Atlassian’s guidelines for writing with AI

How should we use AI to write?

AI is a helpful tool but, without guardrails, its writing can fall flat. We joke about AI tells like em-dash overuse and “it’s not X, it’s Y,” but issues with over-reliance on AI writing go deeper.

• Externally: We’re all already inundated with increasingly AI-generated text. And if you’ve ever received an email clearly written by AI, you’ve likely felt how it erodes trust in the sender.
• Internally: AI-generated content feeds our knowledge graph. Think of it like a constant xerox copy of a photograph—with each copy of a copy, the quality degrades. When the content is generic and unvetted, it solves for no one.

On Atlassian’s Brand team, these areas are top of mind. We want to move fast with AI but still put out content that is useful, fun to read, and trustworthy. Our mission has always been to unleash the potential of teams, and we believe a key part of that is putting out content that represents our best, most actionable, and most unique thinking.

In April, we brought together a group of writers across Atlassian to discuss if, where, and how AI should show up in the writing process. We then pressure-tested those guidelines with our entire Marketing function.

In the spirit of “Open company, no bullshit,” we decided to publish them here as well.

This is an ongoing journey. Rather than a rigid set of rules, these guidelines outline what we aspire towards. We hope they’re useful for your team, too.

Core principles

1. Humans lead all writing and ideation. AI can help validate, edit, and refine, but a human should always write the first and final drafts in their own voice.
2. Everything we publish needs to have a human byline. That human is publicly accountable for the content, including anything AI-influenced.
3. Give AI the proper context. When using AI to edit or refine, always specify the text’s audience, purpose, and tone to avoid generic feedback.
4. Fact check, fact check, fact check. Anything AI says or suggests must be verified by the accountable human before implementing—including any factual information, data, or quotes.
5. Don’t publish anything you wouldn’t read yourself. If you don’t find the piece genuinely helpful or resonant, our audience probably won’t either.
6. Guard against AI erasure. AI defaults to the audiences most represented in its training data, over-representing some perspectives and potentially flattening others. When reviewing AI writing suggestions, pay as much attention to what it removes as what it adds.

Applying these principles at Atlassian