ChatGPT now lets you lock your chats, adding privacy inside your account

ChatGPT now lets you lock your chats, introducing a privacy boundary inside an account that otherwise mixes personal questions with everyday work. OpenAI began rolling out the feature on October 20, 2026. The change addresses an ordinary problem: showing someone a useful conversation without exposing everything else you have discussed.

The immediate benefit is control over what another person sees. The larger issue is what users understand that control to mean. Keeping a conversation away from a borrowed screen differs from limiting how a cloud service processes its contents.

WhatsApp established a familiar precedent with locked conversations. ChatGPT adds another complication because an assistant can remember information and use it in later answers. That makes the central question broader than whether a conversation opens: which privacy boundary does the lock actually create?

ChatGPT now lets you lock your chats without deleting them

Locked Chats introduces an access barrier for conversations users want to keep. It gives saved discussions a different place inside the account rather than requiring users to discard them.

According to OpenAI’s locked chats guidance, selected conversations move outside regular history, pinned chats, and ordinary search. Access requires a four-digit account PIN or supported device authentication. The feature is rolling out gradually, so account availability can differ.

The important product decision is preserving the conversation. Consider a hypothetical user preparing for a difficult family discussion. They might want to revisit an earlier exchange, refine their thoughts, and keep working through the subject. Deleting the conversation would undermine that continuity.

Another hypothetical user might be organizing a surprise trip. They still want their travel preferences and planning notes available when they return. They simply do not want the conversation appearing while showing a family member something unrelated.

These examples explain why a conversation lock deserves a place alongside broader privacy settings. The desired outcome is selective visibility during normal use. The user wants to keep using the assistant without treating every saved conversation as equally suitable for an audience.

That distinction also helps explain the timing of the feature’s relevance. A conversational assistant can hold unfinished thoughts alongside finished answers. Users might happily share a recipe while considering an adjacent personal discussion completely private.

The interface has to accommodate both expectations. An account can belong to one person while its screen becomes temporarily visible to someone else. Treating account access as the only meaningful boundary leaves that everyday situation unresolved.

The announcement does not establish how frequently people encounter this problem. It also does not supply evidence that locks will change adoption or engagement. Those outcomes require observation after rollout.

What changed is more specific. OpenAI has added a way to retain selected conversations behind an additional access step. That creates a clearer distinction between saving information and making it readily visible during ordinary browsing.

For readers, this is the useful starting point. Judge the feature against the exposure it addresses, then examine the other privacy questions separately. A narrowly defined benefit can still be valuable when the boundary matches the situation.

A borrowed screen creates a different privacy problem

The feature puts pressure on the assumption that an unlocked account should reveal every saved conversation. That assumption becomes awkward when personal and shareable material occupy the same interface.

Imagine showing a coworker an answer about spreadsheet formulas. The coworker does not need access to your private planning, personal questions, or unfinished writing. Yet handing over the device creates an audience change without changing the signed-in account.

A conversation lock can make that audience change easier to manage. Its practical value lies in giving users a reason to distinguish material before the moment of sharing. Privacy becomes part of organizing the account, rather than an emergency response afterward.

There is a useful historical reference in messaging. Meta introduced WhatsApp’s Chat Lock in May 2023. It described moving protected conversations into a separate folder behind a device password or biometric authentication.

Meta also highlighted people who occasionally share phones with family members. The parallel is direct: device ownership does not guarantee that only the owner sees the screen. Conversation-level controls address that gap.

The comparison should stop short of treating messaging and AI assistance as equivalent. WhatsApp’s announcement describes a messaging feature. ChatGPT’s feature operates inside an assistant where saved context can affect future responses.

The transferable idea is selective access within a familiar application. The implementation and surrounding data practices still require separate examination. Similar names do not establish identical protection.

For OpenAI, the pressure is therefore partly about communication. A lock symbol carries expectations before anyone reads documentation. Users should be able to understand the feature’s purpose without interpreting it as a promise about every aspect of privacy.

For competing assistants, the announcement raises a design question rather than proving a competitive disadvantage. Can people show useful work while keeping unrelated conversations out of view? Different products can answer that through different controls.

There is no evidence here that another assistant has lost customers because of this launch. Nor does the announcement demonstrate that conversation locks determine product preference. Those would be separate claims requiring separate sources.

The more defensible competitive observation concerns usability. A personal assistant needs an understandable answer when users ask which parts of their account are appropriate to display.

Knowledge workers face the same organizational question across their tools. A personal knowledge base can collect material with very different intended audiences.

Putting those materials together makes retrieval convenient. Deciding what belongs on a shared screen remains a separate responsibility. Chat locks make that distinction more visible inside the assistant itself.

Saved conversations and temporary conversations serve different goals

Keeping a discussion private while preserving it is a different goal from starting a conversation that stays outside history. Confusing those intentions can lead users toward the wrong control.

OpenAI’s current temporary chat guidance says temporary conversations stay outside history unless users save them. While temporary, they do not create or update memories and are not used to improve OpenAI models.

The documentation also distinguishes personalized and unpersonalized temporary conversations. Personalized sessions can use existing context; unpersonalized sessions exclude that personalization. Saving a temporary conversation converts it into a regular chat governed by applicable account settings.

That comparison matters because privacy is often discussed as though users always want less persistence. Sometimes they want the opposite. They want an assistant to preserve a useful exchange while limiting its visibility to nearby people.

Consider a hypothetical career discussion. A user might return repeatedly to draft questions, compare options, or revise an application. The ability to continue the conversation is part of its value.

A temporary conversation answers a different intention when the user does not want that exchange saved in ordinary history. Neither intention is inherently more sensible. The correct choice depends on what the person wants to happen afterward.

This is why the announcement should be understood as an additional category of use. It gives people another way to express how a conversation belongs in their account. That does not make other controls obsolete.

The distinction also changes how a product should explain privacy. Labels organized around technical features can force users to translate their intentions into settings. Labels organized around outcomes can make the decision easier.

A person preparing to share a screen is thinking about visibility. Someone asking a question they do not want retained in history is thinking about persistence. Someone concerned about training is thinking about how submitted information gets used.

Those are related concerns, but they are not interchangeable. A useful interface should make the relationship legible without suggesting that selecting one option automatically settles the others.

There is also a practical lesson for reporting. Calling every private-looking mode “incognito” obscures meaningful differences. Readers need to know whether a feature changes access, history, personalization, or data use.

The strongest interpretation of this launch stays close to the intended outcome. ChatGPT now lets you lock your chats when keeping them available matters. Temporary conversations remain relevant when the desired relationship with saved history differs.

That gives users more choice. It also increases the importance of precise language, because more privacy controls create more opportunities for mistaken assumptions.

ChatGPT memory makes privacy more than a visibility setting

A hidden conversation and forgotten information are separate outcomes. An assistant’s ability to reuse context makes that separation especially important.

OpenAI’s memory controls describe Memory as information used to personalize responses. With Memory enabled, locked chats can use relevant regular memories. New messages in locked conversations do not personalize regular chats.

The same documentation says locking does not remove information remembered beforehand. After permanent unlocking, the full conversation can become available for regular Memory where settings and workspace policies permit.

That behavior introduces a timeline into privacy decisions. A user’s expectations depend partly on when information entered the system and when they changed the conversation’s status. The visible state alone does not tell that entire story.

Consider a hypothetical user who previously discussed moving to another city. Later, they decide that the discussion belongs behind a lock. Their intention might include keeping the subject out of future unrelated answers.

The distinction between conversation access and remembered context becomes relevant immediately. The user needs to understand which action changes which outcome. Otherwise, the interface can look private while their expectations remain broader than the action they took.

This is an analytical concern, rather than evidence of a demonstrated malfunction. The documentation describes separate behaviors. The question is whether people understand those behaviors well enough to make informed choices.

Memory also explains why borrowing the messaging metaphor has limits. A conversation with an assistant can contribute to a continuing relationship with the product. The privacy question extends beyond whether somebody can open the original thread.

A user might care about the appearance of a sensitive topic in a later recommendation. They might care about whether unrelated work receives personal context. They might want continuity within a private discussion while limiting its influence elsewhere.

These intentions require understandable boundaries around context. More broadly, they suggest that personal AI products need ways to explain information movement, not merely information storage.

For developers, this is a question of making state visible. Users should be able to understand what an action changes before relying on it. Good documentation helps, but the decisive moment happens inside the product.

For buyers, it is a question of evaluating the whole experience. A demonstration that a conversation requires authentication does not answer every question about personalization. Both can matter to the same user.

For readers, the practical takeaway is to identify the desired result before judging the control. Keeping a discussion away from someone viewing your screen is a clear goal. Preventing earlier information from influencing later answers is another.

The announcement is meaningful because those goals now sit closer together in the interface. Its success will depend partly on whether users can distinguish them without needing to study the system’s history.

The lock does not rewrite the underlying data relationship

Conversation access and the service’s handling of information remain different privacy questions. Evaluating the feature fairly requires examining both without collapsing them into one claim.

OpenAI explicitly says locking does not introduce end-to-end encryption, where only conversation participants hold the means to read messages. It also leaves retention and the model-improvement setting unchanged.

That limitation defines the feature’s scope. It does not eliminate the usefulness of controlling access on a borrowed screen. It does mean users should avoid treating a lock as a universal answer about sensitive information.

OpenAI’s consumer data controls explain that personal-service content can support model improvement depending on settings. The company also describes limited authorized access for purposes including support, security, legal matters, and permitted model improvement.

Those are service-level considerations. They concern the relationship between the user and the provider, rather than the relationship between the user and someone nearby.

For a hypothetical surprise party discussion, nearby visibility might be the main concern. For confidential work, the organization’s rules for submitting information can be decisive. The same interface feature does not resolve both situations.

Managed workspaces make that distinction explicit. OpenAI’s enterprise privacy controls say organizational retention and compliance requirements still apply. Locked conversations remain included in message logs available through the OpenAI Compliance Platform.

Employees therefore need to distinguish personal discretion from organizational governance. A conversation can deserve privacy from casual viewing while still remaining subject to workplace policies.

That is not an inherent contradiction. Organizations can have legitimate recordkeeping requirements while employees need to avoid exposing unrelated material during a presentation. These requirements operate at different levels.

The skeptical question is whether the product communicates those levels clearly enough. A visible lock can be interpreted more broadly than a policy statement. The burden should not fall entirely on users to reconcile the two.

Recovery also deserves attention. OpenAI says resetting a forgotten PIN permanently deletes protected locked-chat content. That makes recovery communication part of the privacy experience, rather than a minor settings detail.

A user choosing access protection also needs to understand the consequences of losing access. Clear warnings matter because changing a known credential and resetting a forgotten one can imply different outcomes.

The available sources establish documented behavior, not independent security testing. They do not demonstrate resistance to every attack or prove that users will understand each boundary correctly.

The appropriate judgment is therefore limited and practical. This launch adds control over conversation access. Decisions about what to submit still require attention to data settings, organizational requirements, and the sensitivity of the material.

What to watch as Locked Chats reaches more users

The next useful evidence concerns availability, user understanding, and the broader product response. Those signals will reveal more than treating the announcement alone as proof of stronger trust.

First, watch how consistently the feature reaches eligible accounts and devices. A gradual rollout can make early experiences uneven. Readers should distinguish missing availability from a feature behaving differently than documented.

The meaningful evidence would be clear availability guidance supported by reproducible experiences across supported environments. That would strengthen the case that users can make the control part of an ordinary routine.

Persistent uncertainty about where the option exists would weaken that case. A privacy feature becomes less useful when people cannot tell whether they can rely on it during their next presentation or device handoff.

This does not require assuming the rollout will encounter problems. It identifies what successful delivery would look like from a user’s perspective.

Second, watch whether users understand the relationship between access, memory, and data use. Product wording, contextual explanations, and carefully documented user studies would be useful evidence here.

A person who can explain the control’s purpose accurately is better positioned to choose it appropriately. That would strengthen the argument that the feature improves practical control over personal conversations.

Repeated confusion would point toward a communication problem, even if the underlying behavior matches the documentation. Product correctness and user understanding are connected, but they are not identical measures.

Independent evaluations would be particularly useful if they examined ordinary decisions rather than asking people to repeat terminology. Can someone choose the right setting for a borrowed screen, an unsaved question, or confidential workplace material?

Those are distinct scenarios with distinct desired outcomes. Measuring whether people can navigate them would say more than counting how often a lock button gets pressed.

Third, watch how other assistant products respond to selective privacy inside an account. Relevant responses could include conversation access controls or clearer explanations of existing boundaries.

A competing announcement alone would not establish better protection. The comparison should examine what each product changes and what remains governed elsewhere.

If providers explain those distinctions more clearly, this launch will have contributed to a useful design expectation. If conversation locks remain isolated features with confusing implications, the broader lesson will be less encouraging.

For knowledge workers, the larger question is how much material belongs in any personal AI workspace. A useful knowledge management practice starts with understanding the intended audience and future use of information.

That applies whether the material sits in an assistant conversation, a note, or a searchable document collection. Access controls are easier to evaluate when the desired outcome is specific.

Before relying on the feature, decide which problem you want it to solve. Is the concern an unrelated conversation appearing while someone borrows your screen? Is it previously remembered information influencing an answer? Or is it the provider’s handling of submitted content? Those questions deserve separate answers. Review the documented behavior against your situation, especially when a conversation contains work material or information you cannot easily replace. ChatGPT now lets you lock your chats, giving saved discussions another access boundary. Use that boundary deliberately, and check the surrounding settings before assuming the rest of your privacy preferences have changed.