My biggest issue with AI isn’t that Ii’s “too smart”... it’s that we’re pretending it understands things

The more I use large language models, the less I think the real problem is “AI taking over.” My bigger concern is how easily people confuse fluent output with actual understanding. These systems are incredibly good at generating convincing language. That part is real. But convincing language is not the same thing as comprehension. An LLM can tell you a Boeing 747 has four engines because it has statistically associated those words together from huge amounts of training data. That does not mean it understands what a plane is, what flight is, or what an engine actually does in the physical world. That distinction matters more than people admit. A lot of users interact with AI as if it were a knowledgeable entity reasoning through problems the way a human expert would. In reality, the model is predicting likely sequences of tokens based on patterns it has seen before. Sometimes that produces excellent answers. Sometimes it produces complete nonsense delivered with total confidence. And that’s the part I think people underestimate. There’s also the cost side of this that rarely gets discussed outside tech circles. These models are unbelievably expensive to run. Every interaction requires massive compute infrastructure, energy consumption, cooling systems, and data center capacity. The scale is hard to even visualize. We’ve normalized asking AI casual questions the same way we’d ask a search engine, but the backend cost is dramatically different. What worries me is that we may be building an ecosystem where “information-shaped text” becomes more valuable than verified information itself. A polished answer feels authoritative even when it’s partially wrong. And because the delivery is conversational, people often lower their guard compared to reading a normal webpage or forum post. To be clear, I’m not saying LLMs are useless. They’re obviously useful. I use them myself. But I think the public conversation around AI has become strangely polarized between “this is AGI” and “this is worthless,” when the reality is more complicated. These systems are powerful pattern-generation tools. That’s already transformative. But treating statistical language prediction as human-like understanding creates expectations that the technology simply does not meet right now. And considering the environmental and infrastructure costs involved, that distinction probably matters more than ever.