Artificial Intelligence (AI), can feel mysterious, even a little intimidating, if you’re not sure how it actually works. But the story of how we got here is far simpler, and far more human, than most people realise.
Some AI History
Thirty‑five years ago, AI research was focused on defining the structure of language.
The goal was to capture grammar, logic, and meaning using hand‑written rules. But natural language proved far too ambiguous for this approach to scale. Every exception spawned another exception, and the systems became brittle.
Around 30 years ago, when I worked in the steel industry, we were building something different: Expert Systems.
These systems didn’t try to “understand” language or meaning. Instead, they relied on large amounts of plant data and human‑encoded rules. If a certain pattern of sensor readings had historically led to a poor outcome, the system would warn operators and suggest adjustments.
It wasn’t learning in the modern sense, but it was recognising patterns and mapping them to likely consequences. In hindsight, it sat somewhere between deterministic rules and probabilistic reasoning. This was a stepping stone toward what came next.
Today
Today’s large language models (LLMs) take a fundamentally different approach.
Rather than relying on rules or explicit knowledge engineering, they learn by predicting the next word in billions of sentences. Through this simple mechanism, they absorb the statistical patterns of language at a scale no human could ever encode manually.
They don’t “understand” language the way we do; they generate the most probable continuation based on everything they’ve seen.
Humans still play a crucial role.
Specialists review outputs, correct mistakes, and provide preferred or “golden” answers. They define rubrics that describe what a good response looks like. The model doesn’t store these answers; instead, it adjusts its internal probabilities to align with human expectations. This is how modern AI appears to “learn”.
Limitations
Because LLMs are probability machines, they sometimes get things wrong.
If the patterns in the data are flawed, incomplete, or misleading, the model’s output will reflect that. It isn’t reasoning; it’s estimating. When the underlying probability distribution is skewed, the answer will be skewed as well.
And this brings us to the limits of AI – the lottery.
LLMs can analyse historical draws, spot patterns, and even generate number combinations that look statistically interesting. But lotteries are designed to be memoryless. Each draw is completely independent of every draw that came before. The probability resets every time. There is no pattern to learn, no trend to exploit, no hidden structure to uncover.
Any lottery numbers generated by an AI are simply echoes of historical data, not predictions of future outcomes. The next draw is entirely disconnected from all history.
AI is extraordinary, but it is not magic.
It excels where patterns exist.
It fails where randomness rules.
Understanding the boundary is essential as we integrate these tools into real-world decision-making.
If you’ve ever felt unsure about AI or overwhelmed by the hype, I hope this helped clarify things. I’d love to hear your thoughts.
How do you see AI fitting into your world?
In my world, I have used AI to:
- generate photographs and advertising videos for items on thebookhookonline.etsy.com (examples at the end), with extended versions for the YouTube and TikTok channels.
- create sketches from photographs or from text prompts to generate colouring pages.
- correct grammer in my writing.
- advise me on copyright-related publishing issues.
- help with search engine optimisation.
- help with creating some features for this website.
- generate a website that suggests EuroMillions Lotto numbers based on statistical analysis. (YES – I know, the numbers it generates are no more likely to come up than any method, but we all live in hope!)
In all these examples, I have exerted what I would describe as the controlling mind over the AI. This means I describe what I want, validate and accept the result, refine it, or, in some cases, reject it.
Click the link to discover some AI Products.
J K Mullins, 17-March-2026.
