AI Speaks Our Language But Doesn't Think Our Thoughts

So I was on the train last week and could not help but eavesdrop on a family's conversation, which is basically a commuter's constitutional right.

Mum pulls out her phone with a "watch me be tech-savvy" flourish and asks ChatGPT where they should shop for their upcoming vacation. Her teenage kids look suitably impressed, for about 30 seconds. ChatGPT bombs spectacularly, suggesting stores that do not exist or are wildly off-base. Within minutes they have spiraled into an absurd debate about AI, throwing around claims like "gallons of water per prompt", it is actually a few ounces per several dozen prompts, and "basically just hallucinating everything", partially true, these systems do generate incorrect info, but not everything they produce is false.

I am trying not to laugh, thinking about early electricity demonstrations where people paid to watch sparks fly with zero understanding of what was happening. Same energy. No pun intended. Okay fine, pun absolutely intended.

Even Isaac Newton was confused about electricity

Did you know Isaac Newton, literal genius, inventor of calculus, was quite wrong about electricity? He held theories about "ethers" that do not line up with our modern understanding.

In the 18th century, scientists would literally sell tickets to electricity demonstrations. Picture Victorian-era folks paying good money to watch someone make sparks fly or shock himself for entertainment, the 18th-century equivalent of a TikTok stunt account. The scientific establishment nearly lost their minds when someone proved electricity works in a vacuum. "But wait, if there is no ether, then what the heck is it?" Their entire understanding collapsed faster than my motivation when I have to vacuum the living room.

This is exactly where we are with AI. We are fascinated spectators at the electricity show. Even Dario Amodei, Anthropic's CEO, acknowledges we do not really understand how large language models work. We are building things that function in ways we cannot fully explain.

Underestimating intelligence that doesn't look like ours

We humans tend to underestimate intelligences unlike our own while overestimating ourselves. Recent studies have identified about 150-156 distinct codas, the click patterns, in sperm whales' communication "alphabet". English has just 26 letters in its alphabet, though we combine these into thousands of syllables. Most human languages function with between 20-40 basic sounds.

We dismissed whale intelligence for centuries because they do not build cities or write books. We overestimate AI for the opposite reason, because it speaks our language. ChatGPT sounds thoughtful because it uses linguistic patterns we associate with intelligence.

The reality check: AI systems are just extraordinarily good at predicting which word comes next in a sequence. No consciousness or understanding, just sophisticated pattern completion, essentially saying "based on all the text I have seen, when humans write X, they usually follow with Y". When that mum mentioned ChatGPT "hallucinating" store recommendations, she was using terminology that describes a real phenomenon. These systems produce outputs based on statistical patterns with no inherent concern for factual accuracy.

It is like your grandpa's stories that get more impressive with each retelling. There is no fact-checker in his head saying "actually, that fish was 2 kilos, not 20". The difference is your grandpa probably knows he is embellishing, while AI has no self-awareness about its factual errors.

Moving beyond the "ooh, sparks!" phase

For electricity to become useful, people had to move past the "ooh, sparks!" phase. We are at that exact same point with AI. That family on my train never considered that asking "what should we buy for our trip?" without saying where they are going or what they will be doing is like pressing the "I'm Feeling Lucky" button on Google.

The non-obvious principles that actually matter

Think in contexts, not commands. Instead of "best restaurants in Sydney?", try "I'm a food critic writing about Sydney's dining scene focusing on innovative chefs using indigenous ingredients."

Understand token economics. Most LLM companies charge more for output tokens than input tokens, 2-3x more, when you use their APIs, so being verbose in prompts can save money and time if it gets the right answer faster.

Master the balancing act of context. There is real tension between giving enough context and drowning the AI in detail, so know when precision matters and when to leave breathing room.

Use system messages as operating modes. Establish fundamental processing modes like "Evaluate all claims critically" or "When presenting technical concepts, start simple then increase complexity."

Iterative refinement beats perfect prompting. Develop a dialogue pattern, start broad, examine the output, then add constraints, "that's helpful, now focus on X."

Unlock complex concepts with everyday analogies. "Explain quantum computing like I'm a gardener" gives surprisingly insightful explanations that connect to knowledge you already have. The strongest AI users treat these systems as thought partners rather than command terminals.

The current is flowing, we're still learning to see it

When people figured out electricity, it did not become less cool. It became more useful. Understanding it let us build everything from toasters to TikTok.

The same will be true for AI. As we build better mental models, these tools become less mysterious and infinitely more valuable. We are not the first humans to meet a technology we do not understand. We are still in the early, awkward phase, still paying for tickets to the electricity show and posting our amazement on social media.

So next time you reach for AI, write the brief you would give a sharp stranger before they did the task, what you want, the context, and what good looks like. The current is flowing. We are just figuring out how to see it without getting shocked in embarrassing places.