.
It's more like a screen-proof film: it reduces access to chat, and the content remains the same.
Seeing the word "lock," the first reaction of many people is that since it's locked, these conversations will not be remembered by AI, nor will they be used to train models?
This is not true: putting the "interface hidden" directly on "data isolation" is the biggest cognitive error.
It's the memory function.
If you turn on the memory function, lock-in chats may still call on memory already saved in regular chats, so you don't have to introduce yourself again.
At the same time, the new information you learn from locked chats may be remembered by AI and shared with other locked chats.
The good news is, new memories are created in lock-in, not leaked into regular chats. This means that your argument has a unique memory.
ChatGPT's "Memory summary" page displays saved personal background information; once selected, changes can be made by clicking "Make a correction".
But here's another "old bill":
Information that has been remembered by AI before locking in, does not automatically remove it because the dialogue is locked.
For example, a private matter had previously been preserved as a memory in a regular chat, and then, even if the conversation was locked, it could still be used for a regular chat.
If one day you unlock the conversation permanently, the whole chat may be used to create ordinary memories, according to the account and workspace settings.
However, individual memory formed during locking does not automatically incorporate ordinary memory with unlocking.
In a nutshell, old books are not written off for re-locking, and new ones are strictly "earmarked".
And look at the "training" data.
Lock-up chats do not change the old retention policy or automatically close model improvements; whether dialogue is used for training is still determined by the account ' s data control settings.
Here's a concept that needs to be clarified: memory influences how AI personally responds to you; model improvement settings determine whether your conversation will be used to improve the model.
If you want to stop the new conversation for model training, you need to go to Setup-Data Control and manually close the "Model Improvement for All."
Modelling improvement switches in data control.
Once closed, the new dialogue will not be used for modelling training.
However, the official addition: even if the model training switch is closed, the entire section of the dialogue related to this feedback may still be used for the training model if you are willing to provide feedback by means of a compliment, a step.
If you want to keep the conversation out of history, not create or update memories, and not use model training, you can use "temporary chat".
For security clearance purposes, the dialogue may still be held for up to 30 days. If you later switch to normal chat, you follow the normal chat rule.
In ChatGPT, the functions are divided into:
Fearing that people who know each other will be able to scan the screen, they will use "lock and chat" or that they will not mention the pot in their daily conversations, they will be able to fix the "memory set" or that they will feed the data to LLM and turn off the training switch in "data control."
One of the buttons is extremely lethal when using locked chats: to reset the PIN code.
If you just want to change the old PIN code, you can choose "change the PIN code" and chat records will be kept.
Even if the PIN code is forgotten, there is no need to panic immediately: forgetting the password itself does not automatically delete chats or files.
The real trigger for deletion is your choice to "reset PIN code" and confirm execution.
According to Open AI ' s rules, the reset permanently removes all locked chats from the account (not just the current one), saves only the memories in locked chats, and only files protected by the old PIN code.
This was an irreversible clean-up, and even Open AI's client service could not restore the deleted content.
Therefore, before pressing the confirmation key, the scope of the deletion is clearly identified and whether there are important dialogues and documents that need to be retained.
Now, we're talking to AI about more than work.
You use it to write weekly newspapers, check codes, talk about health, feelings, and decisions that are difficult to make.
While some of the questions have passed, some of the dialogues have left important reflections that deserve to be reviewed again and again.
Locking up chats allows these to be kept while reducing the chances of being seen by others.
But we also need to know: chats are locked and old memories are still there. Can the shared link open? What dialogue and documents would be deleted if PIN codes were recoded?
These consequences should be made clear in advance of the operation. Users should not be asked to go to help centres to find answers until they have lost their content.
References:
Editor: Won-woo
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.