The training toggle is a decoy.

In November 2025 a single warning about Gmail spread fast enough that Google had to answer it by name. The claim was that you had been quietly opted in to let your inbox train Google’s AI, and that turning it off meant hunting through settings in two separate places. Millions of people read it. Enough panic followed that on 21 November the Gmail account posted a flat rebuttal: “We have not changed anyone’s settings,” and “we do not use your Gmail content to train our Gemini AI model.”

The uncomfortable part is that Google was telling the truth. Read the company’s Workspace privacy commitments and the promise is plain and inspectable: your content is “not human reviewed or otherwise used for Generative AI model training outside your domain without permission.” Two weeks of fury, a class-action filing, a hundred how-to-opt-out posts, and a corporate denial, all aimed at a switch that mostly did what its label said.

It was the wrong fight. And it is not the first time we have picked it. Back on 1 July 2023 Google edited its privacy policy to swap “language models” for “AI models,” and the same alarm went up. Every round, the argument is about the same verb: train. Will they train on my data. Nobody in either round asked the question that actually decides whether you are stuck: if you leave, can you take what the assistant learned about you?

The verb everyone argues about is the wrong one

Here is the claim I will stand behind, and you are free to think I have it backwards: an assistant that swears on its life it will never train on your inbox, but will not let you export the context it built about you, is a worse trap than one that trains on everything and lets you walk out with the whole thing. The privacy toggle protects an abstraction. The export protects you.

Training is a one-time extraction. A model reads a pattern in aggregate, learns something diffuse from millions of examples, and your individual emails are not sitting inside the weights waiting to be recited. It is a real concern, worth a real setting. But it is not the thing that holds you hostage. What holds you hostage is the other pile of learning, the one nobody put a toggle on: everything the assistant accumulated for you. Who your important clients are. Which projects died and why. The way you like a reply drafted, and which of your deadlines are real versus the ones that quietly slip. The running context that made the thing finally feel like it knew your work.

That second pile is the asset. It is also, in almost every product, locked in a database you cannot read, structured in a shape no competitor can ingest, and available to you only as whatever the “download your data” button decides to hand over. You can win the training argument completely and still own none of it.

Privacy you can check beats privacy you are promised

A no-training pledge is a promise about what a company will refrain from doing inside a building you cannot see. You cannot audit it. You take it on faith, or on a regulator’s word, and you find out you were wrong only in discovery. Portability is the version you can verify on any ordinary afternoon. Either the context comes out in a format another tool can open, or it does not. There is no faith involved.

This is why we built be/brief around the export rather than the pledge. What your associates learn about your work lives in an open format you can read with our app closed, structured so another tool can pick it up and keep going. Not as a compliance gesture, but because the promise we most want you to trust is the one you can test yourself. We would rather keep you because the work is good than because leaving would cost you two years of accumulated judgment.

The training toggle will keep making headlines, because it is legible and it is scary and it fits in a screenshot. Keep it on, by all means. Just notice what the noise is doing. While everyone stands guard at the front door arguing about what the company reads, the thing worth defending is quietly bolted to the floor behind them. Ask a vendor for the training setting and they will show you one proudly. Ask for everything your assistant knows about you, in a file you can open somewhere else, and watch how fast the conversation changes.