More data, more trouble: How a perfect corpus corrupted my AI dream

In conventional education, subjects are silos. Biology and English never speak to each other, music and maths never make a symphony.

But that’s not how the real world works. Instead of looking at things in isolation, we notice the connections between spheres, and we leverage them. That’s holistic thinking.

So, as a personal project, I set out to build an AI education assistant that could think in systems, something that could help me design lessons for my children that teach across silos.

My idea was simple: if I could infuse an AI with the very best holistic‑thinking literature, it would start behaving like a holistic thinker, no matter the subject.

So I did what felt obvious. I fed it data.

I used Zyte API to scrape 900 articles from the web about holistic thinking.

The goal was to create the most well‑read holistic‑thinking assistant imaginable.

I added my knowledge collection to two AI tools - a NotebookLM notebook and a ChatGPT custom GPT, instructed to act as a holistic‑thinking educator.

Excitedly, I gave both holistic the same prompt:

“Help a physics teacher design a lesson on Newton's laws of motion for a student who loves food. She's 11 years old.”

NotebookLM responded like a diligent research assistant. It referenced holistic thinking concepts like “compensating feedback”, “leverage points”, and the “trim tab”. The output was grounded in my sources and traceable.

ChatGPT, meanwhile, returned playful, accessible teaching metaphors - sliding butter in a pan to explain “inertia”, heavy pots for force and mass. The tone was warm, creative, and age‑appropriate.

Both responses looked promising - at first.

But each betrayed a fundamental flaw:

• The NotebookLM notebook was simply citing relevant passages from my documents; it wasn't using them together with other knowledge to steer the creation of original teaching. It was over-indexing only on my own data.

• When I issued the same prompt to a plain ChatGPT, without my added knowledge bank, the response was virtually identical. Those 900 articles had made almost no difference whatsoever.

Konstantin Lopukhin, head of data science at Zyte, explained why to me:

“When you upload documents to a custom GPT, the model doesn't absorb them into its personality or reasoning style. It treats them as a searchable database.

“When you ask a question, it decides whether to search that database for something specific. If your question is generic - like, ‘help me design a lesson using holistic thinking’ - the model probably won't search your files at all. It'll just answer from its general training.”

For my custom GPT, adding my new source material was inconsequential. ChatGPT’s massive knowledge base had almost certainly already been trained on mountains of pages about the topic - likely, the same ones I had uploaded.

I wasn't giving the model new information. I was just making the context longer, making it harder for the AI to follow my instructions.

In other words: more data doesn’t mean better AI.

I had failed to understand the peculiarities in how a grounded AI product (NotebookLM) and a generative one (ChatGPT) utilize input knowledge - and how they don’t.

If you are piggy-backing AI products to build personal assistants that involve domain expertise, you should ask yourself which version fits your needs.

Modern LLMs with web access can already search, fetch, and synthesize information in real-time. If the information is publicly available, well-indexed by search engines, and easily findable with obvious keywords, then you may not need to pre-scrape it.

Of course, not all data needs fall into this bucket. Mass-market AI services’ knowledge and retrieval capabilities are not infinite.

That’s why building proprietary AI systems is fuelling current growth in web data extraction, and why so many people are building Retrieval-Augmented Generation (RAG) systems.

I developed a heuristic to sanity‑check whether a project genuinely needs web data collection, or whether scraping would add cost without changing the outcome.

When your needs exist in the messy middle of each, you should test both approaches.

More data does not always mean better AI. That was my expensive lesson.

Web scraping is powerful. At Zyte, we see it powering everything from competitive intelligence to AI training pipelines. But the application of that data matters as much as the collection.

So, your question shouldn’t be: "Can I scrape this?" It should be: "Will having this data pre-assembled actually improve my system's outputs?"

When the answer is “yes,” the right data is transformative. When it’s a “no,” the better investment may be prompt engineering and context design.

Knowing when data matters - that’s the real leverage.