ChatGPT Work Now Learns Your Writing Style From Gmail, Drive, and Slack

Getting an AI assistant to actually sound like you, rather than sounding like every other ChatGPT output, has been a persistent complaint since generative writing tools went mainstream. OpenAI's answer, shipped September 7, 2026: ChatGPT Work can now learn your writing style directly from the tools you already use — Gmail, Google Drive, Slack, and SharePoint — rather than asking you to describe your voice in a settings field.

TL;DR

What it actually reads and does

Rather than asking users to write out a style guide for themselves — an exercise most people are bad at, because writing style is largely unconscious — ChatGPT Work's personalization feature works backward from evidence: it reads your emails, messages, and files across connected tools and extracts patterns from what you've already written. That includes recurring phrases you reach for, how you typically sign off a message, and small mechanical habits like capitalization choices.

Once learned, that style profile carries into new writing ChatGPT Work generates for you — drafts, replies, and documents — across both web and mobile. The pitch is straightforward: instead of getting generic "AI voice" output that you then have to manually edit into something that sounds like you, the draft starts closer to your voice from the first pass.

OpenAI's announcement thread specified four connectors — Gmail, Google Drive, Slack, and SharePoint — and emphasized that the feature learns from workplace content, not from what you type directly into ChatGPT. That distinction matters: if you already draft most email with ChatGPT, the style learner may be studying its own prior output rather than your original voice.

Writing Style vs Custom Instructions vs Memory

ChatGPT already had personalization layers before September 7. The new feature adds a fourth axis — inferred style from connected apps — and the overlap confuses people:

Writing Style and Custom Instructions are complementary, not redundant. Custom Instructions tell the model what kind of writing you want; Writing Style tries to learn how you actually write from evidence. Memory stores facts; Writing Style stores mannerisms.

Anthropic's parallel is Claude unified memory across Chat and Cowork — facts and context crossing product surfaces — but Claude does not yet offer an equivalent "scan my Gmail and infer my sign-off" feature as of this writing. The competitive gap is automatic voice inference versus manual CLAUDE.md-style rules (what is CLAUDE.md).

Setup

1. Web: Go directly to chatgpt.com/?surface=work&writing_style=setup.
2. In-app: Navigate to Settings → Personalization → Writing style.
3. Connect your tools: Gmail, Google Drive, Slack, and SharePoint are the four sources named in the announcement — connect whichever ones hold a representative sample of your actual writing.
4. Availability: Web, for all paid plans, as of the September 7, 2026 rollout.

Prerequisites and permissions

Before connecting sources, verify three gates:

• Plan access: Writing Style setup requires a paid ChatGPT plan with ChatGPT Work access. Free-tier users will not see the Personalization → Writing style path.
• Connector permissions: Each app uses OAuth with read-only scopes. Organizational Google Workspace or Microsoft 365 admins can block third-party app access — if your IT department restricts ChatGPT connectors, Writing Style cannot read corporate SharePoint or Gmail regardless of your personal settings.
• Representative sample: The feature learns from volume. Connect the apps where your actual writing lives — not the apps where you only read others' messages.

OpenAI separates app permissions from broader personalization controls like Memory. Disconnecting Gmail does not delete Memory entries about your writing preferences, and vice versa.

What the setup flow actually does

Based on OpenAI's announcement and third-party walkthroughs, the pipeline appears to work like this:

1. OAuth connection — You authorize read access to sent emails, documents, and messages in each connected service.
2. Pattern extraction — ChatGPT Work analyzes recurring linguistic habits: vocabulary frequency, greeting and sign-off templates, sentence length distribution, capitalization quirks.
3. Style profile creation — The extracted patterns become a persistent profile applied to new Work-generated text on web and mobile.
4. Ongoing sync — New content in connected apps may update the profile over time (OpenAI has not published refresh cadence or retention policies).

OpenAI has not publicly documented how profiles are represented internally, whether formal and informal writing are weighted differently, or whether users can inspect or edit individual learned patterns — only enable, disable, or disconnect sources.

The honest limitation: which writing gets learned matters

A sharp reply on OpenAI's announcement thread put the real risk plainly: "I have never typed an em dash in my life. If it learns my style and the first draft still has three of them, it learned someone else's." That's a fair test, and it points at a genuine design question OpenAI hasn't fully answered publicly — does the feature weight formal writing (a carefully drafted email) differently from casual writing (a fast Slack message), or does it treat all connected content as equally representative of "your style"? If it's the latter, a user who writes carefully in email but dashes off quick, sloppy Slack messages could end up with a learned voice that skews toward the sloppier sample simply because there's more volume of it.

Another reply raised the more cynical version of the same concern: "ChatGPT training itself on the emails it wrote in the first place" — a fair point for any user who already drafts most of their email with ChatGPT, since a style-learning feature reading that output risks reinforcing ChatGPT's own generic patterns rather than the user's original voice.

Context mixing — formal email vs casual Slack

The design question OpenAI has not fully answered: does Writing Style weight sources differently?

Consider a typical knowledge worker's connected corpus:

If Slack volume outweighs carefully drafted Gmail, a learned profile might produce client emails that open with "hey team" and close with a reaction emoji. The fix — if OpenAI ships it — would be per-source weighting or per-context profiles ("use Slack style for internal drafts, Gmail style for external").

What OpenAI has not clarified

As of September 11, 2026, OpenAI's public documentation leaves several builder-relevant gaps:

• Retention: How long are writing samples stored? Are they used for model training or only for the user's style profile?
• Inspection: Can users view, edit, or delete specific learned patterns (e.g., remove an unwanted sign-off)?
• Multi-user workspaces: If multiple people share a Slack workspace, whose style does the feature learn?
• Regular Chat mode: The announcement scoped Writing Style to ChatGPT Work only. A user reply on OpenAI's thread asked why it was not available in standard Chat mode — OpenAI had not addressed that gap publicly at time of writing.
• Enterprise admin controls: Whether org admins can disable Writing Style org-wide (similar to connector policies) is undocumented.

Users can disconnect apps or delete their learned writing profile through settings, per third-party reporting — but the exact deletion scope (profile only vs. cached samples) is unclear.

How to test whether your style profile actually works

Before relying on Writing Style for client-facing drafts, run a quick validation:

1. The em-dash test — Ask ChatGPT Work to draft a short email on a neutral topic. If you never use em dashes and the draft includes three, the profile learned generic AI habits or someone else's patterns in a shared workspace.
2. The sign-off test — Generate a formal reply. Compare the sign-off against your last 10 sent Gmail messages. Mismatch suggests Slack or template content dominated the profile.
3. The vocabulary test — Use a phrase you never write (e.g., "I hope this email finds you well" if you never open that way). If it appears in output, the learner is applying population defaults, not your corpus.
4. The A/B against manual rules — Draft the same prompt with Writing Style on vs. off but with a mannered-prose-style custom instruction. Compare which output requires fewer edits.

Document what passes. If the profile fails the em-dash test, disconnect Slack temporarily and reconnect only Gmail to see if register improves.

Manual alternatives when auto-learning skews wrong

Writing Style is not the only path to non-generic AI output. If connected-app inference produces the wrong voice, these approaches give you explicit control:

Custom Instructions with concrete examples. Instead of "write like me," paste three paragraphs of your actual writing and instruct: "Match the sentence length, vocabulary level, and sign-off pattern of these examples. Do not add em dashes, rhetorical questions, or metaphorical openings."

The mannered-prose prompt. explainx.ai's mannered-prose guide targets the specific failure mode where models substitute flourish for direct statement — a problem Writing Style does not fix if your corpus already contains AI-drafted content.

Load-bearing tells cleanup. Claude Opus 5's Claudisms catalog the mechanical tells (certain transition phrases, hedging patterns) that make AI text detectable. A style profile that learned from AI-heavy email will reproduce those tells faithfully.

Detection awareness. If your goal is writing that passes human review, cross-check output against explainx.ai's top 10 signs of AI-generated text — learned style that reproduces your Slack casualness may still trigger detectors on formal documents.

Enterprise and team considerations

Writing Style is a work decision, not just a personal toggle:

• Data access scope: Connecting SharePoint grants ChatGPT read access to whatever documents your account can reach — potentially including HR policies, financial drafts, and legal correspondence. That is a broader surface than pasting one document into chat.
• Style ownership on shared accounts: Team ChatGPT Work accounts with shared connectors may produce a blended "team voice" rather than any individual's style.
• Compliance: Regulated industries (finance, healthcare, legal) may have policies against third-party AI reading email archives. Check internal AI usage policies before connecting Gmail or SharePoint.
• Revocation: Users can disconnect apps through settings, but org-level connector policies may override individual choices.

For a full map of which OpenAI product surface handles which task — and where Writing Style fits — see ChatGPT Work vs Codex.

Where this fits next to explainx.ai's own coverage

This sits in the same category explainx.ai covered with Claude Opus 5's "load-bearing Claudisms" and the viral "mannered prose" prompt — both are about closing the gap between generic AI writing and writing that actually sounds like a specific person, just from opposite directions. The mannered-prose prompt is a manual rule you apply explicitly; ChatGPT Work's personalization tries to infer the equivalent automatically from your existing writing. For anyone frustrated by generic AI output, both are worth trying — and comparing which approach actually gets closer to your real voice is a genuinely useful exercise.

It's also worth reading alongside ChatGPT Work vs. Codex for anyone deciding which OpenAI product surface to use for which task — this feature is specifically a ChatGPT Work capability, not something that currently reaches Codex or standard ChatGPT chat.

The broader question — whether AI-assisted writing remains a viable career skill — connects to explainx.ai's coverage of whether AI writing is a safe job. Writing Style automates the voice layer; it does not automate judgment about what to say or to whom.

Builder takeaways

Related on explainx.ai

• The "mannered prose" prompt that fixes AI writing on any model
• Claude Opus 5's load-bearing "Claudisms" and writing tells
• ChatGPT Work vs. Codex: what actually changes
• Top 10 signs of AI-generated text
• Is AI writing a safe job? Mollick, Demirbas, and the wicked-problem debate
• Claude unified memory across Chat and Cowork
• What is CLAUDE.md — persistent memory for Claude Code

Sources

• ChatGPT on X — writing style personalization announcement, September 7, 2026
• RuntimeWire — feature walkthrough and connector details
• OpenAI ChatGPT Work product page — connected services

Feature availability and setup steps reflect OpenAI's September 7, 2026 announcement. Check ChatGPT Work's live settings for current availability, since paid-plan feature rollouts sometimes expand or change scope after initial launch.