ChatGPT is moving towards a more personalised writing experience as OpenAI increases how much it can know about you (the kind of person) by connecting workplace applications. The idea is simple: instead of asking ChatGPT “make this sound like me” every time, the AI can already see examples from emails, messages and work documents you’ve written before and use them to write drafts closer to how you normally write things.
New reports say the new feature can analyze vocabulary, sentence structure and emotional tone when users log in to workplace services. Gmail, Google Drive, Slack and SharePoint are some of the services mentioned in connection with personalization experience.
For many users, this would make AI-generated communication feel much less generic. One of the biggest complaints about AI-written emails and messages is that they sound like “AI voice” — over-polished language, repetitive phrases, or a tone that doesn’t sound natural to the person sending it.
The new approach tries to fix that by giving ChatGPT more context about how an individual actually communicates.
For example, someone who normally writes short, direct emails might get short drafts instead of long formal responses. Someone else uses a warmer, more conversational style, so messages could be written in that tone. The system could identify wording patterns, sentence structures, etc from the content it is allowed to see.
This is especially important in professional environments where people write similar messages all day long. Employees might have to respond to clients, colleagues, managers, and business partners in a consistent personal or professional voice.
OpenAI already supports connections to services (Outlook in this example) so ChatGPT can search for and reference emails and draft replies from messages in a connected mailbox (permissions and workspace settings apply).
Google-connected services have similar controls too. OpenAI docs state that when users connect Gmail, Calendar or Drive “ChatGPT can access information required for supported requests from the user’s Google permissions and workspace settings (where applicable)”. Eligible info accessed through connected Google apps may also help personalize ChatGPT if Memory is enabled.
But that also raises an obvious question: how much personal communication should an AI be allowed to analyse?
Emails and private messages can have sensitive information (confidential business discussions, personal relationships) and details about users’ lives that users don’t expect an AI system to do more than one task with. Connecting an inbox therefore means paying attention to permissions and data controls.
OpenAI says users have control over whether their conversations with ChatGPT can be used to improve models: Users of individual ChatGPT services can turn off “Improve the model for everyone” in Data Controls. Temporary Chats are also not used in training models (they don’t create memories or use existing ones).
There is also another important distinction between personalization and model training: an AI using information from a connected account to answer a user’s request isn’t the same as that information being used to train the underlying model. Rules for data use differ by product, type of account, connected application, and settings.
For businesses, OpenAI says inputs and outputs of ChatGPT Business, Enterprise, and API aren’t used for training by default unless the organization opts in.
The bigger picture is clear: AI assistants are moving from generic tools to systems that know what people like and how they work.
Instead of spitting out an email from a blank prompt, ChatGPT can use more context (what platforms you communicate on), what you work on (projects), and what information is available through connected tools.
That could make AI-generated writing faster, more natural and more useful. But it also makes privacy and permission settings even more important.
The goal isn’t necessarily for ChatGPT to replace the person’s voice. The technology is moving in that direction — AI writing sounds more like what the person has to actually send it — but users also have more control over what information the system can access and how that information is used.
