AI and content: Avoiding disaster

As a purveyor of high-stakes technical content, I am watching the rise of AI with alarm. Our interest in automation and new technologies is on a collision course with our mandate to deliver timely, accurate information. I am not the only one who is concerned; many people are writing on this topic. (Here’s a recent post from Michael Iantosca.)

genAI and authoring efficiency

Generative AI (genAI) can and should serve as a supporting tool for authors. Two obvious use cases are refactoring/converting content and cleaning up grammar and mechanics. Using genAI to create new information is more challenging—you need a good starting point and too often, we do not have one. Automatically generating a datasheet from product specifications is easy, as long as we have accurate source data in a consistent format. My general experience is that we have neither accurate specs nor any consistency in the content that is captured. So creating a datasheet means hunting down the spec, then validating those specs against reality, and only THEN creating a datasheet or similar document. Can we do better? Of course. Will organizations start creating accurate specifications? I am not holding my breath.

GenAI will work best when the underlying data is accurate, well-organized, and uses consistent patterns.

Today, we have underlying data that is sloppy, out-of-date, and incomplete. 

If we want to use genAI for authoring, we have to address our existing content debt.

AI enablement

Content consumers are increasingly using AI to access information, which results in a shift for content creators. AI is a new delivery end point. Instead of producing content for direct consumption (like a PDF file or a collection of webpages), we must create the information that AI uses to provide answers.

The problem with this scenario is that the AI interface now sits between a piece of content and the consumption of that content. In other words, my carefully crafted document is irrelevant because the reader will never see it. Instead, AI consumes the page and delivers content to the human downstream using the AI’s preferred format (like an “AI overview” snippet in Google search or a ChatGPT response).

Nonetheless, I see huge opportunities. Remember that AI is math. A large learning model (LLM) is a mathematical model of text relationships. Therefore, it is your job to produce information that makes the math work.

The first step is to understand your AI customer’s requirements. How should you organize and present the information so that the AI can process it? A few guidelines have emerged:

• Organize information consistently. Always put the same information in the same location. So, for example, a task should have a title (using “How to XXX” or maybe “XXXing the YYY” but never both), an introductory paragraph, and a series of numbered steps.
• Use consistent terminology. Be careful in your use of technical terms, and use them consistently. For example, “car seat,” “safety seat,” and “baby bucket” are all fine in casual discussions, but in a technical document, you should pick one and use it exclusively. 
• Keep sentences short and simple. Use straightforward language. Don’t use florid, complex sentence structures and definitely don’t use weird similes, metaphors, or hyperbole. Let your words flow into the AI whirlpool like a gently scented spring breeze, splashing down to invoke meaningful experiences. (That. Don’t do that.)
• Provide context. Use metadata to further categorize information.

And here is where you see the nexus between AI and structured content. The guidelines that result in more effective content for AI mirror the results of implementing structured content and general best practices for writers.

Roadmap to successful AI

Here’s what I think you should do:

1. Address your content debt problems. Make sure your sources are accurate, up-to-date, and consistently formatted.
2. Move your content into a structured content workflow.
3. Extend the semantics and the metadata of your content for added precision.

Once you are done with these three simple steps (!!), only then should you begin to think about more sophisticated possibilities. These may include:

• More advanced connectors, such as APIs and use of MCP
• Adding retrieval augmented generation (RAG) to your AI processing stack
• Moving content from XML into a knowledge graph
• Further optimizing content to ensure it performs in your AI stack

Questions about AI and content? Let’s talk!