ChatGPT Work's New Writing Style Feature: What Connecting Your Inbox Actually Costs

On September 7, 2026, OpenAI began rolling out a writing style feature for ChatGPT Work: connect Gmail, Google Drive, Slack, and SharePoint, and ChatGPT studies how you actually write across them, then carries that voice into new drafts. No more pasting three old emails into a chat and typing “sound like this.”

That part is a genuinely good idea. People have wanted this for as long as AI drafting tools have existed, because a model’s default voice is nobody’s voice in particular. The part worth slowing down on is what “connect Gmail, Google Drive, Slack, and SharePoint” means in practice, since it is a bigger ask than “read a few writing samples.”

What the Feature Actually Does

The setup lives in ChatGPT settings, under Personalization and then Writing style. You connect one or more of the four supported apps, and OpenAI says the model infers phrases you favor, capitalization habits, formatting choices, formality level, and how you typically open and close a message. Once configured, that inferred style applies to new drafts on web and mobile without you restating your tone each time. BleepingComputer’s report on the rollout has the fuller walkthrough of the setup screens.

Coverage of the rollout has been consistent on the mechanics and less consistent on availability: some outlets describe it as live for ChatGPT Work accounts on the web, others describe it as still in limited testing. Treat “available to everyone right now” as unconfirmed and check your own account’s Personalization settings rather than assuming either way.

The Trade Is Bigger Than a Few Writing Samples

Here is the part the feature name undersells. To infer your writing style from Gmail, the model does not need three representative emails, it needs read access to the account, because it has no other way to find the representative ones. The same is true of Slack, Drive, and SharePoint. “Learn my writing style” and “grant an AI assistant standing read access to my inbox, my chat history, and my company’s shared drive” turn out to be the same permission request wearing two different names.

That distinction matters because those are two different risk profiles. A handful of writing samples is a snapshot you chose. An active connector is ongoing access that persists until you revoke it, covering everything in that account, not just the parts that happen to be good style examples. If your Gmail contains a client’s financial details or your Slack contains a colleague’s performance conversation, that content is now reachable by whatever wrote the style profile, whether or not a single word of it ends up echoed in your next draft. Is it safe to paste work emails into ChatGPT covers the same underlying question for the simpler case of pasting text by hand, and the answer gets more complicated, not less, once the tool is reading the account continuously instead of the one message you pasted.

Why the Access Question Will Not Go Away

This is not a knock on OpenAI specifically. Any assistant that wants to sound like you from real examples faces the same design fork: read broadly and infer, or ask you to hand over a small, deliberate set of samples yourself. The first option is less work for the user and a larger data footprint. The second is more setup and a much smaller one. What employees upload to AI tools is the broader version of this same tradeoff, and it is worth reading before you connect anything, not after.

None of this means the feature is unsafe to use. It means the actual decision is not “do I want AI that sounds like me,” which almost everyone does, but “do I want to get there by giving an assistant standing access to my inbox and chat history, or by choosing a handful of examples myself.” Those are different questions, and the feature’s name only asks the first one.

A Narrower Version of the Same Idea

The alternative is to keep the examples, drop the connector. You pick a few messages you actually sent, in the register you want to sound like, and feed those directly. Nothing else in the account gets touched, and the sample set only grows when you decide to add to it.

Before:

Hi team, wanted to give a quick heads up that the vendor migration is going to slip a bit. We’re working on it and will keep you posted, sorry for any inconvenience!

After:

The vendor migration will slip past this week’s date. QA sign-off is the blocker; I will send a revised date by Thursday once that clears.

The second version is not just shorter. It drops the reflexive apology and the vague “we’re working on it,” and states the actual blocker and the actual next date, which is what a colleague reading a status update wants first. Getting a model to reliably produce that version, in your phrasing, is exactly what a style feature is supposed to buy you. It does not require reading your Slack history to do it; it requires a couple of real examples of you writing status updates that way.

A Wrivio Context for this could say:

Rewrite this update in my style based on the examples below. Match my typical sentence length, greeting, and sign-off. Lead with the outcome, not an apology. Keep every name, date, figure, and commitment exactly as written. Do not invent phrasing I have not actually used in the examples.

Press Ctrl+Shift+Space, paste the update, and check the diff. The examples that teach a Wrivio Context your voice are ones you add on purpose, in Local mode if you want the strongest guarantee, so the writing that trains your style profile never leaves the machine it was typed on. How to make AI-written emails sound like you covers how to pick examples that actually carry your voice instead of just your topic.

Common Questions

Does ChatGPT Work’s writing style feature read my entire Gmail account?

Public reporting on the September 2026 rollout says the feature analyzes content across whichever connected apps you enable, rather than a fixed sample you select in advance, which functionally means it can draw on anything in that account rather than a chosen subset.

Can I get the same style-matching result without connecting Gmail or Slack?

Yes. Pasting a handful of your own real messages directly into a prompt, or into a stored instruction like a Wrivio Context, gives a model the same kind of material to imitate without granting ongoing access to an entire mailbox or chat history.

Does connecting these apps mean my messages train future models?

Reporting on this specific feature does not say either way. If that question matters to you, check your account’s current data-use and training settings before connecting anything, rather than assuming the answer from a news summary.

Is this the same thing as Wrivio’s Styleprint?

The goal is similar, sounding like you instead of a generic AI voice, but the mechanism is different. Styleprint works from examples you choose and add yourself, and in Local mode those examples and the rewrites built from them never leave your device, which is not the same privacy shape as an always-on connector into cloud accounts.

Download Wrivio for Windows to build a writing style from examples you choose yourself, without connecting your inbox to get there.