Everyone is learning that success with AI isn’t quite automatic. A foundation of credible knowledge is critical for getting the most out of AI tools, and a recent experiment delving into millenia of written knowledge explains why.
It all started when Paul Gustafson, president of Expert Support, casually asked his smart speaker about the fate of the Library of Alexandria and ended up stumbling onto something that reframes how we should think about artificial intelligence. Within an hour, using nothing more than a voice assistant, a follow-up email, and Google’s Gemini, he and colleague Paul Carlstrom had a full first draft of a blog post about ten historical figures who built the architecture of human knowledge.
Well, so what? Here’s why the experiment mattered: It exposes why many enterprise AI projects aren’t meeting expectations. The missing ingredient isn’t better prompts, it’s the volume of organized, credible, peer-reviewed human knowledge that AI depends upon for accuracy. Otherwise, generated information fails to be helpful for agents and humans alike.
Understand How a Simple Question Started It All
Gustafson didn’t set out to write an article. Rather, he was just talking with Alexa, curious to know why the modern AI industry is wrestling with a problem humans have tried to solve for millennia. How do you organize what you know so other people can find it, trust it, and build upon it?
He asked a simple question about the history of libraries, wondering if the Library of Alexandria was the first. The answer was a big surprise. The Library of Alexandria, the institution he — like most people — assumed was civilization’s first great knowledge repository, wasn’t the first at all. In 679 BCE, roughly 800 years earlier, an Assyrian king named Ashurbanipal had already collected tablets and organized a vast library. Even more importantly, Ashurbanipal invented the colophon, the ancient equivalent of a book’s copyright page, listing a scribe, a date, and a source.
Follow the Thread AI Hands You, One Question at a Time
Gustafson then realized there must be a number of people who had advanced the state of human knowledge organization. He asked Alexa to send him an email with a transcript of the conversation, and got it. He then handed those emails to Gemini with a simple instruction: Draft a blog post about the ten most important historical figures who advanced this work.
Twenty minutes after his first conversation started, he had a working draft, complete with citations and links back to source material. That speed is the part people have become accustomed to, and delivers the “wow” factor most of us first experienced within weeks of ChatGPT’s debut. But the real insight here was more important.
Trace the Draft Back to the Invisible Library Beneath It
It dawned on Gustafson that the blog post draft only existed because online resources like Wikipedia exist. Of these, Wikipedia may be the best known. It’s a resource thousands of volunteers build and maintain by spending thousands and thousands of cumulative years documenting knowledge infrastructure pioneers like Ashurbanipal, Zenodotus, Cai Lun, and Ibn al-Nadim in a form any machine, or human, could parse.
What’s more, Wikipedia itself rests on a massive foundation of human knowledge and expertise. It’s a foundation created by decades of archaeologists sifting through ruins, historians translating cuneiform tablets, and researchers publishing findings that other researchers then checked, argued over, and confirmed through peer review. None of that labor shows up automatically in a modern LLM’s chat window. All of it shows up in the answer. Strip away that foundation, and an AI model has nothing credible to draw from — it can still generate smooth confident-sounding sentences that appear right, but it can no longer generate credible information people can trust and depend upon.
Recognize Why Many AI Projects Struggle with ROI
There’s a pattern research keeps confirming: A widely cited MIT study and others have found that the vast majority of generative AI pilots inside companies fail to produce measurable returns. It’s not that the models are weak. It’s that organizations often overlook the things that make items and activities like Gustafson’s rapid draft possible.
But without critical foundations, an organization (under mandates to “do something with AI”) can be pressured to deploy a chatbot on top of a knowledge base nobody has validated, curated, or organized in years. And then people wonder why the answers provided aren’t great. The lesson from Ashurbanipal’s scribes and Wikipedia’s volunteer editors is the same lesson enterprises tend to ignore: Intelligence, artificial or otherwise, needs a well-tended, well-organized library to draw upon, not a data swamp.
Ask Who Your Own Custodians of Learning Are
The Library of Alexandria employed a staff called the Custodians of Learning. It’s a great job title, isn’t it? Their job, as you can imagine, was to curate, catalog, and maintain the collection over time. But as time went on, their roles became less valued by those in power. Hello, Dark Ages.
Today, historians generally agree that what ultimately destroyed the library wasn’t a single dramatic fire Julius Caesar ordered. While there may have been a fire, the library didn’t suddenly vanish. The real reason, sadly, was neglect, the slow decay that creeps in when an institution loses interest, turns away, and stops paying for its own upkeep.
What a tragedy. Imagine if the Library of Alexandria had survived into our modern era. How much further along would humanity be? Sadly, we will never know.
That tragic history presents some uncomfortable questions for every enterprise executive chasing AI ROI: Who are the “Custodians of Learning” for your enterprise? What enterprise knowledge makes your organization better and unique?
Who captures and organizes what your best experts know? How does their expertise make you better than your competitors? Who validates your content, retires what’s outdated, and organizes the rest so a person or a model can actually find the relevant know-how the moment they need it? If you can’t answer those questions quickly, your AI investment may be standing on the same kind of wobbly foundation that, over time, let Alexandria crumble.
Understand Where Human Ownership Takes Over
Gustafson and Carlstrom made a deliberate choice about how to tell this story: Start with the machine and then use human judgment for revision. Here’s the initial prompt:
Please identify other significant ancients who helped humanity collect, store, organize, preserve, and use human knowledge. Once the top 10 historical figures have been identified, write a 1000-word blog post in the voice of Paul Gustafson (as defined by his posts at Expert Support) that provides a summary of each, and a relative timeline that begins with Ashurbanipal and ends with Gutenberg.
None of this is groundbreaking. You probably use AI tools in the same way. As it turns out, the Gemini draft was solid — informative, well-sourced, genuinely useful as a starting point.
Carlstrom’s rewrite is shorter, better ordered, and carries a voice and a sense of humor unique to Expert Support. The gap between those two versions is the whole argument in miniature. Generation is not composition. A model can hand you sentences, but only you can decide which ones are true, which ones matter, and which ones sound like your organization rather than like everyone else generating blog posts with AI tools.
Remember That Editing Is Thinking, Not Overhead
As Carlstrom likes to say, writing and editing are forms of thinking, and when you let a machine handle every step of that process, you stop thinking. You can and should use AI aggressively — Gustafson and Carlstrom certainly did — but the moment you hand over full ownership of the output, you also hand over your own agency to fix what’s wrong, clarify what’s vague, and decide what actually deserves to be published. That distinction between using a tool and outsourcing your judgment to it is exactly where AI adoption succeeds or quietly fails inside most companies.
Build the Knowledge Foundation Your AI Actually Needs
None of this is an argument against using AI. Gustafson’s own experience proves just the opposite. But if you want your enterprise AI applications to deliver the “wow factor” he got from a 20-minute exchange with Alexa and Gemini, you’ll need a solid foundation of captured knowledge and expertise, validated content, and a design and process for organizing all of it into something findable and trustworthy.
AI can absolutely help you build that infrastructure, but not without your help. But it cannot replace the decision to build it, and it cannot substitute for the human judgment that decides what’s right, important, and worth keeping. The organizations that prioritize this enterprise knowledge foundation as a prerequisite, rather than an afterthought, will be the first ones to see the returns everyone else will still be chasing.
We’ll close with a quote that Gemini helped us find and cite properly that reminds us of the fate of the Library of Alexandria, and the lessons it teaches us all:
“Those who cannot remember the past are condemned to repeat it.”
— George Santayana (The Life of Reason, Vol. 1, 1905)
Internet Encyclopedia of Philosophy
If you need help sorting out your Alexandria of enterprise knowledge, please let us know. We’d love to explore how we can help your enterprise avoid repeating the mistakes of our ancestors.