Meta Built a Vault. After Years of Living in Your Walls.
Incognito AI chat isn't a privacy innovation — it's a receipt.

There's a version of this announcement that sounds genuinely impressive. Mark Zuckerberg says Meta AI's new incognito chat mode runs inference inside a Trusted Execution Environment — meaning the computation itself happens in a space Meta can't access. Conversations vanish when you close them. Web searches aren't linked to you. Nothing is stored on servers. Not even Meta can read it, Zuckerberg says. He compared it to end-to-end encryption.
Take that at face value and it's a real technical commitment. Trusted Execution Environments aren't marketing language. They're architecture.
But here's the thing about architecture: it tells you what the building was designed for before the renovation.
The Feature That Shouldn't Have Been Optional
Zuckerberg is calling this the first major AI product with no server-side conversation logs. That claim deserves two reads. The first read is a competitive shot at OpenAI, which is currently facing lawsuits over stored chat logs — a detail MacRumors noted in the same breath as the Meta announcement. The second read is a quiet admission that every other major AI product, including Meta's own, has been logging your conversations by default. That's not a revelation exactly. But having the CEO frame your new privacy feature as historically unprecedented does clarify what the history was.
The feature is available in the Meta AI app and on WhatsApp. You turn it on when you want to have what Zuckerberg called a "completely private way to interact with AI." Which implies, without quite saying it, that the non-incognito way was something else entirely.
What the Timing Tells You
Meta didn't invent the concept of private AI conversations because it suddenly cared about your inner life. It built this now, in this climate, while a competitor is being sued over the very data practices Meta is now distancing itself from. That's not cynicism — that's reading a press release with both eyes open.
There's also something worth sitting with in the framing. Zuckerberg said that to get the most from "personal superintelligence" — his phrase, not a modest one — you need to be able to share things you wouldn't want stored. That's a real insight about how people actually use AI assistants. The most useful conversations are often the most sensitive ones. The question of whether you trust the system enough to have them honestly is the whole ballgame.
Meta's answer is: here's a mode for that. Which is fine. It might even be good.
But privacy shouldn't be a mode. It should be the default, and opting into logging should be the feature that requires a tap.
The fact that it works the other way around — that you have to find incognito, activate it, remember it exists — tells you everything about what the product was optimized for before this week. The vault is real. It's just that Meta built it after spending years very comfortably without one.
Keep reading tech.

Helpfulness Was Always the Attack Surface
Mozilla's researchers didn't break Claude Code. They let it do exactly what it was built to do.

Asian AI Built Around the Wall. Now Anthropic Has to Explain the Wall.
When the market you're trying to protect just engineers its own version, export restrictions stop being strategy and start being a countdown.

Tim Cook Blamed Memory. Then He Raised the Price of Everything Else.
Apple's pricing logic is starting to contradict itself out loud.
From the other desks.

Under $25,000, Crank Windows, No Apology
Slate built a truck that costs less than a decent used F-150, and the coverage can't decide if that's genius or a problem.

H. Moser Stopped Explaining Itself. Worn & Wound Noticed.
A writer comes back to the Streamliner Minis with enamel dust still on their hands, and the watches hold up.

Doug Martin's Parents Filed Eight Claims. The Ninth Is Unwritten.
A lawsuit over a former NFL player's death in police custody asks whether institutions built to protect themselves can ever admit what happened.