You Look Like a Thing and I Love You: How AI Works and Why It's Making the World a Weirder Place: Summary & Key Insights

When I use the term ‘artificial intelligence,’ I don’t mean a sentient robot from fiction. I mean the collection of systems we humans design to perform tasks—sometimes narrow, sometimes complex—that seem to require something like thought. Many of those systems are not nearly as capable as popular narratives would have you believe. They don’t think, feel, or reason; they detect patterns. And at their heart, modern AI systems are powered by machine learning—algorithms that learn from examples instead of obeying hand-written rules.

Rule-based systems were once the gold standard of AI. They relied on explicit instructions: if condition A is met, then perform action B. But real life doesn’t fit neatly into such crisp categories, so we developed neural networks—digital approximations of the way biological neurons send signals. These networks adjust themselves internally based on data, seeking connections that humans might not even see. That’s where the magic—and the madness—begins.

Machine learning isn’t about teaching an algorithm what cats are; it’s about feeding it thousands of images labeled as ‘cat’ and ‘not cat,’ then letting it figure out the difference itself. What emerges is a kind of internal model—a dense web of mathematical weights—that allows the system to make educated guesses. It’s marvelous, but it’s also fragile. A neural network trained to identify photos can stumble on an image that’s perfectly ordinary to you and me but utterly confounding to its learned patterns.

Understanding this distinction helps us appreciate AI’s accomplishments without attributing human-like intellect where there is none. Machines don’t think; they calculate patterns at enormous speed. When they surprise us—whether delightfully or disastrously—it’s because they’re following the patterns in their data, not because they understand the task the way humans do.

The quality of an AI system lives and dies by its training data. That’s a truth I emphasize again and again because it’s easy to forget that even the most sophisticated algorithms can’t escape their input. If your data is flawed, biased, or incomplete, your AI will inherit those flaws—and amplify them.

In my online experiments, I’ve fed neural networks all kinds of human-created data: names, recipes, color palettes, even pickup lines. What the AI produces always reflects the sum of its examples and the gaps between them. Train a network on romantic sentences, and you might get odd results like ‘You look like a thing and I love you’—sweet, accidental poetry born from pattern-matching, not emotion. Train it on jokes, and it generates punchlines that rarely make sense but reveal something fascinating about how humor relies on cultural and contextual understanding that machines don’t have.

Real-world consequences are serious. A facial recognition system trained mainly on lighter-skinned faces will misidentify people with darker skin. A hiring AI trained on past corporate data might perpetuate discrimination unintentionally. The algorithm is only as fair as its data, and the data is only as fair as the society that generates it.

Recognizing bias isn’t about rejecting AI—it’s about becoming aware that every dataset is a mirror. When we train AIs, we’re teaching them our habits, prejudices, priorities, and blind spots. Once we see that clearly, we can start designing data with fairness and diversity in mind, giving machines a better chance to reflect the world we want rather than the world we used to have.