Personalization Moved From "You Describe It" to "It Goes and Looks"
Until now, getting a model to write like you meant describing yourself in custom instructions: a bit more formal, no exclamation marks, sign off with "Best." The ceiling on that is obvious — people describe their own writing habits badly, and the description is a static paragraph regardless. What changed is the source of evidence. Instead of a self-description, the model reads emails and documents you actually sent, and pulls out the phrases, sign-offs and capitalization patterns that recur. For the genuinely hard problem in AI writing — not grammar but voice — that is the more targeted route. A grammatically flawless email can still read as obviously not yours, because it doesn't use the handful of words you always use.
Three Things to Settle Before Turning It On
First, retrieval scope. What it reads is your message and document history, and in a corporate environment that is not something an individual should decide by ticking a box in settings. The SharePoint path deserves particular attention: once connected it can search SharePoint and supported OneDrive work or school content, with the boundary set by existing permissions and workspace settings — meaning what it can see is everything your account can already see. Second, retention and training terms. "Learning your style" means this material was read. How long it is retained afterward, whether it feeds training, and whether administrators can audit any of it are questions to confirm in the enterprise agreement beforehand, not after. Third, who decides. The feature is scoped to ChatGPT Work and its real availability depends on administrator settings — which is OpenAI itself treating this as an organizational switch. The right move on the IT side is to set policy first and then enable, rather than letting employees each connect their own accounts.
"Sounds Like You" Has Its Own Cost
One more thing worth flagging. For the past two years, "you can tell this was written by AI" has been a useful signal — recipients used it to gauge how much attention went into a message, and to stay a little more skeptical of the contents. Once the model has your verbal tics and your sign-off, that signal is gone. Which makes review before sending more important, not less. OpenAI's own description keeps that line: this personalizes drafts and does not include automatic approval or sending. Use it as a tool that produces a first draft sounding more like you — not as a ghostwriter you can stop reading.
via: BusinessToday, AlternativeTo, WinCentral