ChatGPT adds a PIN lock for your conversations
Resetting the PIN permanently deletes protected content, while locking a chat leaves model-improvement settings unchanged.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:005:07
Read transcript
ChatGPT is adding a lock for selected conversations—but this is an access barrier, not a new privacy regime. OpenAI began gradually rolling out Locked Chats on October 20, letting users protect a conversation with a four-digit PIN or supported device authentication, including Face ID and Touch ID. The feature is meant for ordinary shared-device moments: show someone one chat without exposing another, or keep a surprise out of view without deleting the conversation. Choose “Lock chat” from a conversation’s three-dot menu, and it moves to a dedicated section, disappearing from regular history, pinned chats and search. It remains available for later, unlike a Temporary Chat. There’s a separate boundary for Memory, too: when Memory is on, new messages in Locked Chats can update memory shared within that section, not regular Memory. But locking a conversation doesn’t erase information already remembered before the lock. The more important distinction is what the feature does not do. OpenAI says Locked Chats are not end-to-end encrypted and are not a separate encrypted vault. Retention settings remain unchanged, as does “Improve the model for everyone.” So locking a chat does not change whether it may be used for model improvement. And resetting the PIN permanently deletes the chats, files and other content protected by the old PIN; it is not a harmless way to set a new code. The release note names Free, Go, Plus and Pro accounts across web, mobile and desktop, with availability rolling out gradually. TechCrunch also reports access for Business, Enterprise and Edu, but that does not mean every account has the feature now. ChatGPT for Healthcare is excluded, as are Temporary, archived, group and shared-link chats, plus conversations with scheduled tasks or meeting recordings. The practical takeaway: this can keep a chat out of casual browsing, but it does not change the data settings behind it. Elsewhere, a different test of AI capability came from the Times championship cryptic crossword. The Guardian reports that Darmok, a program developed by computer scientist Sid Chatterjee, finished the final about three minutes faster than winner Liam Hughes. Its earlier attempt had given up after 18 minutes without solving the puzzle. Chatterjee says the improvement came from stronger models and changes to the solving process, including making Darmok attempt every clue before entering answers. In one clue, “Wolf trap missing its lowest crosspiece?,” crossing letters left a choice between “scarf” and “snarf.” Darmok worked from “snare,” removed its lowest horizontal stroke, and arrived at “snarf”—a solution that tripped up three of the five finalists, including 14-time champion Mark Goodliffe. That is a striking result, not proof that cryptics are uniformly solved: Darmok still struggles more with puzzles by some setters. Chatterjee considers writing a satisfying cryptic even harder, because each clue must read smoothly as ordinary English and work as cryptic instruction. And from solving language puzzles to responding to live performance: the Boston Symphony Orchestra is set to premiere Tod Machover’s “already and not yet” on October 22 at Symphony Hall. A system built at the MIT Media Lab is intended to interact with individual musicians, then help connect their performances as the piece develops. In the second movement, the aim is to bring out each player’s distinct character; in the third, to find common ground across the orchestra. Studio-made electronic layers are part of the work, too, and live sound mixing by Ana Schon remains central to the performance. The system is not simply taking over from musicians or sound engineers. Machover draws a useful line between detecting sound and understanding music. He says current AI can identify audio features and patterns, but is less capable of interpreting performance and expressive intent than earlier systems he developed for Yo-Yo Ma. The orchestra and its technology will be heard live, but the composer’s caveat is that pattern recognition is not the same as musical understanding. Finally, a reported ownership stake puts a striking number on the expectations surrounding Anthropic—but not on any cash Eric Schmidt has received. Bloomberg, as cited by TechCrunch, reports that the former Google chief owns about one percent of the company after investing tens of millions of dollars in its 2021 Series A. Whether he has sold any shares is unclear. At valuations investors have discussed—$1.5 trillion to $2 trillion—that stake would imply a value of $15 billion to $20 billion. Those are hypothetical figures, not proceeds from a sale, and depend on both the eventual valuation and how much Schmidt still owns. For comparison, Anthropic CEO Dario Amodei is reported to hold about two percent; that ownership comparison says nothing by itself about voting power. TechCrunch describes a public listing before year-end as widely expected, not scheduled. The Times of India reports that Anthropic confidentially submitted draft IPO papers in June, but there is no confirmed timetable. Across these stories, the useful question is what a feature or headline actually establishes: a lock doesn’t change data use, a crossword win doesn’t mean puzzle-writing is solved, pattern detection isn’t expressive understanding, and a paper valuation isn’t a realized payout.
