AI and writing: What is missing and what could we lose?

For about a year now, I’ve been researching a biographical subject for a book I’m writing. I’ve spent hours poring over yellowed letters in archives, requested dozens of old books and journals from the bowels of libraries, accessed databases and gone beyond paywalls to search for buried records, and asked numerous people for help. I’ve also found information on the internet, of course, but it pales in comparison to the nuggets I’ve discovered through these deeper efforts.

This has all got me thinking about how much information is simply not available on the surface web—the part of the internet we can access using standard search engines. Taking this further, I’ve been thinking about how much is missing from the subset of the internet that AI bots access when they respond to prompts.

As an experiment, I recently asked ChatGPT to tell me everything it could about my biography subject. The response came quickly and looked promising on the surface, but it didn’t take me long to spot the problems. It had enormous information gaps that my human research has filled. It glazed over or left out some of the most interesting parts of my subject’s life, and it lacked the telling details that will bring the story alive. It was also missing the connections and nuances that my more laborious research process has revealed. The bot kept offering to dig deeper and I kept telling it to do so, but that only generated more superficial responses.

This experiment didn’t surprise me, and it left me feeling reassured that I, as a human writer, can do so much that it can’t. However, it also made me more worried about what’s at risk as AI continues to permeate the writing process.

With intentional irony given the subject of this blog post, I started a new ChatGPT session and asked what information is missed if a writer relies on AI to research and generate ideas. “A lot,” its response began. Further prompts and my own experience highlight five main categories of missing information:

• Unpublished, like the archival documents and privately held records I’m finding for my biography project
• Unrecorded, like the lived experiences of the people I’m interviewing
• Restricted or fragmented, like the information I’m finding in databases or buried in PDFs or scanned documents
• Underrepresented, like the sources that critique some of my subject’s work in colonial settings
• Real-time or changing, like ongoing research or evolving interpretations of my subject’s life and times

Had I thought to use AI at the beginning of my project (which I didn’t), it would have given me a decent starting point for my research, but that’s all. I could never skip the human work, conversations, creative thinking, and time it has taken to gather information outside the AI-sphere, nor would I want to. They are what will make my book meaningful and one that only I can write.

When we’re writing, editing, or reading something that draws on AI, we must remember that this tool is using a slice of the internet that is itself only a small slice of all human knowledge. We must ask what it’s missing, and we must ask what we could lose if we rely on it too much.

Again embracing the irony, I asked ChatGPT what we should do to protect human knowledge as AI use continues to grow. “One of the greatest risks is not that AI will destroy knowledge,” it responded. “It is that people may stop creating, preserving, and consulting original sources because AI-generated summaries feel sufficient. Human knowledge has always been protected by redundancy: many books, many archives, many witnesses, many researchers, many perspectives. The safest future is one in which AI becomes another layer in that system—not the only layer.”